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 <libavformat/avio.h>
  11 #include <libavutil/mathematics.h>
  12 #include <libavutil/rational.h>
  13 #include <libdrm/drm_fourcc.h>
  14 #include <stdint.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 <mutex>
  24 #include <queue>
  25 #include <string>
  26 #include <thread>
  27
  28 #include "context.h"
  29 #include "httpd.h"
  30 #include "timebase.h"
  31
  32 class QOpenGLContext;
  33 class QSurface;
  34
  35 #define CHECK_VASTATUS(va_status, func)                                 \
  36     if (va_status != VA_STATUS_SUCCESS) {                               \
  37         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  38         exit(1);                                                        \
  39     }
  40
  41 //#include "loadsurface.h"
  42
  43 #define NAL_REF_IDC_NONE        0
  44 #define NAL_REF_IDC_LOW         1
  45 #define NAL_REF_IDC_MEDIUM      2
  46 #define NAL_REF_IDC_HIGH        3
  47
  48 #define NAL_NON_IDR             1
  49 #define NAL_IDR                 5
  50 #define NAL_SPS                 7
  51 #define NAL_PPS                 8
  52 #define NAL_SEI                 6
  53
  54 #define SLICE_TYPE_P            0
  55 #define SLICE_TYPE_B            1
  56 #define SLICE_TYPE_I            2
  57 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  58 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  59 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  60
  61
  62 #define ENTROPY_MODE_CAVLC      0
  63 #define ENTROPY_MODE_CABAC      1
  64
  65 #define PROFILE_IDC_BASELINE    66
  66 #define PROFILE_IDC_MAIN        77
  67 #define PROFILE_IDC_HIGH        100
  68
  69 #define BITSTREAM_ALLOCATE_STEPPING     4096
  70
  71 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  72 static  VADisplay va_dpy;
  73 static  VAProfile h264_profile = (VAProfile)~0;
  74 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  75 static  int config_attrib_num = 0, enc_packed_header_idx;
  76
  77 struct GLSurface {
  78         VASurfaceID src_surface, ref_surface;
  79         VABufferID coded_buf;
  80
  81         VAImage surface_image;
  82         GLuint y_tex, cbcr_tex;
  83         EGLImage y_egl_image, cbcr_egl_image;
  84 };
  85 GLSurface gl_surfaces[SURFACE_NUM];
  86
  87 static  VAConfigID config_id;
  88 static  VAContextID context_id;
  89 static  VAEncSequenceParameterBufferH264 seq_param;
  90 static  VAEncPictureParameterBufferH264 pic_param;
  91 static  VAEncSliceParameterBufferH264 slice_param;
  92 static  VAPictureH264 CurrentCurrPic;
  93 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
  94
  95 static  unsigned int MaxFrameNum = (2<<16);
  96 static  unsigned int MaxPicOrderCntLsb = (2<<8);
  97 static  unsigned int Log2MaxFrameNum = 16;
  98 static  unsigned int Log2MaxPicOrderCntLsb = 8;
  99
 100 static  unsigned int num_ref_frames = 2;
 101 static  unsigned int numShortTerm = 0;
 102 static  int constraint_set_flag = 0;
 103 static  int h264_packedheader = 0; /* support pack header? */
 104 static  int h264_maxref = (1<<16|1);
 105 static  int h264_entropy_mode = 1; /* cabac */
 106
 107 static  char *coded_fn = NULL;
 108
 109 static  int frame_width = 176;
 110 static  int frame_height = 144;
 111 static  int frame_width_mbaligned;
 112 static  int frame_height_mbaligned;
 113 static  int frame_rate = 60;
 114 static  unsigned int frame_bitrate = 0;
 115 static  unsigned int frame_slices = 1;
 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 = 60;
 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 static unsigned int
 155 va_swap32(unsigned int val)
 156 {
 157     unsigned char *pval = (unsigned char *)&val;
 158
 159     return ((pval[0] << 24)     |
 160             (pval[1] << 16)     |
 161             (pval[2] << 8)      |
 162             (pval[3] << 0));
 163 }
 164
 165 static void
 166 bitstream_start(bitstream *bs)
 167 {
 168     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 169     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 170     bs->bit_offset = 0;
 171 }
 172
 173 static void
 174 bitstream_end(bitstream *bs)
 175 {
 176     int pos = (bs->bit_offset >> 5);
 177     int bit_offset = (bs->bit_offset & 0x1f);
 178     int bit_left = 32 - bit_offset;
 179
 180     if (bit_offset) {
 181         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 182     }
 183 }
 184
 185 static void
 186 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 187 {
 188     int pos = (bs->bit_offset >> 5);
 189     int bit_offset = (bs->bit_offset & 0x1f);
 190     int bit_left = 32 - bit_offset;
 191
 192     if (!size_in_bits)
 193         return;
 194
 195     bs->bit_offset += size_in_bits;
 196
 197     if (bit_left > size_in_bits) {
 198         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 199     } else {
 200         size_in_bits -= bit_left;
 201         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 202         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 203
 204         if (pos + 1 == bs->max_size_in_dword) {
 205             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 206             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 207         }
 208
 209         bs->buffer[pos + 1] = val;
 210     }
 211 }
 212
 213 static void
 214 bitstream_put_ue(bitstream *bs, unsigned int val)
 215 {
 216     int size_in_bits = 0;
 217     int tmp_val = ++val;
 218
 219     while (tmp_val) {
 220         tmp_val >>= 1;
 221         size_in_bits++;
 222     }
 223
 224     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 225     bitstream_put_ui(bs, val, size_in_bits);
 226 }
 227
 228 static void
 229 bitstream_put_se(bitstream *bs, int val)
 230 {
 231     unsigned int new_val;
 232
 233     if (val <= 0)
 234         new_val = -2 * val;
 235     else
 236         new_val = 2 * val - 1;
 237
 238     bitstream_put_ue(bs, new_val);
 239 }
 240
 241 static void
 242 bitstream_byte_aligning(bitstream *bs, int bit)
 243 {
 244     int bit_offset = (bs->bit_offset & 0x7);
 245     int bit_left = 8 - bit_offset;
 246     int new_val;
 247
 248     if (!bit_offset)
 249         return;
 250
 251     assert(bit == 0 || bit == 1);
 252
 253     if (bit)
 254         new_val = (1 << bit_left) - 1;
 255     else
 256         new_val = 0;
 257
 258     bitstream_put_ui(bs, new_val, bit_left);
 259 }
 260
 261 static void
 262 rbsp_trailing_bits(bitstream *bs)
 263 {
 264     bitstream_put_ui(bs, 1, 1);
 265     bitstream_byte_aligning(bs, 0);
 266 }
 267
 268 static void nal_start_code_prefix(bitstream *bs)
 269 {
 270     bitstream_put_ui(bs, 0x00000001, 32);
 271 }
 272
 273 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 274 {
 275     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 276     bitstream_put_ui(bs, nal_ref_idc, 2);
 277     bitstream_put_ui(bs, nal_unit_type, 5);
 278 }
 279
 280 static void sps_rbsp(bitstream *bs)
 281 {
 282     int profile_idc = PROFILE_IDC_BASELINE;
 283
 284     if (h264_profile  == VAProfileH264High)
 285         profile_idc = PROFILE_IDC_HIGH;
 286     else if (h264_profile  == VAProfileH264Main)
 287         profile_idc = PROFILE_IDC_MAIN;
 288
 289     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 290     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 291     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 292     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 293     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 294     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 295     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 296     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 297
 298     if ( profile_idc == PROFILE_IDC_HIGH) {
 299         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 300         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 301         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 302         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 303         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 304     }
 305
 306     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 307     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 308
 309     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 310         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 311     else {
 312         assert(0);
 313     }
 314
 315     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 316     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 317
 318     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 319     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 320     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 321
 322     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 323         assert(0);
 324     }
 325
 326     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 327     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 328
 329     if (seq_param.frame_cropping_flag) {
 330         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 331         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 332         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 333         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 334     }
 335
 336     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 337     if ( false ) {
 338         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 339     } else {
 340         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 341         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 342         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 343         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 344         {
 345             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 346             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 347             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 348             {
 349                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 350                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 351                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 352             }
 353         }
 354         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 355         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 356         {
 357             bitstream_put_ui(bs, 1, 32);  // FPS
 358             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 359             bitstream_put_ui(bs, 1, 1);
 360         }
 361         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 362         {
 363             // hrd_parameters
 364             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 365             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 366             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 367
 368             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 369             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 370             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 371
 372             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 373             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 374             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 375             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 376         }
 377         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 378         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 379
 380         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 381         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 382     }
 383
 384     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 385 }
 386
 387
 388 static void pps_rbsp(bitstream *bs)
 389 {
 390     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 391     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 392
 393     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 394
 395     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 396
 397     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 398
 399     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 400     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 401
 402     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 403     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 404
 405     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 406     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 407     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 408
 409     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 410     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 411     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 412
 413     /* more_rbsp_data */
 414     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 415     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 416     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 417
 418     rbsp_trailing_bits(bs);
 419 }
 420
 421 static void slice_header(bitstream *bs)
 422 {
 423     int first_mb_in_slice = slice_param.macroblock_address;
 424
 425     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 426     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 427     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 428     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 429
 430     /* frame_mbs_only_flag == 1 */
 431     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 432         /* FIXME: */
 433         assert(0);
 434     }
 435
 436     if (pic_param.pic_fields.bits.idr_pic_flag)
 437         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 438
 439     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 440         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 441         /* pic_order_present_flag == 0 */
 442     } else {
 443         /* FIXME: */
 444         assert(0);
 445     }
 446
 447     /* redundant_pic_cnt_present_flag == 0 */
 448     /* slice type */
 449     if (IS_P_SLICE(slice_param.slice_type)) {
 450         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 451
 452         if (slice_param.num_ref_idx_active_override_flag)
 453             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 454
 455         /* ref_pic_list_reordering */
 456         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 457     } else if (IS_B_SLICE(slice_param.slice_type)) {
 458         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 459
 460         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 461
 462         if (slice_param.num_ref_idx_active_override_flag) {
 463             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 464             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 465         }
 466
 467         /* ref_pic_list_reordering */
 468         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 469         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 470     }
 471
 472     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 473          IS_P_SLICE(slice_param.slice_type)) ||
 474         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 475          IS_B_SLICE(slice_param.slice_type))) {
 476         /* FIXME: fill weight/offset table */
 477         assert(0);
 478     }
 479
 480     /* dec_ref_pic_marking */
 481     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 482         unsigned char no_output_of_prior_pics_flag = 0;
 483         unsigned char long_term_reference_flag = 0;
 484         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 485
 486         if (pic_param.pic_fields.bits.idr_pic_flag) {
 487             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 488             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 489         } else {
 490             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 491         }
 492     }
 493
 494     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 495         !IS_I_SLICE(slice_param.slice_type))
 496         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 497
 498     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 499
 500     /* ignore for SP/SI */
 501
 502     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 503         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 504
 505         if (slice_param.disable_deblocking_filter_idc != 1) {
 506             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 507             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 508         }
 509     }
 510
 511     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 512         bitstream_byte_aligning(bs, 1);
 513     }
 514 }
 515
 516 static int
 517 build_packed_pic_buffer(unsigned char **header_buffer)
 518 {
 519     bitstream bs;
 520
 521     bitstream_start(&bs);
 522     nal_start_code_prefix(&bs);
 523     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 524     pps_rbsp(&bs);
 525     bitstream_end(&bs);
 526
 527     *header_buffer = (unsigned char *)bs.buffer;
 528     return bs.bit_offset;
 529 }
 530
 531 static int
 532 build_packed_seq_buffer(unsigned char **header_buffer)
 533 {
 534     bitstream bs;
 535
 536     bitstream_start(&bs);
 537     nal_start_code_prefix(&bs);
 538     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 539     sps_rbsp(&bs);
 540     bitstream_end(&bs);
 541
 542     *header_buffer = (unsigned char *)bs.buffer;
 543     return bs.bit_offset;
 544 }
 545
 546 static int build_packed_slice_buffer(unsigned char **header_buffer)
 547 {
 548     bitstream bs;
 549     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 550     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 551
 552     bitstream_start(&bs);
 553     nal_start_code_prefix(&bs);
 554
 555     if (IS_I_SLICE(slice_param.slice_type)) {
 556         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 557     } else if (IS_P_SLICE(slice_param.slice_type)) {
 558         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 559     } else {
 560         assert(IS_B_SLICE(slice_param.slice_type));
 561         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 562     }
 563
 564     slice_header(&bs);
 565     bitstream_end(&bs);
 566
 567     *header_buffer = (unsigned char *)bs.buffer;
 568     return bs.bit_offset;
 569 }
 570
 571
 572 /*
 573   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 574   1) period between Frame #X and Frame #N = #X - #N
 575   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 576   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 577   4) intra_period and intra_idr_period take precedence over ip_period
 578   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 579      of I/IDR frames, see bellow examples
 580   -------------------------------------------------------------------
 581   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 582   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 583   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 584   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 585   1            1                ignored    IDR IDR IDR IDR...
 586   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 587   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 588   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 589                                               (PBB)(IBB)(PBB)(IBB)...
 590   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 591                                            IDRPPPPP IPPPPP IPPPPP...
 592   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 593                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 594                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 595                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 596                                            {IDR(PBB)(PBB)}                     (6/6/3)
 597                                            {IDR(PBB)(PBB)}.
 598 */
 599
 600 // General pts/dts strategy:
 601 //
 602 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 603 // bit tricky. We assume first of all that the frame rate never goes _above_
 604 // <frame_rate>, which gives us a frame period N. The decoder can always decode
 605 // in at least this speed, as long at dts <= pts (the frame is not attempted
 606 // presented before it is decoded). Furthermore, we never have longer chains of
 607 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 608 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 609 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 610 // frame, with an order O <= C.)
 611 //
 612 // Many strategies are possible, but we establish these rules:
 613 //
 614 //  - Tip frames have dts = pts - (C-O)*N.
 615 //  - Non-tip frames have dts = dts_last + N.
 616 //
 617 // An example, with C=2 and N=10 and the data flow showed with arrows:
 618 //
 619 //        I  B  P  B  B  P
 620 //   pts: 30 40 50 60 70 80
 621 //        ↓  ↓     ↓
 622 //   dts: 10 30 20 60 50←40
 623 //         |  |  ↑        ↑
 624 //         `--|--'        |
 625 //             `----------'
 626 //
 627 // To show that this works fine also with irregular spacings, let's say that
 628 // the third frame is delayed a bit (something earlier was dropped). Now the
 629 // situation looks like this:
 630 //
 631 //        I  B  P  B  B   P
 632 //   pts: 30 40 80 90 100 110
 633 //        ↓  ↓     ↓
 634 //   dts: 10 30 20 90 50←40
 635 //         |  |  ↑        ↑
 636 //         `--|--'        |
 637 //             `----------'
 638 //
 639 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 640 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 641 //
 642 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 643 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 644 // a tip frame and should be given a dts based on the previous one.
 645 #define FRAME_P 0
 646 #define FRAME_B 1
 647 #define FRAME_I 2
 648 #define FRAME_IDR 7
 649 void encoding2display_order(
 650     unsigned long long encoding_order, int intra_period,
 651     int intra_idr_period, int ip_period,
 652     unsigned long long *displaying_order,
 653     int *frame_type, int *pts_lag)
 654 {
 655     int encoding_order_gop = 0;
 656
 657     *pts_lag = 0;
 658
 659     if (intra_period == 1) { /* all are I/IDR frames */
 660         *displaying_order = encoding_order;
 661         if (intra_idr_period == 0)
 662             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 663         else
 664             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 665         return;
 666     }
 667
 668     if (intra_period == 0)
 669         intra_idr_period = 0;
 670
 671     if (ip_period == 1) {
 672         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 673         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 674         *displaying_order = encoding_order;
 675
 676         if (encoding_order_gop == 0) { /* the first frame */
 677             *frame_type = FRAME_IDR;
 678         } else if (intra_period != 0 && /* have I frames */
 679                    encoding_order_gop >= 2 &&
 680                    (encoding_order_gop % intra_period == 0)) {
 681             *frame_type = FRAME_I;
 682         } else {
 683             *frame_type = FRAME_P;
 684         }
 685         return;
 686     }
 687
 688     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 689     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 690     *pts_lag = -1;  // Most frames are not tip frames.
 691
 692     if (encoding_order_gop == 0) { /* the first frame */
 693         *frame_type = FRAME_IDR;
 694         *displaying_order = encoding_order;
 695         // IDR frames are a special case; I honestly can't find the logic behind
 696         // why this is the right thing, but it seems to line up nicely in practice :-)
 697         *pts_lag = TIMEBASE / frame_rate;
 698     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 699         *frame_type = FRAME_B;
 700         *displaying_order = encoding_order - 1;
 701         if ((encoding_order_gop % ip_period) == 0) {
 702             *pts_lag = 0;  // Last B-frame.
 703         }
 704     } else if (intra_period != 0 && /* have I frames */
 705                encoding_order_gop >= 2 &&
 706                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 707         *frame_type = FRAME_I;
 708         *displaying_order = encoding_order + ip_period - 1;
 709     } else {
 710         *frame_type = FRAME_P;
 711         *displaying_order = encoding_order + ip_period - 1;
 712     }
 713 }
 714
 715
 716 static const char *rc_to_string(int rcmode)
 717 {
 718     switch (rc_mode) {
 719     case VA_RC_NONE:
 720         return "NONE";
 721     case VA_RC_CBR:
 722         return "CBR";
 723     case VA_RC_VBR:
 724         return "VBR";
 725     case VA_RC_VCM:
 726         return "VCM";
 727     case VA_RC_CQP:
 728         return "CQP";
 729     case VA_RC_VBR_CONSTRAINED:
 730         return "VBR_CONSTRAINED";
 731     default:
 732         return "Unknown";
 733     }
 734 }
 735
 736 #if 0
 737 static int process_cmdline(int argc, char *argv[])
 738 {
 739     char c;
 740     const struct option long_opts[] = {
 741         {"help", no_argument, NULL, 0 },
 742         {"bitrate", required_argument, NULL, 1 },
 743         {"minqp", required_argument, NULL, 2 },
 744         {"initialqp", required_argument, NULL, 3 },
 745         {"intra_period", required_argument, NULL, 4 },
 746         {"idr_period", required_argument, NULL, 5 },
 747         {"ip_period", required_argument, NULL, 6 },
 748         {"rcmode", required_argument, NULL, 7 },
 749         {"srcyuv", required_argument, NULL, 9 },
 750         {"recyuv", required_argument, NULL, 10 },
 751         {"fourcc", required_argument, NULL, 11 },
 752         {"syncmode", no_argument, NULL, 12 },
 753         {"enablePSNR", no_argument, NULL, 13 },
 754         {"prit", required_argument, NULL, 14 },
 755         {"priv", required_argument, NULL, 15 },
 756         {"framecount", required_argument, NULL, 16 },
 757         {"entropy", required_argument, NULL, 17 },
 758         {"profile", required_argument, NULL, 18 },
 759         {NULL, no_argument, NULL, 0 }};
 760     int long_index;
 761
 762     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 763         switch (c) {
 764         case 'w':
 765             frame_width = atoi(optarg);
 766             break;
 767         case 'h':
 768             frame_height = atoi(optarg);
 769             break;
 770         case 'n':
 771         case 'f':
 772             frame_rate = atoi(optarg);
 773             break;
 774         case 'o':
 775             coded_fn = strdup(optarg);
 776             break;
 777         case 0:
 778             print_help();
 779             exit(0);
 780         case 1:
 781             frame_bitrate = atoi(optarg);
 782             break;
 783         case 2:
 784             minimal_qp = atoi(optarg);
 785             break;
 786         case 3:
 787             initial_qp = atoi(optarg);
 788             break;
 789         case 4:
 790             intra_period = atoi(optarg);
 791             break;
 792         case 5:
 793             intra_idr_period = atoi(optarg);
 794             break;
 795         case 6:
 796             ip_period = atoi(optarg);
 797             break;
 798         case 7:
 799             rc_mode = string_to_rc(optarg);
 800             if (rc_mode < 0) {
 801                 print_help();
 802                 exit(1);
 803             }
 804             break;
 805         case 9:
 806             srcyuv_fn = strdup(optarg);
 807             break;
 808         case 11:
 809             srcyuv_fourcc = string_to_fourcc(optarg);
 810             if (srcyuv_fourcc <= 0) {
 811                 print_help();
 812                 exit(1);
 813             }
 814             break;
 815         case 13:
 816             calc_psnr = 1;
 817             break;
 818         case 14:
 819             misc_priv_type = strtol(optarg, NULL, 0);
 820             break;
 821         case 15:
 822             misc_priv_value = strtol(optarg, NULL, 0);
 823             break;
 824         case 17:
 825             h264_entropy_mode = atoi(optarg) ? 1: 0;
 826             break;
 827         case 18:
 828             if (strncmp(optarg, "BP", 2) == 0)
 829                 h264_profile = VAProfileH264Baseline;
 830             else if (strncmp(optarg, "MP", 2) == 0)
 831                 h264_profile = VAProfileH264Main;
 832             else if (strncmp(optarg, "HP", 2) == 0)
 833                 h264_profile = VAProfileH264High;
 834             else
 835                 h264_profile = (VAProfile)0;
 836             break;
 837         case ':':
 838         case '?':
 839             print_help();
 840             exit(0);
 841         }
 842     }
 843
 844     if (ip_period < 1) {
 845         printf(" ip_period must be greater than 0\n");
 846         exit(0);
 847     }
 848     if (intra_period != 1 && intra_period % ip_period != 0) {
 849         printf(" intra_period must be a multiplier of ip_period\n");
 850         exit(0);
 851     }
 852     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 853         printf(" intra_idr_period must be a multiplier of intra_period\n");
 854         exit(0);
 855     }
 856
 857     if (frame_bitrate == 0)
 858         frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
 859
 860     if (coded_fn == NULL) {
 861         struct stat buf;
 862         if (stat("/tmp", &buf) == 0)
 863             coded_fn = strdup("/tmp/test.264");
 864         else if (stat("/sdcard", &buf) == 0)
 865             coded_fn = strdup("/sdcard/test.264");
 866         else
 867             coded_fn = strdup("./test.264");
 868     }
 869
 870
 871     frame_width_mbaligned = (frame_width + 15) & (~15);
 872     frame_height_mbaligned = (frame_height + 15) & (~15);
 873     if (frame_width != frame_width_mbaligned ||
 874         frame_height != frame_height_mbaligned) {
 875         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 876                frame_width, frame_height,
 877                frame_width_mbaligned, frame_height_mbaligned
 878                );
 879     }
 880
 881     return 0;
 882 }
 883 #endif
 884
 885 static Display *x11_display;
 886 static Window   x11_window;
 887
 888 VADisplay
 889 va_open_display(void)
 890 {
 891     x11_display = XOpenDisplay(NULL);
 892     if (!x11_display) {
 893         fprintf(stderr, "error: can't connect to X server!\n");
 894         return NULL;
 895     }
 896     return vaGetDisplay(x11_display);
 897 }
 898
 899 void
 900 va_close_display(VADisplay va_dpy)
 901 {
 902     if (!x11_display)
 903         return;
 904
 905     if (x11_window) {
 906         XUnmapWindow(x11_display, x11_window);
 907         XDestroyWindow(x11_display, x11_window);
 908         x11_window = None;
 909     }
 910     XCloseDisplay(x11_display);
 911     x11_display = NULL;
 912 }
 913
 914 static int init_va(void)
 915 {
 916     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 917     VAEntrypoint *entrypoints;
 918     int num_entrypoints, slice_entrypoint;
 919     int support_encode = 0;
 920     int major_ver, minor_ver;
 921     VAStatus va_status;
 922     unsigned int i;
 923
 924     va_dpy = va_open_display();
 925     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 926     CHECK_VASTATUS(va_status, "vaInitialize");
 927
 928     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 929     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 930     if (!entrypoints) {
 931         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 932         exit(1);
 933     }
 934
 935     /* use the highest profile */
 936     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 937         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 938             continue;
 939
 940         h264_profile = profile_list[i];
 941         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 942         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 943             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 944                 support_encode = 1;
 945                 break;
 946             }
 947         }
 948         if (support_encode == 1)
 949             break;
 950     }
 951
 952     if (support_encode == 0) {
 953         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 954         exit(1);
 955     } else {
 956         switch (h264_profile) {
 957             case VAProfileH264Baseline:
 958                 printf("Use profile VAProfileH264Baseline\n");
 959                 ip_period = 1;
 960                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 961                 h264_entropy_mode = 0;
 962                 break;
 963             case VAProfileH264ConstrainedBaseline:
 964                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 965                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 966                 ip_period = 1;
 967                 break;
 968
 969             case VAProfileH264Main:
 970                 printf("Use profile VAProfileH264Main\n");
 971                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 972                 break;
 973
 974             case VAProfileH264High:
 975                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 976                 printf("Use profile VAProfileH264High\n");
 977                 break;
 978             default:
 979                 printf("unknow profile. Set to Baseline");
 980                 h264_profile = VAProfileH264Baseline;
 981                 ip_period = 1;
 982                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 983                 break;
 984         }
 985     }
 986
 987     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 988
 989     /* find out the format for the render target, and rate control mode */
 990     for (i = 0; i < VAConfigAttribTypeMax; i++)
 991         attrib[i].type = (VAConfigAttribType)i;
 992
 993     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 994                                       &attrib[0], VAConfigAttribTypeMax);
 995     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 996     /* check the interested configattrib */
 997     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 998         printf("Not find desired YUV420 RT format\n");
 999         exit(1);
1000     } else {
1001         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1002         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1003         config_attrib_num++;
1004     }
1005
1006     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1007         int tmp = attrib[VAConfigAttribRateControl].value;
1008
1009         printf("Support rate control mode (0x%x):", tmp);
1010
1011         if (tmp & VA_RC_NONE)
1012             printf("NONE ");
1013         if (tmp & VA_RC_CBR)
1014             printf("CBR ");
1015         if (tmp & VA_RC_VBR)
1016             printf("VBR ");
1017         if (tmp & VA_RC_VCM)
1018             printf("VCM ");
1019         if (tmp & VA_RC_CQP)
1020             printf("CQP ");
1021         if (tmp & VA_RC_VBR_CONSTRAINED)
1022             printf("VBR_CONSTRAINED ");
1023
1024         printf("\n");
1025
1026         if (rc_mode == -1 || !(rc_mode & tmp))  {
1027             if (rc_mode != -1) {
1028                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1029             }
1030
1031             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1032                 if (rc_default_modes[i] & tmp) {
1033                     rc_mode = rc_default_modes[i];
1034                     break;
1035                 }
1036             }
1037
1038             printf("RateControl mode: %s\n", rc_to_string(rc_mode));
1039         }
1040
1041         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1042         config_attrib[config_attrib_num].value = rc_mode;
1043         config_attrib_num++;
1044     }
1045
1046
1047     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1048         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1049
1050         printf("Support VAConfigAttribEncPackedHeaders\n");
1051
1052         h264_packedheader = 1;
1053         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1054         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1055
1056         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1057             printf("Support packed sequence headers\n");
1058             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1059         }
1060
1061         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1062             printf("Support packed picture headers\n");
1063             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1064         }
1065
1066         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1067             printf("Support packed slice headers\n");
1068             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1069         }
1070
1071         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1072             printf("Support packed misc headers\n");
1073             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1074         }
1075
1076         enc_packed_header_idx = config_attrib_num;
1077         config_attrib_num++;
1078     }
1079
1080     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1081         int tmp = attrib[VAConfigAttribEncInterlaced].value;
1082
1083         printf("Support VAConfigAttribEncInterlaced\n");
1084
1085         if (tmp & VA_ENC_INTERLACED_FRAME)
1086             printf("support VA_ENC_INTERLACED_FRAME\n");
1087         if (tmp & VA_ENC_INTERLACED_FIELD)
1088             printf("Support VA_ENC_INTERLACED_FIELD\n");
1089         if (tmp & VA_ENC_INTERLACED_MBAFF)
1090             printf("Support VA_ENC_INTERLACED_MBAFF\n");
1091         if (tmp & VA_ENC_INTERLACED_PAFF)
1092             printf("Support VA_ENC_INTERLACED_PAFF\n");
1093
1094         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1095         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1096         config_attrib_num++;
1097     }
1098
1099     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1100         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1101
1102         printf("Support %d RefPicList0 and %d RefPicList1\n",
1103                h264_maxref & 0xffff, (h264_maxref >> 16) & 0xffff );
1104     }
1105
1106     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1107         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1108
1109     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1110         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1111
1112         printf("Support VAConfigAttribEncSliceStructure\n");
1113
1114         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1115             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1116         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1117             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1118         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1119             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1120     }
1121     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1122         printf("Support VAConfigAttribEncMacroblockInfo\n");
1123     }
1124
1125     free(entrypoints);
1126     return 0;
1127 }
1128
1129 static int setup_encode()
1130 {
1131     VAStatus va_status;
1132     VASurfaceID *tmp_surfaceid;
1133     int codedbuf_size, i;
1134     static VASurfaceID src_surface[SURFACE_NUM];
1135     static VASurfaceID ref_surface[SURFACE_NUM];
1136
1137     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1138             &config_attrib[0], config_attrib_num, &config_id);
1139     CHECK_VASTATUS(va_status, "vaCreateConfig");
1140
1141     /* create source surfaces */
1142     va_status = vaCreateSurfaces(va_dpy,
1143                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1144                                  &src_surface[0], SURFACE_NUM,
1145                                  NULL, 0);
1146     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1147
1148     /* create reference surfaces */
1149     va_status = vaCreateSurfaces(va_dpy,
1150                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1151                                  &ref_surface[0], SURFACE_NUM,
1152                                  NULL, 0);
1153     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1154
1155     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1156     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1157     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1158
1159     /* Create a context for this encode pipe */
1160     va_status = vaCreateContext(va_dpy, config_id,
1161                                 frame_width_mbaligned, frame_height_mbaligned,
1162                                 VA_PROGRESSIVE,
1163                                 tmp_surfaceid, 2 * SURFACE_NUM,
1164                                 &context_id);
1165     CHECK_VASTATUS(va_status, "vaCreateContext");
1166     free(tmp_surfaceid);
1167
1168     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1169
1170     for (i = 0; i < SURFACE_NUM; i++) {
1171         /* create coded buffer once for all
1172          * other VA buffers which won't be used again after vaRenderPicture.
1173          * so APP can always vaCreateBuffer for every frame
1174          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1175          * so VA won't maintain the coded buffer
1176          */
1177         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1178                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1179         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1180     }
1181
1182     /* create OpenGL objects */
1183     //glGenFramebuffers(SURFACE_NUM, fbos);
1184
1185     for (i = 0; i < SURFACE_NUM; i++) {
1186         glGenTextures(1, &gl_surfaces[i].y_tex);
1187         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1188     }
1189
1190     for (i = 0; i < SURFACE_NUM; i++) {
1191         gl_surfaces[i].src_surface = src_surface[i];
1192         gl_surfaces[i].ref_surface = ref_surface[i];
1193     }
1194
1195     return 0;
1196 }
1197
1198
1199
1200 #define partition(ref, field, key, ascending)   \
1201     while (i <= j) {                            \
1202         if (ascending) {                        \
1203             while (ref[i].field < key)          \
1204                 i++;                            \
1205             while (ref[j].field > key)          \
1206                 j--;                            \
1207         } else {                                \
1208             while (ref[i].field > key)          \
1209                 i++;                            \
1210             while (ref[j].field < key)          \
1211                 j--;                            \
1212         }                                       \
1213         if (i <= j) {                           \
1214             tmp = ref[i];                       \
1215             ref[i] = ref[j];                    \
1216             ref[j] = tmp;                       \
1217             i++;                                \
1218             j--;                                \
1219         }                                       \
1220     }                                           \
1221
1222 static void sort_one(VAPictureH264 ref[], int left, int right,
1223                      int ascending, int frame_idx)
1224 {
1225     int i = left, j = right;
1226     unsigned int key;
1227     VAPictureH264 tmp;
1228
1229     if (frame_idx) {
1230         key = ref[(left + right) / 2].frame_idx;
1231         partition(ref, frame_idx, key, ascending);
1232     } else {
1233         key = ref[(left + right) / 2].TopFieldOrderCnt;
1234         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1235     }
1236
1237     /* recursion */
1238     if (left < j)
1239         sort_one(ref, left, j, ascending, frame_idx);
1240
1241     if (i < right)
1242         sort_one(ref, i, right, ascending, frame_idx);
1243 }
1244
1245 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1246                      int partition_ascending, int list0_ascending, int list1_ascending)
1247 {
1248     int i = left, j = right;
1249     VAPictureH264 tmp;
1250
1251     if (frame_idx) {
1252         partition(ref, frame_idx, key, partition_ascending);
1253     } else {
1254         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1255     }
1256
1257
1258     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1259     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1260 }
1261
1262 static int update_ReferenceFrames(void)
1263 {
1264     int i;
1265
1266     if (current_frame_type == FRAME_B)
1267         return 0;
1268
1269     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1270     numShortTerm++;
1271     if (numShortTerm > num_ref_frames)
1272         numShortTerm = num_ref_frames;
1273     for (i=numShortTerm-1; i>0; i--)
1274         ReferenceFrames[i] = ReferenceFrames[i-1];
1275     ReferenceFrames[0] = CurrentCurrPic;
1276
1277     if (current_frame_type != FRAME_B)
1278         current_frame_num++;
1279     if (current_frame_num > MaxFrameNum)
1280         current_frame_num = 0;
1281
1282     return 0;
1283 }
1284
1285
1286 static int update_RefPicList(void)
1287 {
1288     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1289
1290     if (current_frame_type == FRAME_P) {
1291         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1292         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1293     }
1294
1295     if (current_frame_type == FRAME_B) {
1296         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1297         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1298                  1, 0, 1);
1299
1300         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1301         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1302                  0, 1, 0);
1303     }
1304
1305     return 0;
1306 }
1307
1308
1309 static int render_sequence(void)
1310 {
1311     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1312     VAStatus va_status;
1313     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1314     VAEncMiscParameterRateControl *misc_rate_ctrl;
1315
1316     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1317     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1318     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1319     seq_param.bits_per_second = frame_bitrate;
1320
1321     seq_param.intra_period = intra_period;
1322     seq_param.intra_idr_period = intra_idr_period;
1323     seq_param.ip_period = ip_period;
1324
1325     seq_param.max_num_ref_frames = num_ref_frames;
1326     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1327     seq_param.time_scale = TIMEBASE * 2;
1328     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1329     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1330     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1331     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1332     seq_param.seq_fields.bits.chroma_format_idc = 1;
1333     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1334
1335     if (frame_width != frame_width_mbaligned ||
1336         frame_height != frame_height_mbaligned) {
1337         seq_param.frame_cropping_flag = 1;
1338         seq_param.frame_crop_left_offset = 0;
1339         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1340         seq_param.frame_crop_top_offset = 0;
1341         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1342     }
1343
1344     va_status = vaCreateBuffer(va_dpy, context_id,
1345                                VAEncSequenceParameterBufferType,
1346                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1347     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1348
1349     va_status = vaCreateBuffer(va_dpy, context_id,
1350                                VAEncMiscParameterBufferType,
1351                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1352                                1, NULL, &rc_param_buf);
1353     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1354
1355     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1356     misc_param->type = VAEncMiscParameterTypeRateControl;
1357     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1358     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1359     misc_rate_ctrl->bits_per_second = frame_bitrate;
1360     misc_rate_ctrl->target_percentage = 66;
1361     misc_rate_ctrl->window_size = 1000;
1362     misc_rate_ctrl->initial_qp = initial_qp;
1363     misc_rate_ctrl->min_qp = minimal_qp;
1364     misc_rate_ctrl->basic_unit_size = 0;
1365     vaUnmapBuffer(va_dpy, rc_param_buf);
1366
1367     render_id[0] = seq_param_buf;
1368     render_id[1] = rc_param_buf;
1369
1370     va_status = vaRenderPicture(va_dpy, context_id, &render_id[0], 2);
1371     CHECK_VASTATUS(va_status, "vaRenderPicture");;
1372
1373     if (misc_priv_type != 0) {
1374         va_status = vaCreateBuffer(va_dpy, context_id,
1375                                    VAEncMiscParameterBufferType,
1376                                    sizeof(VAEncMiscParameterBuffer),
1377                                    1, NULL, &misc_param_tmpbuf);
1378         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1379         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1380         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1381         misc_param_tmp->data[0] = misc_priv_value;
1382         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1383
1384         va_status = vaRenderPicture(va_dpy, context_id, &misc_param_tmpbuf, 1);
1385     }
1386
1387     return 0;
1388 }
1389
1390 static int calc_poc(int pic_order_cnt_lsb)
1391 {
1392     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1393     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1394     int PicOrderCntMsb, TopFieldOrderCnt;
1395
1396     if (current_frame_type == FRAME_IDR)
1397         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1398     else {
1399         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1400         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1401     }
1402
1403     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1404         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1405         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1406     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1407              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1408         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1409     else
1410         PicOrderCntMsb = prevPicOrderCntMsb;
1411
1412     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1413
1414     if (current_frame_type != FRAME_B) {
1415         PicOrderCntMsb_ref = PicOrderCntMsb;
1416         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1417     }
1418
1419     return TopFieldOrderCnt;
1420 }
1421
1422 static int render_picture(void)
1423 {
1424     VABufferID pic_param_buf;
1425     VAStatus va_status;
1426     int i = 0;
1427
1428     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1429     pic_param.CurrPic.frame_idx = current_frame_num;
1430     pic_param.CurrPic.flags = 0;
1431     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1432     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1433     CurrentCurrPic = pic_param.CurrPic;
1434
1435     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1436         update_RefPicList(); /* calc RefPicList */
1437         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1438         if (current_frame_type == FRAME_P) {
1439             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1440         } else if (current_frame_type == FRAME_B) {
1441             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1442             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1443         }
1444     } else {
1445         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1446         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1447             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1448             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1449         }
1450     }
1451
1452     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1453     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1454     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1455     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1456     pic_param.frame_num = current_frame_num;
1457     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1458     pic_param.last_picture = false;  // FIXME
1459     pic_param.pic_init_qp = initial_qp;
1460
1461     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1462                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1463     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1464
1465     va_status = vaRenderPicture(va_dpy, context_id, &pic_param_buf, 1);
1466     CHECK_VASTATUS(va_status, "vaRenderPicture");
1467
1468     return 0;
1469 }
1470
1471 static int render_packedsequence(void)
1472 {
1473     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1474     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1475     unsigned int length_in_bits;
1476     unsigned char *packedseq_buffer = NULL;
1477     VAStatus va_status;
1478
1479     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1480
1481     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1482
1483     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1484     packedheader_param_buffer.has_emulation_bytes = 0;
1485     va_status = vaCreateBuffer(va_dpy,
1486                                context_id,
1487                                VAEncPackedHeaderParameterBufferType,
1488                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1489                                &packedseq_para_bufid);
1490     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1491
1492     va_status = vaCreateBuffer(va_dpy,
1493                                context_id,
1494                                VAEncPackedHeaderDataBufferType,
1495                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1496                                &packedseq_data_bufid);
1497     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1498
1499     render_id[0] = packedseq_para_bufid;
1500     render_id[1] = packedseq_data_bufid;
1501     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1502     CHECK_VASTATUS(va_status, "vaRenderPicture");
1503
1504     free(packedseq_buffer);
1505
1506     return 0;
1507 }
1508
1509
1510 static int render_packedpicture(void)
1511 {
1512     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1513     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1514     unsigned int length_in_bits;
1515     unsigned char *packedpic_buffer = NULL;
1516     VAStatus va_status;
1517
1518     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1519     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1520     packedheader_param_buffer.bit_length = length_in_bits;
1521     packedheader_param_buffer.has_emulation_bytes = 0;
1522
1523     va_status = vaCreateBuffer(va_dpy,
1524                                context_id,
1525                                VAEncPackedHeaderParameterBufferType,
1526                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1527                                &packedpic_para_bufid);
1528     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1529
1530     va_status = vaCreateBuffer(va_dpy,
1531                                context_id,
1532                                VAEncPackedHeaderDataBufferType,
1533                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1534                                &packedpic_data_bufid);
1535     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1536
1537     render_id[0] = packedpic_para_bufid;
1538     render_id[1] = packedpic_data_bufid;
1539     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1540     CHECK_VASTATUS(va_status, "vaRenderPicture");
1541
1542     free(packedpic_buffer);
1543
1544     return 0;
1545 }
1546
1547 static void render_packedslice()
1548 {
1549     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1550     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1551     unsigned int length_in_bits;
1552     unsigned char *packedslice_buffer = NULL;
1553     VAStatus va_status;
1554
1555     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1556     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1557     packedheader_param_buffer.bit_length = length_in_bits;
1558     packedheader_param_buffer.has_emulation_bytes = 0;
1559
1560     va_status = vaCreateBuffer(va_dpy,
1561                                context_id,
1562                                VAEncPackedHeaderParameterBufferType,
1563                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1564                                &packedslice_para_bufid);
1565     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1566
1567     va_status = vaCreateBuffer(va_dpy,
1568                                context_id,
1569                                VAEncPackedHeaderDataBufferType,
1570                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1571                                &packedslice_data_bufid);
1572     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1573
1574     render_id[0] = packedslice_para_bufid;
1575     render_id[1] = packedslice_data_bufid;
1576     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1577     CHECK_VASTATUS(va_status, "vaRenderPicture");
1578
1579     free(packedslice_buffer);
1580 }
1581
1582 static int render_slice(void)
1583 {
1584     VABufferID slice_param_buf;
1585     VAStatus va_status;
1586     int i;
1587
1588     update_RefPicList();
1589
1590     /* one frame, one slice */
1591     slice_param.macroblock_address = 0;
1592     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1593     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1594     if (current_frame_type == FRAME_IDR) {
1595         if (current_frame_encoding != 0)
1596             ++slice_param.idr_pic_id;
1597     } else if (current_frame_type == FRAME_P) {
1598         int refpiclist0_max = h264_maxref & 0xffff;
1599         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1600
1601         for (i = refpiclist0_max; i < 32; i++) {
1602             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1603             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1604         }
1605     } else if (current_frame_type == FRAME_B) {
1606         int refpiclist0_max = h264_maxref & 0xffff;
1607         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1608
1609         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1610         for (i = refpiclist0_max; i < 32; i++) {
1611             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1612             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1613         }
1614
1615         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1616         for (i = refpiclist1_max; i < 32; i++) {
1617             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1618             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1619         }
1620     }
1621
1622     slice_param.slice_alpha_c0_offset_div2 = 0;
1623     slice_param.slice_beta_offset_div2 = 0;
1624     slice_param.direct_spatial_mv_pred_flag = 1;
1625     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1626
1627
1628     if (h264_packedheader &&
1629         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1630         render_packedslice();
1631
1632     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1633                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1634     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1635
1636     va_status = vaRenderPicture(va_dpy, context_id, &slice_param_buf, 1);
1637     CHECK_VASTATUS(va_status, "vaRenderPicture");
1638
1639     return 0;
1640 }
1641
1642
1643
1644 int H264Encoder::save_codeddata(storage_task task)
1645 {
1646     VACodedBufferSegment *buf_list = NULL;
1647     VAStatus va_status;
1648     unsigned int coded_size = 0;
1649
1650     string data;
1651
1652     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / frame_rate);  // So we never get negative dts.
1653
1654     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1655     CHECK_VASTATUS(va_status, "vaMapBuffer");
1656     while (buf_list != NULL) {
1657         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1658         buf_list = (VACodedBufferSegment *) buf_list->next;
1659
1660         frame_size += coded_size;
1661     }
1662     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1663
1664     {
1665         // Add video.
1666         AVPacket pkt;
1667         memset(&pkt, 0, sizeof(pkt));
1668         pkt.buf = nullptr;
1669         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1670         pkt.size = data.size();
1671         pkt.stream_index = 0;
1672         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1673             pkt.flags = AV_PKT_FLAG_KEY;
1674         } else {
1675             pkt.flags = 0;
1676         }
1677         //pkt.duration = 1;
1678         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1679     }
1680     // Encode and add all audio frames up to and including the pts of this video frame.
1681     // (They can never be queued to us after the video frame they belong to, only before.)
1682     for ( ;; ) {
1683         int64_t audio_pts;
1684         std::vector<float> audio;
1685         {
1686              unique_lock<mutex> lock(frame_queue_mutex);
1687              if (pending_audio_frames.empty()) break;
1688              auto it = pending_audio_frames.begin();
1689              if (it->first > task.pts) break;
1690              audio_pts = it->first;
1691              audio = move(it->second);
1692              pending_audio_frames.erase(it);
1693         }
1694
1695         AVFrame *frame = avcodec_alloc_frame();
1696         frame->nb_samples = audio.size() / 2;
1697         frame->format = AV_SAMPLE_FMT_FLT;
1698         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1699
1700         unique_ptr<float[]> planar_samples(new float[audio.size()]);
1701         avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_FLTP, (const uint8_t*)planar_samples.get(), audio.size() * sizeof(float), 0);
1702         for (int i = 0; i < frame->nb_samples; ++i) {
1703             planar_samples[i] = audio[i * 2 + 0];
1704             planar_samples[i + frame->nb_samples] = audio[i * 2 + 1];
1705         }
1706
1707         AVPacket pkt;
1708         av_init_packet(&pkt);
1709         pkt.data = nullptr;
1710         pkt.size = 0;
1711         int got_output;
1712         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1713         if (got_output) {
1714             pkt.stream_index = 1;
1715             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1716         }
1717         // TODO: Delayed frames.
1718         avcodec_free_frame(&frame);
1719     }
1720
1721 #if 0
1722     printf("\r      "); /* return back to startpoint */
1723     switch (encode_order % 4) {
1724         case 0:
1725             printf("|");
1726             break;
1727         case 1:
1728             printf("/");
1729             break;
1730         case 2:
1731             printf("-");
1732             break;
1733         case 3:
1734             printf("\\");
1735             break;
1736     }
1737     printf("%08lld", encode_order);
1738     printf("(%06d bytes coded)", coded_size);
1739 #endif
1740
1741     return 0;
1742 }
1743
1744
1745 // this is weird. but it seems to put a new frame onto the queue
1746 void H264Encoder::storage_task_enqueue(storage_task task)
1747 {
1748         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1749         storage_task_queue.push(move(task));
1750         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1751         storage_task_queue_changed.notify_all();
1752 }
1753
1754 void H264Encoder::storage_task_thread()
1755 {
1756         for ( ;; ) {
1757                 storage_task current;
1758                 {
1759                         // wait until there's an encoded frame
1760                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1761                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1762                         if (storage_thread_should_quit) return;
1763                         current = move(storage_task_queue.front());
1764                         storage_task_queue.pop();
1765                 }
1766
1767                 VAStatus va_status;
1768
1769                 // waits for data, then saves it to disk.
1770                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1771                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1772                 save_codeddata(move(current));
1773
1774                 {
1775                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1776                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1777                         storage_task_queue_changed.notify_all();
1778                 }
1779         }
1780 }
1781
1782 static int release_encode()
1783 {
1784     int i;
1785
1786     for (i = 0; i < SURFACE_NUM; i++) {
1787         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1788         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1789         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1790     }
1791
1792     vaDestroyContext(va_dpy, context_id);
1793     vaDestroyConfig(va_dpy, config_id);
1794
1795     return 0;
1796 }
1797
1798 static int deinit_va()
1799 {
1800     vaTerminate(va_dpy);
1801
1802     va_close_display(va_dpy);
1803
1804     return 0;
1805 }
1806
1807
1808 static int print_input()
1809 {
1810     printf("\n\nINPUT:Try to encode H264...\n");
1811     if (rc_mode != -1)
1812         printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1813     printf("INPUT: Resolution   : %dx%dframes\n", frame_width, frame_height);
1814     printf("INPUT: FrameRate    : %d\n", frame_rate);
1815     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
1816     printf("INPUT: Slieces      : %d\n", frame_slices);
1817     printf("INPUT: IntraPeriod  : %d\n", intra_period);
1818     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
1819     printf("INPUT: IpPeriod     : %d\n", ip_period);
1820     printf("INPUT: Initial QP   : %d\n", initial_qp);
1821     printf("INPUT: Min QP       : %d\n", minimal_qp);
1822     printf("INPUT: Coded Clip   : %s\n", coded_fn);
1823
1824     printf("\n\n"); /* return back to startpoint */
1825
1826     return 0;
1827 }
1828
1829 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
1830         : current_storage_frame(0), surface(surface), httpd(httpd)
1831 {
1832         AVCodec *codec_audio = avcodec_find_encoder(AV_CODEC_ID_MP3);
1833         context_audio = avcodec_alloc_context3(codec_audio);
1834         context_audio->bit_rate = 256000;
1835         context_audio->sample_rate = 48000;
1836         context_audio->sample_fmt = AV_SAMPLE_FMT_FLTP;
1837         context_audio->channels = 2;
1838         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1839         context_audio->time_base = AVRational{1, TIMEBASE};
1840         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1841                 fprintf(stderr, "Could not open codec\n");
1842                 exit(1);
1843         }
1844
1845         frame_width = width;
1846         frame_height = height;
1847         frame_width_mbaligned = (frame_width + 15) & (~15);
1848         frame_height_mbaligned = (frame_height + 15) & (~15);
1849         frame_bitrate = 15000000;  // / 60;
1850         current_frame_encoding = 0;
1851
1852         print_input();
1853
1854         init_va();
1855         setup_encode();
1856
1857         // No frames are ready yet.
1858         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1859
1860         memset(&seq_param, 0, sizeof(seq_param));
1861         memset(&pic_param, 0, sizeof(pic_param));
1862         memset(&slice_param, 0, sizeof(slice_param));
1863
1864         storage_thread = std::thread(&H264Encoder::storage_task_thread, this);
1865
1866         copy_thread = std::thread([this]{
1867                 //SDL_GL_MakeCurrent(window, context);
1868                 QOpenGLContext *context = create_context();
1869                 eglBindAPI(EGL_OPENGL_API);
1870                 if (!make_current(context, this->surface)) {
1871                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1872                                 eglGetError());
1873                         exit(1);
1874                 }
1875                 copy_thread_func();
1876         });
1877 }
1878
1879 H264Encoder::~H264Encoder()
1880 {
1881         {
1882                 unique_lock<mutex> lock(storage_task_queue_mutex);
1883                 storage_thread_should_quit = true;
1884                 storage_task_queue_changed.notify_all();
1885         }
1886         {
1887                 unique_lock<mutex> lock(frame_queue_mutex);
1888                 copy_thread_should_quit = true;
1889                 frame_queue_nonempty.notify_one();
1890         }
1891         storage_thread.join();
1892         copy_thread.join();
1893
1894         release_encode();
1895         deinit_va();
1896 }
1897
1898 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1899 {
1900         {
1901                 // Wait until this frame slot is done encoding.
1902                 std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1903                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1904                 if (storage_thread_should_quit) return false;
1905         }
1906
1907         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1908         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1909         *y_tex = surf->y_tex;
1910         *cbcr_tex = surf->cbcr_tex;
1911
1912         VASurfaceID surface = surf->src_surface;
1913         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1914         CHECK_VASTATUS(va_status, "vaDeriveImage");
1915
1916         VABufferInfo buf_info;
1917         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1918         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1919         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1920
1921         // Create Y image.
1922         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1923         EGLint y_attribs[] = {
1924                 EGL_WIDTH, frame_width,
1925                 EGL_HEIGHT, frame_height,
1926                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1927                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1928                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1929                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1930                 EGL_NONE
1931         };
1932
1933         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1934         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1935
1936         // Associate Y image to a texture.
1937         glBindTexture(GL_TEXTURE_2D, *y_tex);
1938         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1939
1940         // Create CbCr image.
1941         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1942         EGLint cbcr_attribs[] = {
1943                 EGL_WIDTH, frame_width,
1944                 EGL_HEIGHT, frame_height,
1945                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1946                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1947                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1948                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1949                 EGL_NONE
1950         };
1951
1952         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1953         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1954
1955         // Associate CbCr image to a texture.
1956         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1957         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1958
1959         return true;
1960 }
1961
1962 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
1963 {
1964         {
1965                 unique_lock<mutex> lock(frame_queue_mutex);
1966                 pending_audio_frames[pts] = move(audio);
1967         }
1968         frame_queue_nonempty.notify_one();
1969 }
1970
1971
1972 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
1973 {
1974         {
1975                 unique_lock<mutex> lock(frame_queue_mutex);
1976                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1977                 ++current_storage_frame;
1978         }
1979         frame_queue_nonempty.notify_one();
1980 }
1981
1982 void H264Encoder::copy_thread_func()
1983 {
1984         int64_t last_dts = -1;
1985         for ( ;; ) {
1986                 PendingFrame frame;
1987                 int pts_lag;
1988                 encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
1989                                        &current_frame_display, &current_frame_type, &pts_lag);
1990                 if (current_frame_type == FRAME_IDR) {
1991                         numShortTerm = 0;
1992                         current_frame_num = 0;
1993                         current_IDR_display = current_frame_display;
1994                 }
1995
1996                 {
1997                         unique_lock<mutex> lock(frame_queue_mutex);
1998                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; });
1999                         if (copy_thread_should_quit) return;
2000                         frame = move(pending_video_frames[current_frame_display]);
2001                         pending_video_frames.erase(current_frame_display);
2002                 }
2003
2004                 // Wait for the GPU to be done with the frame.
2005                 glClientWaitSync(frame.fence.get(), 0, 0);
2006
2007                 // Release back any input frames we needed to render this frame.
2008                 frame.input_frames.clear();
2009
2010                 // Unmap the image.
2011                 GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
2012                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2013                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2014                 VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2015                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2016                 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2017                 CHECK_VASTATUS(va_status, "vaDestroyImage");
2018
2019                 VASurfaceID surface = surf->src_surface;
2020
2021                 // Schedule the frame for encoding.
2022                 va_status = vaBeginPicture(va_dpy, context_id, surface);
2023                 CHECK_VASTATUS(va_status, "vaBeginPicture");
2024
2025                 if (current_frame_type == FRAME_IDR) {
2026                         render_sequence();
2027                         render_picture();
2028                         if (h264_packedheader) {
2029                                 render_packedsequence();
2030                                 render_packedpicture();
2031                         }
2032                 } else {
2033                         //render_sequence();
2034                         render_picture();
2035                 }
2036                 render_slice();
2037
2038                 va_status = vaEndPicture(va_dpy, context_id);
2039                 CHECK_VASTATUS(va_status, "vaEndPicture");
2040
2041                 // Determine the pts and dts of this frame.
2042                 int64_t pts = frame.pts;
2043                 int64_t dts;
2044                 if (pts_lag == -1) {
2045                         assert(last_dts != -1);
2046                         dts = last_dts + (TIMEBASE / frame_rate);
2047                 } else {
2048                         dts = pts - pts_lag;
2049                 }
2050                 last_dts = dts;
2051
2052                 // so now the data is done encoding (well, async job kicked off)...
2053                 // we send that to the storage thread
2054                 storage_task tmp;
2055                 tmp.display_order = current_frame_display;
2056                 tmp.frame_type = current_frame_type;
2057                 tmp.pts = pts;
2058                 tmp.dts = dts;
2059                 storage_task_enqueue(move(tmp));
2060
2061                 update_ReferenceFrames();
2062                 ++current_frame_encoding;
2063         }
2064 }