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  char *coded_fn = NULL;
 112
 113 static  int frame_width = 176;
 114 static  int frame_height = 144;
 115 static  int frame_width_mbaligned;
 116 static  int frame_height_mbaligned;
 117 static  int frame_rate = FPS;
 118 static  unsigned int frame_bitrate = 0;
 119 static  unsigned int frame_slices = 1;
 120 static  double frame_size = 0;
 121 static  int initial_qp = 15;
 122 //static  int initial_qp = 28;
 123 static  int minimal_qp = 0;
 124 static  int intra_period = 30;
 125 static  int intra_idr_period = FPS;
 126 static  int ip_period = 3;
 127 static  int rc_mode = -1;
 128 static  int rc_default_modes[] = {
 129     VA_RC_VBR,
 130     VA_RC_CQP,
 131     VA_RC_VBR_CONSTRAINED,
 132     VA_RC_CBR,
 133     VA_RC_VCM,
 134     VA_RC_NONE,
 135 };
 136 static  unsigned long long current_frame_encoding = 0;
 137 static  unsigned long long current_frame_display = 0;
 138 static  unsigned long long current_IDR_display = 0;
 139 static  unsigned int current_frame_num = 0;
 140 static  int current_frame_type;
 141
 142 static  int misc_priv_type = 0;
 143 static  int misc_priv_value = 0;
 144
 145 /* thread to save coded data */
 146 #define SRC_SURFACE_FREE        0
 147 #define SRC_SURFACE_IN_ENCODING 1
 148
 149 struct __bitstream {
 150     unsigned int *buffer;
 151     int bit_offset;
 152     int max_size_in_dword;
 153 };
 154 typedef struct __bitstream bitstream;
 155
 156 using namespace std;
 157
 158 static unsigned int
 159 va_swap32(unsigned int val)
 160 {
 161     unsigned char *pval = (unsigned char *)&val;
 162
 163     return ((pval[0] << 24)     |
 164             (pval[1] << 16)     |
 165             (pval[2] << 8)      |
 166             (pval[3] << 0));
 167 }
 168
 169 static void
 170 bitstream_start(bitstream *bs)
 171 {
 172     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 173     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 174     bs->bit_offset = 0;
 175 }
 176
 177 static void
 178 bitstream_end(bitstream *bs)
 179 {
 180     int pos = (bs->bit_offset >> 5);
 181     int bit_offset = (bs->bit_offset & 0x1f);
 182     int bit_left = 32 - bit_offset;
 183
 184     if (bit_offset) {
 185         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 186     }
 187 }
 188
 189 static void
 190 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 191 {
 192     int pos = (bs->bit_offset >> 5);
 193     int bit_offset = (bs->bit_offset & 0x1f);
 194     int bit_left = 32 - bit_offset;
 195
 196     if (!size_in_bits)
 197         return;
 198
 199     bs->bit_offset += size_in_bits;
 200
 201     if (bit_left > size_in_bits) {
 202         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 203     } else {
 204         size_in_bits -= bit_left;
 205         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 206         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 207
 208         if (pos + 1 == bs->max_size_in_dword) {
 209             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 210             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 211         }
 212
 213         bs->buffer[pos + 1] = val;
 214     }
 215 }
 216
 217 static void
 218 bitstream_put_ue(bitstream *bs, unsigned int val)
 219 {
 220     int size_in_bits = 0;
 221     int tmp_val = ++val;
 222
 223     while (tmp_val) {
 224         tmp_val >>= 1;
 225         size_in_bits++;
 226     }
 227
 228     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 229     bitstream_put_ui(bs, val, size_in_bits);
 230 }
 231
 232 static void
 233 bitstream_put_se(bitstream *bs, int val)
 234 {
 235     unsigned int new_val;
 236
 237     if (val <= 0)
 238         new_val = -2 * val;
 239     else
 240         new_val = 2 * val - 1;
 241
 242     bitstream_put_ue(bs, new_val);
 243 }
 244
 245 static void
 246 bitstream_byte_aligning(bitstream *bs, int bit)
 247 {
 248     int bit_offset = (bs->bit_offset & 0x7);
 249     int bit_left = 8 - bit_offset;
 250     int new_val;
 251
 252     if (!bit_offset)
 253         return;
 254
 255     assert(bit == 0 || bit == 1);
 256
 257     if (bit)
 258         new_val = (1 << bit_left) - 1;
 259     else
 260         new_val = 0;
 261
 262     bitstream_put_ui(bs, new_val, bit_left);
 263 }
 264
 265 static void
 266 rbsp_trailing_bits(bitstream *bs)
 267 {
 268     bitstream_put_ui(bs, 1, 1);
 269     bitstream_byte_aligning(bs, 0);
 270 }
 271
 272 static void nal_start_code_prefix(bitstream *bs)
 273 {
 274     bitstream_put_ui(bs, 0x00000001, 32);
 275 }
 276
 277 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 278 {
 279     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 280     bitstream_put_ui(bs, nal_ref_idc, 2);
 281     bitstream_put_ui(bs, nal_unit_type, 5);
 282 }
 283
 284 static void sps_rbsp(bitstream *bs)
 285 {
 286     int profile_idc = PROFILE_IDC_BASELINE;
 287
 288     if (h264_profile  == VAProfileH264High)
 289         profile_idc = PROFILE_IDC_HIGH;
 290     else if (h264_profile  == VAProfileH264Main)
 291         profile_idc = PROFILE_IDC_MAIN;
 292
 293     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 294     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 295     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 296     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 297     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 298     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 299     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 300     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 301
 302     if ( profile_idc == PROFILE_IDC_HIGH) {
 303         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 304         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 305         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 306         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 307         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 308     }
 309
 310     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 311     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 312
 313     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 314         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 315     else {
 316         assert(0);
 317     }
 318
 319     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 320     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 321
 322     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 323     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 324     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 325
 326     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 327         assert(0);
 328     }
 329
 330     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 331     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 332
 333     if (seq_param.frame_cropping_flag) {
 334         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 335         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 336         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 337         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 338     }
 339
 340     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 341     if ( false ) {
 342         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 343     } else {
 344         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 345         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 346         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 347         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 348         {
 349             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 350             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 351             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 352             {
 353                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 354                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 355                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 356             }
 357         }
 358         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 359         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 360         {
 361             bitstream_put_ui(bs, 1, 32);  // FPS
 362             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 363             bitstream_put_ui(bs, 1, 1);
 364         }
 365         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 366         {
 367             // hrd_parameters
 368             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 369             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 370             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 371
 372             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 373             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 374             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 375
 376             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 377             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 378             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 379             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 380         }
 381         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 382         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 383
 384         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 385         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 386     }
 387
 388     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 389 }
 390
 391
 392 static void pps_rbsp(bitstream *bs)
 393 {
 394     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 395     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 396
 397     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 398
 399     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 400
 401     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 402
 403     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 404     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 405
 406     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 407     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 408
 409     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 410     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 411     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 412
 413     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 414     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 415     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 416
 417     /* more_rbsp_data */
 418     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 419     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 420     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 421
 422     rbsp_trailing_bits(bs);
 423 }
 424
 425 static void slice_header(bitstream *bs)
 426 {
 427     int first_mb_in_slice = slice_param.macroblock_address;
 428
 429     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 430     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 431     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 432     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 433
 434     /* frame_mbs_only_flag == 1 */
 435     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 436         /* FIXME: */
 437         assert(0);
 438     }
 439
 440     if (pic_param.pic_fields.bits.idr_pic_flag)
 441         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 442
 443     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 444         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 445         /* pic_order_present_flag == 0 */
 446     } else {
 447         /* FIXME: */
 448         assert(0);
 449     }
 450
 451     /* redundant_pic_cnt_present_flag == 0 */
 452     /* slice type */
 453     if (IS_P_SLICE(slice_param.slice_type)) {
 454         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 455
 456         if (slice_param.num_ref_idx_active_override_flag)
 457             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 458
 459         /* ref_pic_list_reordering */
 460         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 461     } else if (IS_B_SLICE(slice_param.slice_type)) {
 462         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 463
 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             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 469         }
 470
 471         /* ref_pic_list_reordering */
 472         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 473         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 474     }
 475
 476     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 477          IS_P_SLICE(slice_param.slice_type)) ||
 478         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 479          IS_B_SLICE(slice_param.slice_type))) {
 480         /* FIXME: fill weight/offset table */
 481         assert(0);
 482     }
 483
 484     /* dec_ref_pic_marking */
 485     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 486         unsigned char no_output_of_prior_pics_flag = 0;
 487         unsigned char long_term_reference_flag = 0;
 488         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 489
 490         if (pic_param.pic_fields.bits.idr_pic_flag) {
 491             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 492             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 493         } else {
 494             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 495         }
 496     }
 497
 498     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 499         !IS_I_SLICE(slice_param.slice_type))
 500         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 501
 502     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 503
 504     /* ignore for SP/SI */
 505
 506     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 507         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 508
 509         if (slice_param.disable_deblocking_filter_idc != 1) {
 510             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 511             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 512         }
 513     }
 514
 515     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 516         bitstream_byte_aligning(bs, 1);
 517     }
 518 }
 519
 520 static int
 521 build_packed_pic_buffer(unsigned char **header_buffer)
 522 {
 523     bitstream bs;
 524
 525     bitstream_start(&bs);
 526     nal_start_code_prefix(&bs);
 527     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 528     pps_rbsp(&bs);
 529     bitstream_end(&bs);
 530
 531     *header_buffer = (unsigned char *)bs.buffer;
 532     return bs.bit_offset;
 533 }
 534
 535 static int
 536 build_packed_seq_buffer(unsigned char **header_buffer)
 537 {
 538     bitstream bs;
 539
 540     bitstream_start(&bs);
 541     nal_start_code_prefix(&bs);
 542     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 543     sps_rbsp(&bs);
 544     bitstream_end(&bs);
 545
 546     *header_buffer = (unsigned char *)bs.buffer;
 547     return bs.bit_offset;
 548 }
 549
 550 static int build_packed_slice_buffer(unsigned char **header_buffer)
 551 {
 552     bitstream bs;
 553     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 554     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 555
 556     bitstream_start(&bs);
 557     nal_start_code_prefix(&bs);
 558
 559     if (IS_I_SLICE(slice_param.slice_type)) {
 560         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 561     } else if (IS_P_SLICE(slice_param.slice_type)) {
 562         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 563     } else {
 564         assert(IS_B_SLICE(slice_param.slice_type));
 565         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 566     }
 567
 568     slice_header(&bs);
 569     bitstream_end(&bs);
 570
 571     *header_buffer = (unsigned char *)bs.buffer;
 572     return bs.bit_offset;
 573 }
 574
 575
 576 /*
 577   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 578   1) period between Frame #X and Frame #N = #X - #N
 579   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 580   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 581   4) intra_period and intra_idr_period take precedence over ip_period
 582   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 583      of I/IDR frames, see bellow examples
 584   -------------------------------------------------------------------
 585   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 586   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 587   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 588   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 589   1            1                ignored    IDR IDR IDR IDR...
 590   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 591   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 592   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 593                                               (PBB)(IBB)(PBB)(IBB)...
 594   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 595                                            IDRPPPPP IPPPPP IPPPPP...
 596   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 597                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 598                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 599                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 600                                            {IDR(PBB)(PBB)}                     (6/6/3)
 601                                            {IDR(PBB)(PBB)}.
 602 */
 603
 604 // General pts/dts strategy:
 605 //
 606 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 607 // bit tricky. We assume first of all that the frame rate never goes _above_
 608 // <frame_rate>, which gives us a frame period N. The decoder can always decode
 609 // in at least this speed, as long at dts <= pts (the frame is not attempted
 610 // presented before it is decoded). Furthermore, we never have longer chains of
 611 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 612 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 613 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 614 // frame, with an order O <= C.)
 615 //
 616 // Many strategies are possible, but we establish these rules:
 617 //
 618 //  - Tip frames have dts = pts - (C-O)*N.
 619 //  - Non-tip frames have dts = dts_last + N.
 620 //
 621 // An example, with C=2 and N=10 and the data flow showed with arrows:
 622 //
 623 //        I  B  P  B  B  P
 624 //   pts: 30 40 50 60 70 80
 625 //        ↓  ↓     ↓
 626 //   dts: 10 30 20 60 50←40
 627 //         |  |  ↑        ↑
 628 //         `--|--'        |
 629 //             `----------'
 630 //
 631 // To show that this works fine also with irregular spacings, let's say that
 632 // the third frame is delayed a bit (something earlier was dropped). Now the
 633 // situation looks like this:
 634 //
 635 //        I  B  P  B  B   P
 636 //   pts: 30 40 80 90 100 110
 637 //        ↓  ↓     ↓
 638 //   dts: 10 30 20 90 50←40
 639 //         |  |  ↑        ↑
 640 //         `--|--'        |
 641 //             `----------'
 642 //
 643 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 644 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 645 //
 646 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 647 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 648 // a tip frame and should be given a dts based on the previous one.
 649 #define FRAME_P 0
 650 #define FRAME_B 1
 651 #define FRAME_I 2
 652 #define FRAME_IDR 7
 653 void encoding2display_order(
 654     unsigned long long encoding_order, int intra_period,
 655     int intra_idr_period, int ip_period,
 656     unsigned long long *displaying_order,
 657     int *frame_type, int *pts_lag)
 658 {
 659     int encoding_order_gop = 0;
 660
 661     *pts_lag = 0;
 662
 663     if (intra_period == 1) { /* all are I/IDR frames */
 664         *displaying_order = encoding_order;
 665         if (intra_idr_period == 0)
 666             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 667         else
 668             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 669         return;
 670     }
 671
 672     if (intra_period == 0)
 673         intra_idr_period = 0;
 674
 675     if (ip_period == 1) {
 676         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 677         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 678         *displaying_order = encoding_order;
 679
 680         if (encoding_order_gop == 0) { /* the first frame */
 681             *frame_type = FRAME_IDR;
 682         } else if (intra_period != 0 && /* have I frames */
 683                    encoding_order_gop >= 2 &&
 684                    (encoding_order_gop % intra_period == 0)) {
 685             *frame_type = FRAME_I;
 686         } else {
 687             *frame_type = FRAME_P;
 688         }
 689         return;
 690     }
 691
 692     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 693     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 694     *pts_lag = -1;  // Most frames are not tip frames.
 695
 696     if (encoding_order_gop == 0) { /* the first frame */
 697         *frame_type = FRAME_IDR;
 698         *displaying_order = encoding_order;
 699         // IDR frames are a special case; I honestly can't find the logic behind
 700         // why this is the right thing, but it seems to line up nicely in practice :-)
 701         *pts_lag = TIMEBASE / frame_rate;
 702     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 703         *frame_type = FRAME_B;
 704         *displaying_order = encoding_order - 1;
 705         if ((encoding_order_gop % ip_period) == 0) {
 706             *pts_lag = 0;  // Last B-frame.
 707         }
 708     } else if (intra_period != 0 && /* have I frames */
 709                encoding_order_gop >= 2 &&
 710                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 711         *frame_type = FRAME_I;
 712         *displaying_order = encoding_order + ip_period - 1;
 713     } else {
 714         *frame_type = FRAME_P;
 715         *displaying_order = encoding_order + ip_period - 1;
 716     }
 717 }
 718
 719
 720 static const char *rc_to_string(int rcmode)
 721 {
 722     switch (rc_mode) {
 723     case VA_RC_NONE:
 724         return "NONE";
 725     case VA_RC_CBR:
 726         return "CBR";
 727     case VA_RC_VBR:
 728         return "VBR";
 729     case VA_RC_VCM:
 730         return "VCM";
 731     case VA_RC_CQP:
 732         return "CQP";
 733     case VA_RC_VBR_CONSTRAINED:
 734         return "VBR_CONSTRAINED";
 735     default:
 736         return "Unknown";
 737     }
 738 }
 739
 740 #if 0
 741 static int process_cmdline(int argc, char *argv[])
 742 {
 743     char c;
 744     const struct option long_opts[] = {
 745         {"help", no_argument, NULL, 0 },
 746         {"bitrate", required_argument, NULL, 1 },
 747         {"minqp", required_argument, NULL, 2 },
 748         {"initialqp", required_argument, NULL, 3 },
 749         {"intra_period", required_argument, NULL, 4 },
 750         {"idr_period", required_argument, NULL, 5 },
 751         {"ip_period", required_argument, NULL, 6 },
 752         {"rcmode", required_argument, NULL, 7 },
 753         {"srcyuv", required_argument, NULL, 9 },
 754         {"recyuv", required_argument, NULL, 10 },
 755         {"fourcc", required_argument, NULL, 11 },
 756         {"syncmode", no_argument, NULL, 12 },
 757         {"enablePSNR", no_argument, NULL, 13 },
 758         {"prit", required_argument, NULL, 14 },
 759         {"priv", required_argument, NULL, 15 },
 760         {"framecount", required_argument, NULL, 16 },
 761         {"entropy", required_argument, NULL, 17 },
 762         {"profile", required_argument, NULL, 18 },
 763         {NULL, no_argument, NULL, 0 }};
 764     int long_index;
 765
 766     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 767         switch (c) {
 768         case 'w':
 769             frame_width = atoi(optarg);
 770             break;
 771         case 'h':
 772             frame_height = atoi(optarg);
 773             break;
 774         case 'n':
 775         case 'f':
 776             frame_rate = atoi(optarg);
 777             break;
 778         case 'o':
 779             coded_fn = strdup(optarg);
 780             break;
 781         case 0:
 782             print_help();
 783             exit(0);
 784         case 1:
 785             frame_bitrate = atoi(optarg);
 786             break;
 787         case 2:
 788             minimal_qp = atoi(optarg);
 789             break;
 790         case 3:
 791             initial_qp = atoi(optarg);
 792             break;
 793         case 4:
 794             intra_period = atoi(optarg);
 795             break;
 796         case 5:
 797             intra_idr_period = atoi(optarg);
 798             break;
 799         case 6:
 800             ip_period = atoi(optarg);
 801             break;
 802         case 7:
 803             rc_mode = string_to_rc(optarg);
 804             if (rc_mode < 0) {
 805                 print_help();
 806                 exit(1);
 807             }
 808             break;
 809         case 9:
 810             srcyuv_fn = strdup(optarg);
 811             break;
 812         case 11:
 813             srcyuv_fourcc = string_to_fourcc(optarg);
 814             if (srcyuv_fourcc <= 0) {
 815                 print_help();
 816                 exit(1);
 817             }
 818             break;
 819         case 13:
 820             calc_psnr = 1;
 821             break;
 822         case 14:
 823             misc_priv_type = strtol(optarg, NULL, 0);
 824             break;
 825         case 15:
 826             misc_priv_value = strtol(optarg, NULL, 0);
 827             break;
 828         case 17:
 829             h264_entropy_mode = atoi(optarg) ? 1: 0;
 830             break;
 831         case 18:
 832             if (strncmp(optarg, "BP", 2) == 0)
 833                 h264_profile = VAProfileH264Baseline;
 834             else if (strncmp(optarg, "MP", 2) == 0)
 835                 h264_profile = VAProfileH264Main;
 836             else if (strncmp(optarg, "HP", 2) == 0)
 837                 h264_profile = VAProfileH264High;
 838             else
 839                 h264_profile = (VAProfile)0;
 840             break;
 841         case ':':
 842         case '?':
 843             print_help();
 844             exit(0);
 845         }
 846     }
 847
 848     if (ip_period < 1) {
 849         printf(" ip_period must be greater than 0\n");
 850         exit(0);
 851     }
 852     if (intra_period != 1 && intra_period % ip_period != 0) {
 853         printf(" intra_period must be a multiplier of ip_period\n");
 854         exit(0);
 855     }
 856     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 857         printf(" intra_idr_period must be a multiplier of intra_period\n");
 858         exit(0);
 859     }
 860
 861     if (frame_bitrate == 0)
 862         frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
 863
 864     if (coded_fn == NULL) {
 865         struct stat buf;
 866         if (stat("/tmp", &buf) == 0)
 867             coded_fn = strdup("/tmp/test.264");
 868         else if (stat("/sdcard", &buf) == 0)
 869             coded_fn = strdup("/sdcard/test.264");
 870         else
 871             coded_fn = strdup("./test.264");
 872     }
 873
 874
 875     frame_width_mbaligned = (frame_width + 15) & (~15);
 876     frame_height_mbaligned = (frame_height + 15) & (~15);
 877     if (frame_width != frame_width_mbaligned ||
 878         frame_height != frame_height_mbaligned) {
 879         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 880                frame_width, frame_height,
 881                frame_width_mbaligned, frame_height_mbaligned
 882                );
 883     }
 884
 885     return 0;
 886 }
 887 #endif
 888
 889 static Display *x11_display;
 890 static Window   x11_window;
 891
 892 VADisplay
 893 va_open_display(void)
 894 {
 895     x11_display = XOpenDisplay(NULL);
 896     if (!x11_display) {
 897         fprintf(stderr, "error: can't connect to X server!\n");
 898         return NULL;
 899     }
 900     return vaGetDisplay(x11_display);
 901 }
 902
 903 void
 904 va_close_display(VADisplay va_dpy)
 905 {
 906     if (!x11_display)
 907         return;
 908
 909     if (x11_window) {
 910         XUnmapWindow(x11_display, x11_window);
 911         XDestroyWindow(x11_display, x11_window);
 912         x11_window = None;
 913     }
 914     XCloseDisplay(x11_display);
 915     x11_display = NULL;
 916 }
 917
 918 static int init_va(void)
 919 {
 920     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 921     VAEntrypoint *entrypoints;
 922     int num_entrypoints, slice_entrypoint;
 923     int support_encode = 0;
 924     int major_ver, minor_ver;
 925     VAStatus va_status;
 926     unsigned int i;
 927
 928     va_dpy = va_open_display();
 929     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 930     CHECK_VASTATUS(va_status, "vaInitialize");
 931
 932     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 933     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 934     if (!entrypoints) {
 935         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 936         exit(1);
 937     }
 938
 939     /* use the highest profile */
 940     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 941         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 942             continue;
 943
 944         h264_profile = profile_list[i];
 945         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 946         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 947             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 948                 support_encode = 1;
 949                 break;
 950             }
 951         }
 952         if (support_encode == 1)
 953             break;
 954     }
 955
 956     if (support_encode == 0) {
 957         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 958         exit(1);
 959     } else {
 960         switch (h264_profile) {
 961             case VAProfileH264Baseline:
 962                 printf("Use profile VAProfileH264Baseline\n");
 963                 ip_period = 1;
 964                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 965                 h264_entropy_mode = 0;
 966                 break;
 967             case VAProfileH264ConstrainedBaseline:
 968                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 969                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 970                 ip_period = 1;
 971                 break;
 972
 973             case VAProfileH264Main:
 974                 printf("Use profile VAProfileH264Main\n");
 975                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 976                 break;
 977
 978             case VAProfileH264High:
 979                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 980                 printf("Use profile VAProfileH264High\n");
 981                 break;
 982             default:
 983                 printf("unknow profile. Set to Baseline");
 984                 h264_profile = VAProfileH264Baseline;
 985                 ip_period = 1;
 986                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 987                 break;
 988         }
 989     }
 990
 991     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 992
 993     /* find out the format for the render target, and rate control mode */
 994     for (i = 0; i < VAConfigAttribTypeMax; i++)
 995         attrib[i].type = (VAConfigAttribType)i;
 996
 997     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 998                                       &attrib[0], VAConfigAttribTypeMax);
 999     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1000     /* check the interested configattrib */
1001     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1002         printf("Not find desired YUV420 RT format\n");
1003         exit(1);
1004     } else {
1005         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1006         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1007         config_attrib_num++;
1008     }
1009
1010     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1011         int tmp = attrib[VAConfigAttribRateControl].value;
1012
1013         printf("Support rate control mode (0x%x):", tmp);
1014
1015         if (tmp & VA_RC_NONE)
1016             printf("NONE ");
1017         if (tmp & VA_RC_CBR)
1018             printf("CBR ");
1019         if (tmp & VA_RC_VBR)
1020             printf("VBR ");
1021         if (tmp & VA_RC_VCM)
1022             printf("VCM ");
1023         if (tmp & VA_RC_CQP)
1024             printf("CQP ");
1025         if (tmp & VA_RC_VBR_CONSTRAINED)
1026             printf("VBR_CONSTRAINED ");
1027
1028         printf("\n");
1029
1030         if (rc_mode == -1 || !(rc_mode & tmp))  {
1031             if (rc_mode != -1) {
1032                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1033             }
1034
1035             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1036                 if (rc_default_modes[i] & tmp) {
1037                     rc_mode = rc_default_modes[i];
1038                     break;
1039                 }
1040             }
1041
1042             printf("RateControl mode: %s\n", rc_to_string(rc_mode));
1043         }
1044
1045         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1046         config_attrib[config_attrib_num].value = rc_mode;
1047         config_attrib_num++;
1048     }
1049
1050
1051     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1052         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1053
1054         printf("Support VAConfigAttribEncPackedHeaders\n");
1055
1056         h264_packedheader = 1;
1057         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1058         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1059
1060         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1061             printf("Support packed sequence headers\n");
1062             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1063         }
1064
1065         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1066             printf("Support packed picture headers\n");
1067             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1068         }
1069
1070         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1071             printf("Support packed slice headers\n");
1072             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1073         }
1074
1075         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1076             printf("Support packed misc headers\n");
1077             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1078         }
1079
1080         enc_packed_header_idx = config_attrib_num;
1081         config_attrib_num++;
1082     }
1083
1084     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1085         int tmp = attrib[VAConfigAttribEncInterlaced].value;
1086
1087         printf("Support VAConfigAttribEncInterlaced\n");
1088
1089         if (tmp & VA_ENC_INTERLACED_FRAME)
1090             printf("support VA_ENC_INTERLACED_FRAME\n");
1091         if (tmp & VA_ENC_INTERLACED_FIELD)
1092             printf("Support VA_ENC_INTERLACED_FIELD\n");
1093         if (tmp & VA_ENC_INTERLACED_MBAFF)
1094             printf("Support VA_ENC_INTERLACED_MBAFF\n");
1095         if (tmp & VA_ENC_INTERLACED_PAFF)
1096             printf("Support VA_ENC_INTERLACED_PAFF\n");
1097
1098         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1099         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1100         config_attrib_num++;
1101     }
1102
1103     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1104         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1105
1106         printf("Support %d RefPicList0 and %d RefPicList1\n",
1107                h264_maxref & 0xffff, (h264_maxref >> 16) & 0xffff );
1108     }
1109
1110     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1111         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1112
1113     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1114         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1115
1116         printf("Support VAConfigAttribEncSliceStructure\n");
1117
1118         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1119             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1120         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1121             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1122         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1123             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1124     }
1125     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1126         printf("Support VAConfigAttribEncMacroblockInfo\n");
1127     }
1128
1129     free(entrypoints);
1130     return 0;
1131 }
1132
1133 static int setup_encode()
1134 {
1135     VAStatus va_status;
1136     VASurfaceID *tmp_surfaceid;
1137     int codedbuf_size, i;
1138     static VASurfaceID src_surface[SURFACE_NUM];
1139     static VASurfaceID ref_surface[SURFACE_NUM];
1140
1141     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1142             &config_attrib[0], config_attrib_num, &config_id);
1143     CHECK_VASTATUS(va_status, "vaCreateConfig");
1144
1145     /* create source surfaces */
1146     va_status = vaCreateSurfaces(va_dpy,
1147                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1148                                  &src_surface[0], SURFACE_NUM,
1149                                  NULL, 0);
1150     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1151
1152     /* create reference surfaces */
1153     va_status = vaCreateSurfaces(va_dpy,
1154                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1155                                  &ref_surface[0], SURFACE_NUM,
1156                                  NULL, 0);
1157     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1158
1159     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1160     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1161     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1162
1163     /* Create a context for this encode pipe */
1164     va_status = vaCreateContext(va_dpy, config_id,
1165                                 frame_width_mbaligned, frame_height_mbaligned,
1166                                 VA_PROGRESSIVE,
1167                                 tmp_surfaceid, 2 * SURFACE_NUM,
1168                                 &context_id);
1169     CHECK_VASTATUS(va_status, "vaCreateContext");
1170     free(tmp_surfaceid);
1171
1172     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1173
1174     for (i = 0; i < SURFACE_NUM; i++) {
1175         /* create coded buffer once for all
1176          * other VA buffers which won't be used again after vaRenderPicture.
1177          * so APP can always vaCreateBuffer for every frame
1178          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1179          * so VA won't maintain the coded buffer
1180          */
1181         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1182                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1183         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1184     }
1185
1186     /* create OpenGL objects */
1187     //glGenFramebuffers(SURFACE_NUM, fbos);
1188
1189     for (i = 0; i < SURFACE_NUM; i++) {
1190         glGenTextures(1, &gl_surfaces[i].y_tex);
1191         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1192     }
1193
1194     for (i = 0; i < SURFACE_NUM; i++) {
1195         gl_surfaces[i].src_surface = src_surface[i];
1196         gl_surfaces[i].ref_surface = ref_surface[i];
1197     }
1198
1199     return 0;
1200 }
1201
1202
1203
1204 #define partition(ref, field, key, ascending)   \
1205     while (i <= j) {                            \
1206         if (ascending) {                        \
1207             while (ref[i].field < key)          \
1208                 i++;                            \
1209             while (ref[j].field > key)          \
1210                 j--;                            \
1211         } else {                                \
1212             while (ref[i].field > key)          \
1213                 i++;                            \
1214             while (ref[j].field < key)          \
1215                 j--;                            \
1216         }                                       \
1217         if (i <= j) {                           \
1218             tmp = ref[i];                       \
1219             ref[i] = ref[j];                    \
1220             ref[j] = tmp;                       \
1221             i++;                                \
1222             j--;                                \
1223         }                                       \
1224     }                                           \
1225
1226 static void sort_one(VAPictureH264 ref[], int left, int right,
1227                      int ascending, int frame_idx)
1228 {
1229     int i = left, j = right;
1230     unsigned int key;
1231     VAPictureH264 tmp;
1232
1233     if (frame_idx) {
1234         key = ref[(left + right) / 2].frame_idx;
1235         partition(ref, frame_idx, key, ascending);
1236     } else {
1237         key = ref[(left + right) / 2].TopFieldOrderCnt;
1238         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1239     }
1240
1241     /* recursion */
1242     if (left < j)
1243         sort_one(ref, left, j, ascending, frame_idx);
1244
1245     if (i < right)
1246         sort_one(ref, i, right, ascending, frame_idx);
1247 }
1248
1249 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1250                      int partition_ascending, int list0_ascending, int list1_ascending)
1251 {
1252     int i = left, j = right;
1253     VAPictureH264 tmp;
1254
1255     if (frame_idx) {
1256         partition(ref, frame_idx, key, partition_ascending);
1257     } else {
1258         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1259     }
1260
1261
1262     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1263     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1264 }
1265
1266 static int update_ReferenceFrames(void)
1267 {
1268     int i;
1269
1270     if (current_frame_type == FRAME_B)
1271         return 0;
1272
1273     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1274     numShortTerm++;
1275     if (numShortTerm > num_ref_frames)
1276         numShortTerm = num_ref_frames;
1277     for (i=numShortTerm-1; i>0; i--)
1278         ReferenceFrames[i] = ReferenceFrames[i-1];
1279     ReferenceFrames[0] = CurrentCurrPic;
1280
1281     if (current_frame_type != FRAME_B)
1282         current_frame_num++;
1283     if (current_frame_num > MaxFrameNum)
1284         current_frame_num = 0;
1285
1286     return 0;
1287 }
1288
1289
1290 static int update_RefPicList(void)
1291 {
1292     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1293
1294     if (current_frame_type == FRAME_P) {
1295         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1296         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1297     }
1298
1299     if (current_frame_type == FRAME_B) {
1300         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1301         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1302                  1, 0, 1);
1303
1304         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1305         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1306                  0, 1, 0);
1307     }
1308
1309     return 0;
1310 }
1311
1312
1313 static int render_sequence(void)
1314 {
1315     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1316     VAStatus va_status;
1317     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1318     VAEncMiscParameterRateControl *misc_rate_ctrl;
1319
1320     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1321     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1322     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1323     seq_param.bits_per_second = frame_bitrate;
1324
1325     seq_param.intra_period = intra_period;
1326     seq_param.intra_idr_period = intra_idr_period;
1327     seq_param.ip_period = ip_period;
1328
1329     seq_param.max_num_ref_frames = num_ref_frames;
1330     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1331     seq_param.time_scale = TIMEBASE * 2;
1332     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1333     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1334     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1335     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1336     seq_param.seq_fields.bits.chroma_format_idc = 1;
1337     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1338
1339     if (frame_width != frame_width_mbaligned ||
1340         frame_height != frame_height_mbaligned) {
1341         seq_param.frame_cropping_flag = 1;
1342         seq_param.frame_crop_left_offset = 0;
1343         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1344         seq_param.frame_crop_top_offset = 0;
1345         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1346     }
1347
1348     va_status = vaCreateBuffer(va_dpy, context_id,
1349                                VAEncSequenceParameterBufferType,
1350                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1351     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1352
1353     va_status = vaCreateBuffer(va_dpy, context_id,
1354                                VAEncMiscParameterBufferType,
1355                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1356                                1, NULL, &rc_param_buf);
1357     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1358
1359     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1360     misc_param->type = VAEncMiscParameterTypeRateControl;
1361     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1362     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1363     misc_rate_ctrl->bits_per_second = frame_bitrate;
1364     misc_rate_ctrl->target_percentage = 66;
1365     misc_rate_ctrl->window_size = 1000;
1366     misc_rate_ctrl->initial_qp = initial_qp;
1367     misc_rate_ctrl->min_qp = minimal_qp;
1368     misc_rate_ctrl->basic_unit_size = 0;
1369     vaUnmapBuffer(va_dpy, rc_param_buf);
1370
1371     render_id[0] = seq_param_buf;
1372     render_id[1] = rc_param_buf;
1373
1374     va_status = vaRenderPicture(va_dpy, context_id, &render_id[0], 2);
1375     CHECK_VASTATUS(va_status, "vaRenderPicture");;
1376
1377     if (misc_priv_type != 0) {
1378         va_status = vaCreateBuffer(va_dpy, context_id,
1379                                    VAEncMiscParameterBufferType,
1380                                    sizeof(VAEncMiscParameterBuffer),
1381                                    1, NULL, &misc_param_tmpbuf);
1382         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1383         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1384         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1385         misc_param_tmp->data[0] = misc_priv_value;
1386         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1387
1388         va_status = vaRenderPicture(va_dpy, context_id, &misc_param_tmpbuf, 1);
1389     }
1390
1391     return 0;
1392 }
1393
1394 static int calc_poc(int pic_order_cnt_lsb)
1395 {
1396     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1397     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1398     int PicOrderCntMsb, TopFieldOrderCnt;
1399
1400     if (current_frame_type == FRAME_IDR)
1401         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1402     else {
1403         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1404         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1405     }
1406
1407     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1408         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1409         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1410     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1411              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1412         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1413     else
1414         PicOrderCntMsb = prevPicOrderCntMsb;
1415
1416     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1417
1418     if (current_frame_type != FRAME_B) {
1419         PicOrderCntMsb_ref = PicOrderCntMsb;
1420         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1421     }
1422
1423     return TopFieldOrderCnt;
1424 }
1425
1426 static int render_picture(void)
1427 {
1428     VABufferID pic_param_buf;
1429     VAStatus va_status;
1430     int i = 0;
1431
1432     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1433     pic_param.CurrPic.frame_idx = current_frame_num;
1434     pic_param.CurrPic.flags = 0;
1435     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1436     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1437     CurrentCurrPic = pic_param.CurrPic;
1438
1439     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1440         update_RefPicList(); /* calc RefPicList */
1441         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1442         if (current_frame_type == FRAME_P) {
1443             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1444         } else if (current_frame_type == FRAME_B) {
1445             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1446             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1447         }
1448     } else {
1449         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1450         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1451             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1452             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1453         }
1454     }
1455
1456     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1457     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1458     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1459     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1460     pic_param.frame_num = current_frame_num;
1461     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1462     pic_param.last_picture = false;  // FIXME
1463     pic_param.pic_init_qp = initial_qp;
1464
1465     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1466                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1467     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1468
1469     va_status = vaRenderPicture(va_dpy, context_id, &pic_param_buf, 1);
1470     CHECK_VASTATUS(va_status, "vaRenderPicture");
1471
1472     return 0;
1473 }
1474
1475 static int render_packedsequence(void)
1476 {
1477     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1478     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1479     unsigned int length_in_bits;
1480     unsigned char *packedseq_buffer = NULL;
1481     VAStatus va_status;
1482
1483     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1484
1485     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1486
1487     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1488     packedheader_param_buffer.has_emulation_bytes = 0;
1489     va_status = vaCreateBuffer(va_dpy,
1490                                context_id,
1491                                VAEncPackedHeaderParameterBufferType,
1492                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1493                                &packedseq_para_bufid);
1494     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1495
1496     va_status = vaCreateBuffer(va_dpy,
1497                                context_id,
1498                                VAEncPackedHeaderDataBufferType,
1499                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1500                                &packedseq_data_bufid);
1501     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1502
1503     render_id[0] = packedseq_para_bufid;
1504     render_id[1] = packedseq_data_bufid;
1505     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1506     CHECK_VASTATUS(va_status, "vaRenderPicture");
1507
1508     free(packedseq_buffer);
1509
1510     return 0;
1511 }
1512
1513
1514 static int render_packedpicture(void)
1515 {
1516     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1517     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1518     unsigned int length_in_bits;
1519     unsigned char *packedpic_buffer = NULL;
1520     VAStatus va_status;
1521
1522     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1523     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1524     packedheader_param_buffer.bit_length = length_in_bits;
1525     packedheader_param_buffer.has_emulation_bytes = 0;
1526
1527     va_status = vaCreateBuffer(va_dpy,
1528                                context_id,
1529                                VAEncPackedHeaderParameterBufferType,
1530                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1531                                &packedpic_para_bufid);
1532     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1533
1534     va_status = vaCreateBuffer(va_dpy,
1535                                context_id,
1536                                VAEncPackedHeaderDataBufferType,
1537                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1538                                &packedpic_data_bufid);
1539     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1540
1541     render_id[0] = packedpic_para_bufid;
1542     render_id[1] = packedpic_data_bufid;
1543     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1544     CHECK_VASTATUS(va_status, "vaRenderPicture");
1545
1546     free(packedpic_buffer);
1547
1548     return 0;
1549 }
1550
1551 static void render_packedslice()
1552 {
1553     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1554     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1555     unsigned int length_in_bits;
1556     unsigned char *packedslice_buffer = NULL;
1557     VAStatus va_status;
1558
1559     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1560     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1561     packedheader_param_buffer.bit_length = length_in_bits;
1562     packedheader_param_buffer.has_emulation_bytes = 0;
1563
1564     va_status = vaCreateBuffer(va_dpy,
1565                                context_id,
1566                                VAEncPackedHeaderParameterBufferType,
1567                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1568                                &packedslice_para_bufid);
1569     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1570
1571     va_status = vaCreateBuffer(va_dpy,
1572                                context_id,
1573                                VAEncPackedHeaderDataBufferType,
1574                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1575                                &packedslice_data_bufid);
1576     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1577
1578     render_id[0] = packedslice_para_bufid;
1579     render_id[1] = packedslice_data_bufid;
1580     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1581     CHECK_VASTATUS(va_status, "vaRenderPicture");
1582
1583     free(packedslice_buffer);
1584 }
1585
1586 static int render_slice(void)
1587 {
1588     VABufferID slice_param_buf;
1589     VAStatus va_status;
1590     int i;
1591
1592     update_RefPicList();
1593
1594     /* one frame, one slice */
1595     slice_param.macroblock_address = 0;
1596     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1597     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1598     if (current_frame_type == FRAME_IDR) {
1599         if (current_frame_encoding != 0)
1600             ++slice_param.idr_pic_id;
1601     } else if (current_frame_type == FRAME_P) {
1602         int refpiclist0_max = h264_maxref & 0xffff;
1603         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1604
1605         for (i = refpiclist0_max; i < 32; i++) {
1606             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1607             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1608         }
1609     } else if (current_frame_type == FRAME_B) {
1610         int refpiclist0_max = h264_maxref & 0xffff;
1611         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1612
1613         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1614         for (i = refpiclist0_max; i < 32; i++) {
1615             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1616             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1617         }
1618
1619         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1620         for (i = refpiclist1_max; i < 32; i++) {
1621             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1622             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1623         }
1624     }
1625
1626     slice_param.slice_alpha_c0_offset_div2 = 0;
1627     slice_param.slice_beta_offset_div2 = 0;
1628     slice_param.direct_spatial_mv_pred_flag = 1;
1629     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1630
1631
1632     if (h264_packedheader &&
1633         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1634         render_packedslice();
1635
1636     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1637                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1638     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1639
1640     va_status = vaRenderPicture(va_dpy, context_id, &slice_param_buf, 1);
1641     CHECK_VASTATUS(va_status, "vaRenderPicture");
1642
1643     return 0;
1644 }
1645
1646
1647
1648 int H264Encoder::save_codeddata(storage_task task)
1649 {
1650     VACodedBufferSegment *buf_list = NULL;
1651     VAStatus va_status;
1652     unsigned int coded_size = 0;
1653
1654     string data;
1655
1656     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / frame_rate);  // So we never get negative dts.
1657
1658     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1659     CHECK_VASTATUS(va_status, "vaMapBuffer");
1660     while (buf_list != NULL) {
1661         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1662         buf_list = (VACodedBufferSegment *) buf_list->next;
1663
1664         frame_size += coded_size;
1665     }
1666     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1667
1668     {
1669         // Add video.
1670         AVPacket pkt;
1671         memset(&pkt, 0, sizeof(pkt));
1672         pkt.buf = nullptr;
1673         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1674         pkt.size = data.size();
1675         pkt.stream_index = 0;
1676         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1677             pkt.flags = AV_PKT_FLAG_KEY;
1678         } else {
1679             pkt.flags = 0;
1680         }
1681         //pkt.duration = 1;
1682         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1683     }
1684     // Encode and add all audio frames up to and including the pts of this video frame.
1685     // (They can never be queued to us after the video frame they belong to, only before.)
1686     for ( ;; ) {
1687         int64_t audio_pts;
1688         std::vector<float> audio;
1689         {
1690              unique_lock<mutex> lock(frame_queue_mutex);
1691              if (pending_audio_frames.empty()) break;
1692              auto it = pending_audio_frames.begin();
1693              if (it->first > task.pts) break;
1694              audio_pts = it->first;
1695              audio = move(it->second);
1696              pending_audio_frames.erase(it);
1697         }
1698
1699         AVFrame *frame = avcodec_alloc_frame();
1700         frame->nb_samples = audio.size() / 2;
1701         frame->format = AV_SAMPLE_FMT_FLT;
1702         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1703
1704         unique_ptr<float[]> planar_samples(new float[audio.size()]);
1705         avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_FLTP, (const uint8_t*)planar_samples.get(), audio.size() * sizeof(float), 0);
1706         for (int i = 0; i < frame->nb_samples; ++i) {
1707             planar_samples[i] = audio[i * 2 + 0];
1708             planar_samples[i + frame->nb_samples] = audio[i * 2 + 1];
1709         }
1710
1711         AVPacket pkt;
1712         av_init_packet(&pkt);
1713         pkt.data = nullptr;
1714         pkt.size = 0;
1715         int got_output;
1716         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1717         if (got_output) {
1718             pkt.stream_index = 1;
1719             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1720         }
1721         // TODO: Delayed frames.
1722         avcodec_free_frame(&frame);
1723     }
1724
1725 #if 0
1726     printf("\r      "); /* return back to startpoint */
1727     switch (encode_order % 4) {
1728         case 0:
1729             printf("|");
1730             break;
1731         case 1:
1732             printf("/");
1733             break;
1734         case 2:
1735             printf("-");
1736             break;
1737         case 3:
1738             printf("\\");
1739             break;
1740     }
1741     printf("%08lld", encode_order);
1742     printf("(%06d bytes coded)", coded_size);
1743 #endif
1744
1745     return 0;
1746 }
1747
1748
1749 // this is weird. but it seems to put a new frame onto the queue
1750 void H264Encoder::storage_task_enqueue(storage_task task)
1751 {
1752         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1753         storage_task_queue.push(move(task));
1754         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1755         storage_task_queue_changed.notify_all();
1756 }
1757
1758 void H264Encoder::storage_task_thread()
1759 {
1760         for ( ;; ) {
1761                 storage_task current;
1762                 {
1763                         // wait until there's an encoded frame
1764                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1765                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1766                         if (storage_thread_should_quit) return;
1767                         current = move(storage_task_queue.front());
1768                         storage_task_queue.pop();
1769                 }
1770
1771                 VAStatus va_status;
1772
1773                 // waits for data, then saves it to disk.
1774                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1775                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1776                 save_codeddata(move(current));
1777
1778                 {
1779                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1780                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1781                         storage_task_queue_changed.notify_all();
1782                 }
1783         }
1784 }
1785
1786 static int release_encode()
1787 {
1788     int i;
1789
1790     for (i = 0; i < SURFACE_NUM; i++) {
1791         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1792         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1793         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1794     }
1795
1796     vaDestroyContext(va_dpy, context_id);
1797     vaDestroyConfig(va_dpy, config_id);
1798
1799     return 0;
1800 }
1801
1802 static int deinit_va()
1803 {
1804     vaTerminate(va_dpy);
1805
1806     va_close_display(va_dpy);
1807
1808     return 0;
1809 }
1810
1811
1812 static int print_input()
1813 {
1814     printf("\n\nINPUT:Try to encode H264...\n");
1815     if (rc_mode != -1)
1816         printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1817     printf("INPUT: Resolution   : %dx%dframes\n", frame_width, frame_height);
1818     printf("INPUT: FrameRate    : %d\n", frame_rate);
1819     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
1820     printf("INPUT: Slieces      : %d\n", frame_slices);
1821     printf("INPUT: IntraPeriod  : %d\n", intra_period);
1822     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
1823     printf("INPUT: IpPeriod     : %d\n", ip_period);
1824     printf("INPUT: Initial QP   : %d\n", initial_qp);
1825     printf("INPUT: Min QP       : %d\n", minimal_qp);
1826     printf("INPUT: Coded Clip   : %s\n", coded_fn);
1827
1828     printf("\n\n"); /* return back to startpoint */
1829
1830     return 0;
1831 }
1832
1833 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
1834         : current_storage_frame(0), surface(surface), httpd(httpd)
1835 {
1836         AVCodec *codec_audio = avcodec_find_encoder(AV_CODEC_ID_MP3);
1837         context_audio = avcodec_alloc_context3(codec_audio);
1838         context_audio->bit_rate = 256000;
1839         context_audio->sample_rate = OUTPUT_FREQUENCY;
1840         context_audio->sample_fmt = AV_SAMPLE_FMT_FLTP;
1841         context_audio->channels = 2;
1842         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1843         context_audio->time_base = AVRational{1, TIMEBASE};
1844         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1845                 fprintf(stderr, "Could not open codec\n");
1846                 exit(1);
1847         }
1848
1849         frame_width = width;
1850         frame_height = height;
1851         frame_width_mbaligned = (frame_width + 15) & (~15);
1852         frame_height_mbaligned = (frame_height + 15) & (~15);
1853         frame_bitrate = 15000000;  // / 60;
1854         current_frame_encoding = 0;
1855
1856         print_input();
1857
1858         init_va();
1859         setup_encode();
1860
1861         // No frames are ready yet.
1862         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1863
1864         memset(&seq_param, 0, sizeof(seq_param));
1865         memset(&pic_param, 0, sizeof(pic_param));
1866         memset(&slice_param, 0, sizeof(slice_param));
1867
1868         storage_thread = std::thread(&H264Encoder::storage_task_thread, this);
1869
1870         copy_thread = std::thread([this]{
1871                 //SDL_GL_MakeCurrent(window, context);
1872                 QOpenGLContext *context = create_context(this->surface);
1873                 eglBindAPI(EGL_OPENGL_API);
1874                 if (!make_current(context, this->surface)) {
1875                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1876                                 eglGetError());
1877                         exit(1);
1878                 }
1879                 copy_thread_func();
1880         });
1881 }
1882
1883 H264Encoder::~H264Encoder()
1884 {
1885         {
1886                 unique_lock<mutex> lock(storage_task_queue_mutex);
1887                 storage_thread_should_quit = true;
1888                 storage_task_queue_changed.notify_all();
1889         }
1890         {
1891                 unique_lock<mutex> lock(frame_queue_mutex);
1892                 copy_thread_should_quit = true;
1893                 frame_queue_nonempty.notify_one();
1894         }
1895         storage_thread.join();
1896         copy_thread.join();
1897
1898         release_encode();
1899         deinit_va();
1900 }
1901
1902 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1903 {
1904         {
1905                 // Wait until this frame slot is done encoding.
1906                 std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1907                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1908                 if (storage_thread_should_quit) return false;
1909         }
1910
1911         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1912         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1913         *y_tex = surf->y_tex;
1914         *cbcr_tex = surf->cbcr_tex;
1915
1916         VASurfaceID surface = surf->src_surface;
1917         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1918         CHECK_VASTATUS(va_status, "vaDeriveImage");
1919
1920         VABufferInfo buf_info;
1921         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1922         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1923         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1924
1925         // Create Y image.
1926         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1927         EGLint y_attribs[] = {
1928                 EGL_WIDTH, frame_width,
1929                 EGL_HEIGHT, frame_height,
1930                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1931                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1932                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1933                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1934                 EGL_NONE
1935         };
1936
1937         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1938         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1939
1940         // Associate Y image to a texture.
1941         glBindTexture(GL_TEXTURE_2D, *y_tex);
1942         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1943
1944         // Create CbCr image.
1945         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1946         EGLint cbcr_attribs[] = {
1947                 EGL_WIDTH, frame_width,
1948                 EGL_HEIGHT, frame_height,
1949                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1950                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1951                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1952                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1953                 EGL_NONE
1954         };
1955
1956         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1957         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1958
1959         // Associate CbCr image to a texture.
1960         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1961         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1962
1963         return true;
1964 }
1965
1966 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
1967 {
1968         {
1969                 unique_lock<mutex> lock(frame_queue_mutex);
1970                 pending_audio_frames[pts] = move(audio);
1971         }
1972         frame_queue_nonempty.notify_one();
1973 }
1974
1975
1976 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
1977 {
1978         {
1979                 unique_lock<mutex> lock(frame_queue_mutex);
1980                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1981                 ++current_storage_frame;
1982         }
1983         frame_queue_nonempty.notify_one();
1984 }
1985
1986 void H264Encoder::copy_thread_func()
1987 {
1988         int64_t last_dts = -1;
1989         for ( ;; ) {
1990                 PendingFrame frame;
1991                 int pts_lag;
1992                 encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
1993                                        &current_frame_display, &current_frame_type, &pts_lag);
1994                 if (current_frame_type == FRAME_IDR) {
1995                         numShortTerm = 0;
1996                         current_frame_num = 0;
1997                         current_IDR_display = current_frame_display;
1998                 }
1999
2000                 {
2001                         unique_lock<mutex> lock(frame_queue_mutex);
2002                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; });
2003                         if (copy_thread_should_quit) return;
2004                         frame = move(pending_video_frames[current_frame_display]);
2005                         pending_video_frames.erase(current_frame_display);
2006                 }
2007
2008                 // Wait for the GPU to be done with the frame.
2009                 glClientWaitSync(frame.fence.get(), 0, 0);
2010
2011                 // Release back any input frames we needed to render this frame.
2012                 frame.input_frames.clear();
2013
2014                 // Unmap the image.
2015                 GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
2016                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2017                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2018                 VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2019                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2020                 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2021                 CHECK_VASTATUS(va_status, "vaDestroyImage");
2022
2023                 VASurfaceID surface = surf->src_surface;
2024
2025                 // Schedule the frame for encoding.
2026                 va_status = vaBeginPicture(va_dpy, context_id, surface);
2027                 CHECK_VASTATUS(va_status, "vaBeginPicture");
2028
2029                 if (current_frame_type == FRAME_IDR) {
2030                         render_sequence();
2031                         render_picture();
2032                         if (h264_packedheader) {
2033                                 render_packedsequence();
2034                                 render_packedpicture();
2035                         }
2036                 } else {
2037                         //render_sequence();
2038                         render_picture();
2039                 }
2040                 render_slice();
2041
2042                 va_status = vaEndPicture(va_dpy, context_id);
2043                 CHECK_VASTATUS(va_status, "vaEndPicture");
2044
2045                 // Determine the pts and dts of this frame.
2046                 int64_t pts = frame.pts;
2047                 int64_t dts;
2048                 if (pts_lag == -1) {
2049                         assert(last_dts != -1);
2050                         dts = last_dts + (TIMEBASE / frame_rate);
2051                 } else {
2052                         dts = pts - pts_lag;
2053                 }
2054                 last_dts = dts;
2055
2056                 // so now the data is done encoding (well, async job kicked off)...
2057                 // we send that to the storage thread
2058                 storage_task tmp;
2059                 tmp.display_order = current_frame_display;
2060                 tmp.frame_type = current_frame_type;
2061                 tmp.pts = pts;
2062                 tmp.dts = dts;
2063                 storage_task_enqueue(move(tmp));
2064
2065                 update_ReferenceFrames();
2066                 ++current_frame_encoding;
2067         }
2068 }