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 "timebase.h"
  30
  31 class QOpenGLContext;
  32 class QSurface;
  33
  34 #define CHECK_VASTATUS(va_status, func)                                 \
  35     if (va_status != VA_STATUS_SUCCESS) {                               \
  36         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  37         exit(1);                                                        \
  38     }
  39
  40 //#include "loadsurface.h"
  41
  42 #define NAL_REF_IDC_NONE        0
  43 #define NAL_REF_IDC_LOW         1
  44 #define NAL_REF_IDC_MEDIUM      2
  45 #define NAL_REF_IDC_HIGH        3
  46
  47 #define NAL_NON_IDR             1
  48 #define NAL_IDR                 5
  49 #define NAL_SPS                 7
  50 #define NAL_PPS                 8
  51 #define NAL_SEI                 6
  52
  53 #define SLICE_TYPE_P            0
  54 #define SLICE_TYPE_B            1
  55 #define SLICE_TYPE_I            2
  56 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  57 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  58 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  59
  60
  61 #define ENTROPY_MODE_CAVLC      0
  62 #define ENTROPY_MODE_CABAC      1
  63
  64 #define PROFILE_IDC_BASELINE    66
  65 #define PROFILE_IDC_MAIN        77
  66 #define PROFILE_IDC_HIGH        100
  67
  68 #define BITSTREAM_ALLOCATE_STEPPING     4096
  69
  70 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  71 static  VADisplay va_dpy;
  72 static  VAProfile h264_profile = (VAProfile)~0;
  73 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  74 static  int config_attrib_num = 0, enc_packed_header_idx;
  75
  76 struct GLSurface {
  77         VASurfaceID src_surface, ref_surface;
  78         VABufferID coded_buf;
  79
  80         VAImage surface_image;
  81         GLuint y_tex, cbcr_tex;
  82         EGLImage y_egl_image, cbcr_egl_image;
  83 };
  84 GLSurface gl_surfaces[SURFACE_NUM];
  85
  86 static  VAConfigID config_id;
  87 static  VAContextID context_id;
  88 static  VAEncSequenceParameterBufferH264 seq_param;
  89 static  VAEncPictureParameterBufferH264 pic_param;
  90 static  VAEncSliceParameterBufferH264 slice_param;
  91 static  VAPictureH264 CurrentCurrPic;
  92 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
  93
  94 static  unsigned int MaxFrameNum = (2<<16);
  95 static  unsigned int MaxPicOrderCntLsb = (2<<8);
  96 static  unsigned int Log2MaxFrameNum = 16;
  97 static  unsigned int Log2MaxPicOrderCntLsb = 8;
  98
  99 static  unsigned int num_ref_frames = 2;
 100 static  unsigned int numShortTerm = 0;
 101 static  int constraint_set_flag = 0;
 102 static  int h264_packedheader = 0; /* support pack header? */
 103 static  int h264_maxref = (1<<16|1);
 104 static  int h264_entropy_mode = 1; /* cabac */
 105
 106 static  char *coded_fn = NULL;
 107
 108 static  int frame_width = 176;
 109 static  int frame_height = 144;
 110 static  int frame_width_mbaligned;
 111 static  int frame_height_mbaligned;
 112 static  int frame_rate = 60;
 113 static  unsigned int frame_bitrate = 0;
 114 static  unsigned int frame_slices = 1;
 115 static  double frame_size = 0;
 116 static  int initial_qp = 15;
 117 //static  int initial_qp = 28;
 118 static  int minimal_qp = 0;
 119 static  int intra_period = 30;
 120 static  int intra_idr_period = 60;
 121 static  int ip_period = 3;
 122 static  int rc_mode = -1;
 123 static  int rc_default_modes[] = {
 124     VA_RC_VBR,
 125     VA_RC_CQP,
 126     VA_RC_VBR_CONSTRAINED,
 127     VA_RC_CBR,
 128     VA_RC_VCM,
 129     VA_RC_NONE,
 130 };
 131 static  unsigned long long current_frame_encoding = 0;
 132 static  unsigned long long current_frame_display = 0;
 133 static  unsigned long long current_IDR_display = 0;
 134 static  unsigned int current_frame_num = 0;
 135 static  int current_frame_type;
 136
 137 static  int misc_priv_type = 0;
 138 static  int misc_priv_value = 0;
 139
 140 /* thread to save coded data */
 141 #define SRC_SURFACE_FREE        0
 142 #define SRC_SURFACE_IN_ENCODING 1
 143
 144 struct __bitstream {
 145     unsigned int *buffer;
 146     int bit_offset;
 147     int max_size_in_dword;
 148 };
 149 typedef struct __bitstream bitstream;
 150
 151 using namespace std;
 152
 153 static unsigned int
 154 va_swap32(unsigned int val)
 155 {
 156     unsigned char *pval = (unsigned char *)&val;
 157
 158     return ((pval[0] << 24)     |
 159             (pval[1] << 16)     |
 160             (pval[2] << 8)      |
 161             (pval[3] << 0));
 162 }
 163
 164 static void
 165 bitstream_start(bitstream *bs)
 166 {
 167     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 168     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 169     bs->bit_offset = 0;
 170 }
 171
 172 static void
 173 bitstream_end(bitstream *bs)
 174 {
 175     int pos = (bs->bit_offset >> 5);
 176     int bit_offset = (bs->bit_offset & 0x1f);
 177     int bit_left = 32 - bit_offset;
 178
 179     if (bit_offset) {
 180         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 181     }
 182 }
 183
 184 static void
 185 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 186 {
 187     int pos = (bs->bit_offset >> 5);
 188     int bit_offset = (bs->bit_offset & 0x1f);
 189     int bit_left = 32 - bit_offset;
 190
 191     if (!size_in_bits)
 192         return;
 193
 194     bs->bit_offset += size_in_bits;
 195
 196     if (bit_left > size_in_bits) {
 197         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 198     } else {
 199         size_in_bits -= bit_left;
 200         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 201         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 202
 203         if (pos + 1 == bs->max_size_in_dword) {
 204             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 205             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 206         }
 207
 208         bs->buffer[pos + 1] = val;
 209     }
 210 }
 211
 212 static void
 213 bitstream_put_ue(bitstream *bs, unsigned int val)
 214 {
 215     int size_in_bits = 0;
 216     int tmp_val = ++val;
 217
 218     while (tmp_val) {
 219         tmp_val >>= 1;
 220         size_in_bits++;
 221     }
 222
 223     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 224     bitstream_put_ui(bs, val, size_in_bits);
 225 }
 226
 227 static void
 228 bitstream_put_se(bitstream *bs, int val)
 229 {
 230     unsigned int new_val;
 231
 232     if (val <= 0)
 233         new_val = -2 * val;
 234     else
 235         new_val = 2 * val - 1;
 236
 237     bitstream_put_ue(bs, new_val);
 238 }
 239
 240 static void
 241 bitstream_byte_aligning(bitstream *bs, int bit)
 242 {
 243     int bit_offset = (bs->bit_offset & 0x7);
 244     int bit_left = 8 - bit_offset;
 245     int new_val;
 246
 247     if (!bit_offset)
 248         return;
 249
 250     assert(bit == 0 || bit == 1);
 251
 252     if (bit)
 253         new_val = (1 << bit_left) - 1;
 254     else
 255         new_val = 0;
 256
 257     bitstream_put_ui(bs, new_val, bit_left);
 258 }
 259
 260 static void
 261 rbsp_trailing_bits(bitstream *bs)
 262 {
 263     bitstream_put_ui(bs, 1, 1);
 264     bitstream_byte_aligning(bs, 0);
 265 }
 266
 267 static void nal_start_code_prefix(bitstream *bs)
 268 {
 269     bitstream_put_ui(bs, 0x00000001, 32);
 270 }
 271
 272 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 273 {
 274     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 275     bitstream_put_ui(bs, nal_ref_idc, 2);
 276     bitstream_put_ui(bs, nal_unit_type, 5);
 277 }
 278
 279 static void sps_rbsp(bitstream *bs)
 280 {
 281     int profile_idc = PROFILE_IDC_BASELINE;
 282
 283     if (h264_profile  == VAProfileH264High)
 284         profile_idc = PROFILE_IDC_HIGH;
 285     else if (h264_profile  == VAProfileH264Main)
 286         profile_idc = PROFILE_IDC_MAIN;
 287
 288     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 289     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 290     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 291     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 292     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 293     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 294     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 295     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 296
 297     if ( profile_idc == PROFILE_IDC_HIGH) {
 298         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 299         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 300         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 301         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 302         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 303     }
 304
 305     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 306     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 307
 308     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 309         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 310     else {
 311         assert(0);
 312     }
 313
 314     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 315     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 316
 317     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 318     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 319     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 320
 321     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 322         assert(0);
 323     }
 324
 325     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 326     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 327
 328     if (seq_param.frame_cropping_flag) {
 329         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 330         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 331         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 332         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 333     }
 334
 335     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 336     if ( false ) {
 337         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 338     } else {
 339         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 340         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 341         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 342         bitstream_put_ui(bs, 0, 1); /* video_signal_type_present_flag */
 343         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 344         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 345         {
 346             bitstream_put_ui(bs, 1, 32);  // FPS
 347             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 348             bitstream_put_ui(bs, 1, 1);
 349         }
 350         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 351         {
 352             // hrd_parameters
 353             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 354             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 355             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 356
 357             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 358             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 359             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 360
 361             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 362             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 363             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 364             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 365         }
 366         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 367         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 368
 369         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 370         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 371     }
 372
 373     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 374 }
 375
 376
 377 static void pps_rbsp(bitstream *bs)
 378 {
 379     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 380     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 381
 382     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 383
 384     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 385
 386     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 387
 388     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 389     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 390
 391     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 392     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 393
 394     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 395     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 396     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 397
 398     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 399     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 400     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 401
 402     /* more_rbsp_data */
 403     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 404     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 405     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 406
 407     rbsp_trailing_bits(bs);
 408 }
 409
 410 static void slice_header(bitstream *bs)
 411 {
 412     int first_mb_in_slice = slice_param.macroblock_address;
 413
 414     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 415     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 416     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 417     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 418
 419     /* frame_mbs_only_flag == 1 */
 420     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 421         /* FIXME: */
 422         assert(0);
 423     }
 424
 425     if (pic_param.pic_fields.bits.idr_pic_flag)
 426         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 427
 428     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 429         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 430         /* pic_order_present_flag == 0 */
 431     } else {
 432         /* FIXME: */
 433         assert(0);
 434     }
 435
 436     /* redundant_pic_cnt_present_flag == 0 */
 437     /* slice type */
 438     if (IS_P_SLICE(slice_param.slice_type)) {
 439         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 440
 441         if (slice_param.num_ref_idx_active_override_flag)
 442             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 443
 444         /* ref_pic_list_reordering */
 445         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 446     } else if (IS_B_SLICE(slice_param.slice_type)) {
 447         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 448
 449         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 450
 451         if (slice_param.num_ref_idx_active_override_flag) {
 452             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 453             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 454         }
 455
 456         /* ref_pic_list_reordering */
 457         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 458         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 459     }
 460
 461     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 462          IS_P_SLICE(slice_param.slice_type)) ||
 463         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 464          IS_B_SLICE(slice_param.slice_type))) {
 465         /* FIXME: fill weight/offset table */
 466         assert(0);
 467     }
 468
 469     /* dec_ref_pic_marking */
 470     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 471         unsigned char no_output_of_prior_pics_flag = 0;
 472         unsigned char long_term_reference_flag = 0;
 473         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 474
 475         if (pic_param.pic_fields.bits.idr_pic_flag) {
 476             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 477             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 478         } else {
 479             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 480         }
 481     }
 482
 483     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 484         !IS_I_SLICE(slice_param.slice_type))
 485         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 486
 487     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 488
 489     /* ignore for SP/SI */
 490
 491     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 492         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 493
 494         if (slice_param.disable_deblocking_filter_idc != 1) {
 495             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 496             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 497         }
 498     }
 499
 500     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 501         bitstream_byte_aligning(bs, 1);
 502     }
 503 }
 504
 505 static int
 506 build_packed_pic_buffer(unsigned char **header_buffer)
 507 {
 508     bitstream bs;
 509
 510     bitstream_start(&bs);
 511     nal_start_code_prefix(&bs);
 512     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 513     pps_rbsp(&bs);
 514     bitstream_end(&bs);
 515
 516     *header_buffer = (unsigned char *)bs.buffer;
 517     return bs.bit_offset;
 518 }
 519
 520 static int
 521 build_packed_seq_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_SPS);
 528     sps_rbsp(&bs);
 529     bitstream_end(&bs);
 530
 531     *header_buffer = (unsigned char *)bs.buffer;
 532     return bs.bit_offset;
 533 }
 534
 535 static int build_packed_slice_buffer(unsigned char **header_buffer)
 536 {
 537     bitstream bs;
 538     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 539     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 540
 541     bitstream_start(&bs);
 542     nal_start_code_prefix(&bs);
 543
 544     if (IS_I_SLICE(slice_param.slice_type)) {
 545         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 546     } else if (IS_P_SLICE(slice_param.slice_type)) {
 547         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 548     } else {
 549         assert(IS_B_SLICE(slice_param.slice_type));
 550         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 551     }
 552
 553     slice_header(&bs);
 554     bitstream_end(&bs);
 555
 556     *header_buffer = (unsigned char *)bs.buffer;
 557     return bs.bit_offset;
 558 }
 559
 560
 561 /*
 562   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 563   1) period between Frame #X and Frame #N = #X - #N
 564   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 565   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 566   4) intra_period and intra_idr_period take precedence over ip_period
 567   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 568      of I/IDR frames, see bellow examples
 569   -------------------------------------------------------------------
 570   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 571   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 572   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 573   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 574   1            1                ignored    IDR IDR IDR IDR...
 575   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 576   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 577   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 578                                               (PBB)(IBB)(PBB)(IBB)...
 579   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 580                                            IDRPPPPP IPPPPP IPPPPP...
 581   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 582                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 583                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 584                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 585                                            {IDR(PBB)(PBB)}                     (6/6/3)
 586                                            {IDR(PBB)(PBB)}.
 587 */
 588
 589 // General pts/dts strategy:
 590 //
 591 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 592 // bit tricky. We assume first of all that the frame rate never goes _above_
 593 // <frame_rate>, which gives us a frame period N. The decoder can always decode
 594 // in at least this speed, as long at dts <= pts (the frame is not attempted
 595 // presented before it is decoded). Furthermore, we never have longer chains of
 596 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 597 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 598 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 599 // frame, with an order O <= C.)
 600 //
 601 // Many strategies are possible, but we establish these rules:
 602 //
 603 //  - Tip frames have dts = pts - (C-O)*N.
 604 //  - Non-tip frames have dts = dts_last + N.
 605 //
 606 // An example, with C=2 and N=10 and the data flow showed with arrows:
 607 //
 608 //        I  B  P  B  B  P
 609 //   pts: 30 40 50 60 70 80
 610 //        ↓  ↓     ↓
 611 //   dts: 10 30 20 60 50←40
 612 //         |  |  ↑        ↑
 613 //         `--|--'        |
 614 //             `----------'
 615 //
 616 // To show that this works fine also with irregular spacings, let's say that
 617 // the third frame is delayed a bit (something earlier was dropped). Now the
 618 // situation looks like this:
 619 //
 620 //        I  B  P  B  B   P
 621 //   pts: 30 40 80 90 100 110
 622 //        ↓  ↓     ↓
 623 //   dts: 10 30 20 90 50←40
 624 //         |  |  ↑        ↑
 625 //         `--|--'        |
 626 //             `----------'
 627 //
 628 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 629 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 630 //
 631 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 632 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 633 // a tip frame and should be given a dts based on the previous one.
 634 #define FRAME_P 0
 635 #define FRAME_B 1
 636 #define FRAME_I 2
 637 #define FRAME_IDR 7
 638 void encoding2display_order(
 639     unsigned long long encoding_order, int intra_period,
 640     int intra_idr_period, int ip_period,
 641     unsigned long long *displaying_order,
 642     int *frame_type, int *pts_lag)
 643 {
 644     int encoding_order_gop = 0;
 645
 646     *pts_lag = 0;
 647
 648     if (intra_period == 1) { /* all are I/IDR frames */
 649         *displaying_order = encoding_order;
 650         if (intra_idr_period == 0)
 651             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 652         else
 653             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 654         return;
 655     }
 656
 657     if (intra_period == 0)
 658         intra_idr_period = 0;
 659
 660     if (ip_period == 1) {
 661         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 662         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 663         *displaying_order = encoding_order;
 664
 665         if (encoding_order_gop == 0) { /* the first frame */
 666             *frame_type = FRAME_IDR;
 667         } else if (intra_period != 0 && /* have I frames */
 668                    encoding_order_gop >= 2 &&
 669                    (encoding_order_gop % intra_period == 0)) {
 670             *frame_type = FRAME_I;
 671         } else {
 672             *frame_type = FRAME_P;
 673         }
 674         return;
 675     }
 676
 677     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 678     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 679     *pts_lag = -1;  // Most frames are not tip frames.
 680
 681     if (encoding_order_gop == 0) { /* the first frame */
 682         *frame_type = FRAME_IDR;
 683         *displaying_order = encoding_order;
 684         // IDR frames are a special case; I honestly can't find the logic behind
 685         // why this is the right thing, but it seems to line up nicely in practice :-)
 686         *pts_lag = TIMEBASE / frame_rate;
 687     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 688         *frame_type = FRAME_B;
 689         *displaying_order = encoding_order - 1;
 690         if ((encoding_order_gop % ip_period) == 0) {
 691             *pts_lag = 0;  // Last B-frame.
 692         }
 693     } else if (intra_period != 0 && /* have I frames */
 694                encoding_order_gop >= 2 &&
 695                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 696         *frame_type = FRAME_I;
 697         *displaying_order = encoding_order + ip_period - 1;
 698     } else {
 699         *frame_type = FRAME_P;
 700         *displaying_order = encoding_order + ip_period - 1;
 701     }
 702 }
 703
 704
 705 static const char *rc_to_string(int rcmode)
 706 {
 707     switch (rc_mode) {
 708     case VA_RC_NONE:
 709         return "NONE";
 710     case VA_RC_CBR:
 711         return "CBR";
 712     case VA_RC_VBR:
 713         return "VBR";
 714     case VA_RC_VCM:
 715         return "VCM";
 716     case VA_RC_CQP:
 717         return "CQP";
 718     case VA_RC_VBR_CONSTRAINED:
 719         return "VBR_CONSTRAINED";
 720     default:
 721         return "Unknown";
 722     }
 723 }
 724
 725 #if 0
 726 static int process_cmdline(int argc, char *argv[])
 727 {
 728     char c;
 729     const struct option long_opts[] = {
 730         {"help", no_argument, NULL, 0 },
 731         {"bitrate", required_argument, NULL, 1 },
 732         {"minqp", required_argument, NULL, 2 },
 733         {"initialqp", required_argument, NULL, 3 },
 734         {"intra_period", required_argument, NULL, 4 },
 735         {"idr_period", required_argument, NULL, 5 },
 736         {"ip_period", required_argument, NULL, 6 },
 737         {"rcmode", required_argument, NULL, 7 },
 738         {"srcyuv", required_argument, NULL, 9 },
 739         {"recyuv", required_argument, NULL, 10 },
 740         {"fourcc", required_argument, NULL, 11 },
 741         {"syncmode", no_argument, NULL, 12 },
 742         {"enablePSNR", no_argument, NULL, 13 },
 743         {"prit", required_argument, NULL, 14 },
 744         {"priv", required_argument, NULL, 15 },
 745         {"framecount", required_argument, NULL, 16 },
 746         {"entropy", required_argument, NULL, 17 },
 747         {"profile", required_argument, NULL, 18 },
 748         {NULL, no_argument, NULL, 0 }};
 749     int long_index;
 750
 751     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 752         switch (c) {
 753         case 'w':
 754             frame_width = atoi(optarg);
 755             break;
 756         case 'h':
 757             frame_height = atoi(optarg);
 758             break;
 759         case 'n':
 760         case 'f':
 761             frame_rate = atoi(optarg);
 762             break;
 763         case 'o':
 764             coded_fn = strdup(optarg);
 765             break;
 766         case 0:
 767             print_help();
 768             exit(0);
 769         case 1:
 770             frame_bitrate = atoi(optarg);
 771             break;
 772         case 2:
 773             minimal_qp = atoi(optarg);
 774             break;
 775         case 3:
 776             initial_qp = atoi(optarg);
 777             break;
 778         case 4:
 779             intra_period = atoi(optarg);
 780             break;
 781         case 5:
 782             intra_idr_period = atoi(optarg);
 783             break;
 784         case 6:
 785             ip_period = atoi(optarg);
 786             break;
 787         case 7:
 788             rc_mode = string_to_rc(optarg);
 789             if (rc_mode < 0) {
 790                 print_help();
 791                 exit(1);
 792             }
 793             break;
 794         case 9:
 795             srcyuv_fn = strdup(optarg);
 796             break;
 797         case 11:
 798             srcyuv_fourcc = string_to_fourcc(optarg);
 799             if (srcyuv_fourcc <= 0) {
 800                 print_help();
 801                 exit(1);
 802             }
 803             break;
 804         case 13:
 805             calc_psnr = 1;
 806             break;
 807         case 14:
 808             misc_priv_type = strtol(optarg, NULL, 0);
 809             break;
 810         case 15:
 811             misc_priv_value = strtol(optarg, NULL, 0);
 812             break;
 813         case 17:
 814             h264_entropy_mode = atoi(optarg) ? 1: 0;
 815             break;
 816         case 18:
 817             if (strncmp(optarg, "BP", 2) == 0)
 818                 h264_profile = VAProfileH264Baseline;
 819             else if (strncmp(optarg, "MP", 2) == 0)
 820                 h264_profile = VAProfileH264Main;
 821             else if (strncmp(optarg, "HP", 2) == 0)
 822                 h264_profile = VAProfileH264High;
 823             else
 824                 h264_profile = (VAProfile)0;
 825             break;
 826         case ':':
 827         case '?':
 828             print_help();
 829             exit(0);
 830         }
 831     }
 832
 833     if (ip_period < 1) {
 834         printf(" ip_period must be greater than 0\n");
 835         exit(0);
 836     }
 837     if (intra_period != 1 && intra_period % ip_period != 0) {
 838         printf(" intra_period must be a multiplier of ip_period\n");
 839         exit(0);
 840     }
 841     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 842         printf(" intra_idr_period must be a multiplier of intra_period\n");
 843         exit(0);
 844     }
 845
 846     if (frame_bitrate == 0)
 847         frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
 848
 849     if (coded_fn == NULL) {
 850         struct stat buf;
 851         if (stat("/tmp", &buf) == 0)
 852             coded_fn = strdup("/tmp/test.264");
 853         else if (stat("/sdcard", &buf) == 0)
 854             coded_fn = strdup("/sdcard/test.264");
 855         else
 856             coded_fn = strdup("./test.264");
 857     }
 858
 859
 860     frame_width_mbaligned = (frame_width + 15) & (~15);
 861     frame_height_mbaligned = (frame_height + 15) & (~15);
 862     if (frame_width != frame_width_mbaligned ||
 863         frame_height != frame_height_mbaligned) {
 864         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 865                frame_width, frame_height,
 866                frame_width_mbaligned, frame_height_mbaligned
 867                );
 868     }
 869
 870     return 0;
 871 }
 872 #endif
 873
 874 static Display *x11_display;
 875 static Window   x11_window;
 876
 877 VADisplay
 878 va_open_display(void)
 879 {
 880     x11_display = XOpenDisplay(NULL);
 881     if (!x11_display) {
 882         fprintf(stderr, "error: can't connect to X server!\n");
 883         return NULL;
 884     }
 885     return vaGetDisplay(x11_display);
 886 }
 887
 888 void
 889 va_close_display(VADisplay va_dpy)
 890 {
 891     if (!x11_display)
 892         return;
 893
 894     if (x11_window) {
 895         XUnmapWindow(x11_display, x11_window);
 896         XDestroyWindow(x11_display, x11_window);
 897         x11_window = None;
 898     }
 899     XCloseDisplay(x11_display);
 900     x11_display = NULL;
 901 }
 902
 903 static int init_va(void)
 904 {
 905     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 906     VAEntrypoint *entrypoints;
 907     int num_entrypoints, slice_entrypoint;
 908     int support_encode = 0;
 909     int major_ver, minor_ver;
 910     VAStatus va_status;
 911     unsigned int i;
 912
 913     va_dpy = va_open_display();
 914     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 915     CHECK_VASTATUS(va_status, "vaInitialize");
 916
 917     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 918     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 919     if (!entrypoints) {
 920         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 921         exit(1);
 922     }
 923
 924     /* use the highest profile */
 925     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 926         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 927             continue;
 928
 929         h264_profile = profile_list[i];
 930         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 931         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 932             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 933                 support_encode = 1;
 934                 break;
 935             }
 936         }
 937         if (support_encode == 1)
 938             break;
 939     }
 940
 941     if (support_encode == 0) {
 942         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 943         exit(1);
 944     } else {
 945         switch (h264_profile) {
 946             case VAProfileH264Baseline:
 947                 printf("Use profile VAProfileH264Baseline\n");
 948                 ip_period = 1;
 949                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 950                 h264_entropy_mode = 0;
 951                 break;
 952             case VAProfileH264ConstrainedBaseline:
 953                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 954                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 955                 ip_period = 1;
 956                 break;
 957
 958             case VAProfileH264Main:
 959                 printf("Use profile VAProfileH264Main\n");
 960                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 961                 break;
 962
 963             case VAProfileH264High:
 964                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 965                 printf("Use profile VAProfileH264High\n");
 966                 break;
 967             default:
 968                 printf("unknow profile. Set to Baseline");
 969                 h264_profile = VAProfileH264Baseline;
 970                 ip_period = 1;
 971                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 972                 break;
 973         }
 974     }
 975
 976     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 977
 978     /* find out the format for the render target, and rate control mode */
 979     for (i = 0; i < VAConfigAttribTypeMax; i++)
 980         attrib[i].type = (VAConfigAttribType)i;
 981
 982     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 983                                       &attrib[0], VAConfigAttribTypeMax);
 984     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 985     /* check the interested configattrib */
 986     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 987         printf("Not find desired YUV420 RT format\n");
 988         exit(1);
 989     } else {
 990         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 991         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 992         config_attrib_num++;
 993     }
 994
 995     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 996         int tmp = attrib[VAConfigAttribRateControl].value;
 997
 998         printf("Support rate control mode (0x%x):", tmp);
 999
1000         if (tmp & VA_RC_NONE)
1001             printf("NONE ");
1002         if (tmp & VA_RC_CBR)
1003             printf("CBR ");
1004         if (tmp & VA_RC_VBR)
1005             printf("VBR ");
1006         if (tmp & VA_RC_VCM)
1007             printf("VCM ");
1008         if (tmp & VA_RC_CQP)
1009             printf("CQP ");
1010         if (tmp & VA_RC_VBR_CONSTRAINED)
1011             printf("VBR_CONSTRAINED ");
1012
1013         printf("\n");
1014
1015         if (rc_mode == -1 || !(rc_mode & tmp))  {
1016             if (rc_mode != -1) {
1017                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1018             }
1019
1020             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1021                 if (rc_default_modes[i] & tmp) {
1022                     rc_mode = rc_default_modes[i];
1023                     break;
1024                 }
1025             }
1026
1027             printf("RateControl mode: %s\n", rc_to_string(rc_mode));
1028         }
1029
1030         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1031         config_attrib[config_attrib_num].value = rc_mode;
1032         config_attrib_num++;
1033     }
1034
1035
1036     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1037         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1038
1039         printf("Support VAConfigAttribEncPackedHeaders\n");
1040
1041         h264_packedheader = 1;
1042         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1043         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1044
1045         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1046             printf("Support packed sequence headers\n");
1047             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1048         }
1049
1050         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1051             printf("Support packed picture headers\n");
1052             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1053         }
1054
1055         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1056             printf("Support packed slice headers\n");
1057             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1058         }
1059
1060         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1061             printf("Support packed misc headers\n");
1062             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1063         }
1064
1065         enc_packed_header_idx = config_attrib_num;
1066         config_attrib_num++;
1067     }
1068
1069     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1070         int tmp = attrib[VAConfigAttribEncInterlaced].value;
1071
1072         printf("Support VAConfigAttribEncInterlaced\n");
1073
1074         if (tmp & VA_ENC_INTERLACED_FRAME)
1075             printf("support VA_ENC_INTERLACED_FRAME\n");
1076         if (tmp & VA_ENC_INTERLACED_FIELD)
1077             printf("Support VA_ENC_INTERLACED_FIELD\n");
1078         if (tmp & VA_ENC_INTERLACED_MBAFF)
1079             printf("Support VA_ENC_INTERLACED_MBAFF\n");
1080         if (tmp & VA_ENC_INTERLACED_PAFF)
1081             printf("Support VA_ENC_INTERLACED_PAFF\n");
1082
1083         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1084         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1085         config_attrib_num++;
1086     }
1087
1088     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1089         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1090
1091         printf("Support %d RefPicList0 and %d RefPicList1\n",
1092                h264_maxref & 0xffff, (h264_maxref >> 16) & 0xffff );
1093     }
1094
1095     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1096         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1097
1098     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1099         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1100
1101         printf("Support VAConfigAttribEncSliceStructure\n");
1102
1103         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1104             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1105         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1106             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1107         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1108             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1109     }
1110     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1111         printf("Support VAConfigAttribEncMacroblockInfo\n");
1112     }
1113
1114     free(entrypoints);
1115     return 0;
1116 }
1117
1118 static int setup_encode()
1119 {
1120     VAStatus va_status;
1121     VASurfaceID *tmp_surfaceid;
1122     int codedbuf_size, i;
1123     static VASurfaceID src_surface[SURFACE_NUM];
1124     static VASurfaceID ref_surface[SURFACE_NUM];
1125
1126     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1127             &config_attrib[0], config_attrib_num, &config_id);
1128     CHECK_VASTATUS(va_status, "vaCreateConfig");
1129
1130     /* create source surfaces */
1131     va_status = vaCreateSurfaces(va_dpy,
1132                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1133                                  &src_surface[0], SURFACE_NUM,
1134                                  NULL, 0);
1135     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1136
1137     /* create reference surfaces */
1138     va_status = vaCreateSurfaces(va_dpy,
1139                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1140                                  &ref_surface[0], SURFACE_NUM,
1141                                  NULL, 0);
1142     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1143
1144     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1145     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1146     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1147
1148     /* Create a context for this encode pipe */
1149     va_status = vaCreateContext(va_dpy, config_id,
1150                                 frame_width_mbaligned, frame_height_mbaligned,
1151                                 VA_PROGRESSIVE,
1152                                 tmp_surfaceid, 2 * SURFACE_NUM,
1153                                 &context_id);
1154     CHECK_VASTATUS(va_status, "vaCreateContext");
1155     free(tmp_surfaceid);
1156
1157     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1158
1159     for (i = 0; i < SURFACE_NUM; i++) {
1160         /* create coded buffer once for all
1161          * other VA buffers which won't be used again after vaRenderPicture.
1162          * so APP can always vaCreateBuffer for every frame
1163          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1164          * so VA won't maintain the coded buffer
1165          */
1166         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1167                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1168         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1169     }
1170
1171     /* create OpenGL objects */
1172     //glGenFramebuffers(SURFACE_NUM, fbos);
1173
1174     for (i = 0; i < SURFACE_NUM; i++) {
1175         glGenTextures(1, &gl_surfaces[i].y_tex);
1176         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1177     }
1178
1179     for (i = 0; i < SURFACE_NUM; i++) {
1180         gl_surfaces[i].src_surface = src_surface[i];
1181         gl_surfaces[i].ref_surface = ref_surface[i];
1182     }
1183
1184     return 0;
1185 }
1186
1187
1188
1189 #define partition(ref, field, key, ascending)   \
1190     while (i <= j) {                            \
1191         if (ascending) {                        \
1192             while (ref[i].field < key)          \
1193                 i++;                            \
1194             while (ref[j].field > key)          \
1195                 j--;                            \
1196         } else {                                \
1197             while (ref[i].field > key)          \
1198                 i++;                            \
1199             while (ref[j].field < key)          \
1200                 j--;                            \
1201         }                                       \
1202         if (i <= j) {                           \
1203             tmp = ref[i];                       \
1204             ref[i] = ref[j];                    \
1205             ref[j] = tmp;                       \
1206             i++;                                \
1207             j--;                                \
1208         }                                       \
1209     }                                           \
1210
1211 static void sort_one(VAPictureH264 ref[], int left, int right,
1212                      int ascending, int frame_idx)
1213 {
1214     int i = left, j = right;
1215     unsigned int key;
1216     VAPictureH264 tmp;
1217
1218     if (frame_idx) {
1219         key = ref[(left + right) / 2].frame_idx;
1220         partition(ref, frame_idx, key, ascending);
1221     } else {
1222         key = ref[(left + right) / 2].TopFieldOrderCnt;
1223         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1224     }
1225
1226     /* recursion */
1227     if (left < j)
1228         sort_one(ref, left, j, ascending, frame_idx);
1229
1230     if (i < right)
1231         sort_one(ref, i, right, ascending, frame_idx);
1232 }
1233
1234 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1235                      int partition_ascending, int list0_ascending, int list1_ascending)
1236 {
1237     int i = left, j = right;
1238     VAPictureH264 tmp;
1239
1240     if (frame_idx) {
1241         partition(ref, frame_idx, key, partition_ascending);
1242     } else {
1243         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1244     }
1245
1246
1247     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1248     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1249 }
1250
1251 static int update_ReferenceFrames(void)
1252 {
1253     int i;
1254
1255     if (current_frame_type == FRAME_B)
1256         return 0;
1257
1258     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1259     numShortTerm++;
1260     if (numShortTerm > num_ref_frames)
1261         numShortTerm = num_ref_frames;
1262     for (i=numShortTerm-1; i>0; i--)
1263         ReferenceFrames[i] = ReferenceFrames[i-1];
1264     ReferenceFrames[0] = CurrentCurrPic;
1265
1266     if (current_frame_type != FRAME_B)
1267         current_frame_num++;
1268     if (current_frame_num > MaxFrameNum)
1269         current_frame_num = 0;
1270
1271     return 0;
1272 }
1273
1274
1275 static int update_RefPicList(void)
1276 {
1277     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1278
1279     if (current_frame_type == FRAME_P) {
1280         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1281         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1282     }
1283
1284     if (current_frame_type == FRAME_B) {
1285         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1286         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1287                  1, 0, 1);
1288
1289         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1290         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1291                  0, 1, 0);
1292     }
1293
1294     return 0;
1295 }
1296
1297
1298 static int render_sequence(void)
1299 {
1300     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1301     VAStatus va_status;
1302     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1303     VAEncMiscParameterRateControl *misc_rate_ctrl;
1304
1305     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1306     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1307     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1308     seq_param.bits_per_second = frame_bitrate;
1309
1310     seq_param.intra_period = intra_period;
1311     seq_param.intra_idr_period = intra_idr_period;
1312     seq_param.ip_period = ip_period;
1313
1314     seq_param.max_num_ref_frames = num_ref_frames;
1315     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1316     seq_param.time_scale = TIMEBASE * 2;
1317     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1318     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1319     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1320     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1321     seq_param.seq_fields.bits.chroma_format_idc = 1;
1322     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1323
1324     if (frame_width != frame_width_mbaligned ||
1325         frame_height != frame_height_mbaligned) {
1326         seq_param.frame_cropping_flag = 1;
1327         seq_param.frame_crop_left_offset = 0;
1328         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1329         seq_param.frame_crop_top_offset = 0;
1330         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1331     }
1332
1333     va_status = vaCreateBuffer(va_dpy, context_id,
1334                                VAEncSequenceParameterBufferType,
1335                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1336     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1337
1338     va_status = vaCreateBuffer(va_dpy, context_id,
1339                                VAEncMiscParameterBufferType,
1340                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1341                                1, NULL, &rc_param_buf);
1342     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1343
1344     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1345     misc_param->type = VAEncMiscParameterTypeRateControl;
1346     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1347     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1348     misc_rate_ctrl->bits_per_second = frame_bitrate;
1349     misc_rate_ctrl->target_percentage = 66;
1350     misc_rate_ctrl->window_size = 1000;
1351     misc_rate_ctrl->initial_qp = initial_qp;
1352     misc_rate_ctrl->min_qp = minimal_qp;
1353     misc_rate_ctrl->basic_unit_size = 0;
1354     vaUnmapBuffer(va_dpy, rc_param_buf);
1355
1356     render_id[0] = seq_param_buf;
1357     render_id[1] = rc_param_buf;
1358
1359     va_status = vaRenderPicture(va_dpy, context_id, &render_id[0], 2);
1360     CHECK_VASTATUS(va_status, "vaRenderPicture");;
1361
1362     if (misc_priv_type != 0) {
1363         va_status = vaCreateBuffer(va_dpy, context_id,
1364                                    VAEncMiscParameterBufferType,
1365                                    sizeof(VAEncMiscParameterBuffer),
1366                                    1, NULL, &misc_param_tmpbuf);
1367         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1368         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1369         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1370         misc_param_tmp->data[0] = misc_priv_value;
1371         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1372
1373         va_status = vaRenderPicture(va_dpy, context_id, &misc_param_tmpbuf, 1);
1374     }
1375
1376     return 0;
1377 }
1378
1379 static int calc_poc(int pic_order_cnt_lsb)
1380 {
1381     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1382     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1383     int PicOrderCntMsb, TopFieldOrderCnt;
1384
1385     if (current_frame_type == FRAME_IDR)
1386         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1387     else {
1388         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1389         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1390     }
1391
1392     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1393         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1394         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1395     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1396              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1397         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1398     else
1399         PicOrderCntMsb = prevPicOrderCntMsb;
1400
1401     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1402
1403     if (current_frame_type != FRAME_B) {
1404         PicOrderCntMsb_ref = PicOrderCntMsb;
1405         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1406     }
1407
1408     return TopFieldOrderCnt;
1409 }
1410
1411 static int render_picture(void)
1412 {
1413     VABufferID pic_param_buf;
1414     VAStatus va_status;
1415     int i = 0;
1416
1417     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1418     pic_param.CurrPic.frame_idx = current_frame_num;
1419     pic_param.CurrPic.flags = 0;
1420     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1421     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1422     CurrentCurrPic = pic_param.CurrPic;
1423
1424     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1425         update_RefPicList(); /* calc RefPicList */
1426         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1427         if (current_frame_type == FRAME_P) {
1428             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1429         } else if (current_frame_type == FRAME_B) {
1430             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1431             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1432         }
1433     } else {
1434         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1435         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1436             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1437             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1438         }
1439     }
1440
1441     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1442     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1443     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1444     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1445     pic_param.frame_num = current_frame_num;
1446     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1447     pic_param.last_picture = false;  // FIXME
1448     pic_param.pic_init_qp = initial_qp;
1449
1450     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1451                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1452     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1453
1454     va_status = vaRenderPicture(va_dpy, context_id, &pic_param_buf, 1);
1455     CHECK_VASTATUS(va_status, "vaRenderPicture");
1456
1457     return 0;
1458 }
1459
1460 static int render_packedsequence(void)
1461 {
1462     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1463     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1464     unsigned int length_in_bits;
1465     unsigned char *packedseq_buffer = NULL;
1466     VAStatus va_status;
1467
1468     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1469
1470     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1471
1472     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1473     packedheader_param_buffer.has_emulation_bytes = 0;
1474     va_status = vaCreateBuffer(va_dpy,
1475                                context_id,
1476                                VAEncPackedHeaderParameterBufferType,
1477                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1478                                &packedseq_para_bufid);
1479     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1480
1481     va_status = vaCreateBuffer(va_dpy,
1482                                context_id,
1483                                VAEncPackedHeaderDataBufferType,
1484                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1485                                &packedseq_data_bufid);
1486     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1487
1488     render_id[0] = packedseq_para_bufid;
1489     render_id[1] = packedseq_data_bufid;
1490     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1491     CHECK_VASTATUS(va_status, "vaRenderPicture");
1492
1493     free(packedseq_buffer);
1494
1495     return 0;
1496 }
1497
1498
1499 static int render_packedpicture(void)
1500 {
1501     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1502     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1503     unsigned int length_in_bits;
1504     unsigned char *packedpic_buffer = NULL;
1505     VAStatus va_status;
1506
1507     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1508     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1509     packedheader_param_buffer.bit_length = length_in_bits;
1510     packedheader_param_buffer.has_emulation_bytes = 0;
1511
1512     va_status = vaCreateBuffer(va_dpy,
1513                                context_id,
1514                                VAEncPackedHeaderParameterBufferType,
1515                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1516                                &packedpic_para_bufid);
1517     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1518
1519     va_status = vaCreateBuffer(va_dpy,
1520                                context_id,
1521                                VAEncPackedHeaderDataBufferType,
1522                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1523                                &packedpic_data_bufid);
1524     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1525
1526     render_id[0] = packedpic_para_bufid;
1527     render_id[1] = packedpic_data_bufid;
1528     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1529     CHECK_VASTATUS(va_status, "vaRenderPicture");
1530
1531     free(packedpic_buffer);
1532
1533     return 0;
1534 }
1535
1536 static void render_packedslice()
1537 {
1538     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1539     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1540     unsigned int length_in_bits;
1541     unsigned char *packedslice_buffer = NULL;
1542     VAStatus va_status;
1543
1544     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1545     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1546     packedheader_param_buffer.bit_length = length_in_bits;
1547     packedheader_param_buffer.has_emulation_bytes = 0;
1548
1549     va_status = vaCreateBuffer(va_dpy,
1550                                context_id,
1551                                VAEncPackedHeaderParameterBufferType,
1552                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1553                                &packedslice_para_bufid);
1554     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1555
1556     va_status = vaCreateBuffer(va_dpy,
1557                                context_id,
1558                                VAEncPackedHeaderDataBufferType,
1559                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1560                                &packedslice_data_bufid);
1561     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1562
1563     render_id[0] = packedslice_para_bufid;
1564     render_id[1] = packedslice_data_bufid;
1565     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1566     CHECK_VASTATUS(va_status, "vaRenderPicture");
1567
1568     free(packedslice_buffer);
1569 }
1570
1571 static int render_slice(void)
1572 {
1573     VABufferID slice_param_buf;
1574     VAStatus va_status;
1575     int i;
1576
1577     update_RefPicList();
1578
1579     /* one frame, one slice */
1580     slice_param.macroblock_address = 0;
1581     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1582     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1583     if (current_frame_type == FRAME_IDR) {
1584         if (current_frame_encoding != 0)
1585             ++slice_param.idr_pic_id;
1586     } else if (current_frame_type == FRAME_P) {
1587         int refpiclist0_max = h264_maxref & 0xffff;
1588         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1589
1590         for (i = refpiclist0_max; i < 32; i++) {
1591             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1592             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1593         }
1594     } else if (current_frame_type == FRAME_B) {
1595         int refpiclist0_max = h264_maxref & 0xffff;
1596         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1597
1598         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1599         for (i = refpiclist0_max; i < 32; i++) {
1600             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1601             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1602         }
1603
1604         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1605         for (i = refpiclist1_max; i < 32; i++) {
1606             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1607             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1608         }
1609     }
1610
1611     slice_param.slice_alpha_c0_offset_div2 = 0;
1612     slice_param.slice_beta_offset_div2 = 0;
1613     slice_param.direct_spatial_mv_pred_flag = 1;
1614     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1615
1616
1617     if (h264_packedheader &&
1618         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1619         render_packedslice();
1620
1621     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1622                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1623     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1624
1625     va_status = vaRenderPicture(va_dpy, context_id, &slice_param_buf, 1);
1626     CHECK_VASTATUS(va_status, "vaRenderPicture");
1627
1628     return 0;
1629 }
1630
1631
1632
1633 int H264Encoder::save_codeddata(storage_task task)
1634 {
1635     VACodedBufferSegment *buf_list = NULL;
1636     VAStatus va_status;
1637     unsigned int coded_size = 0;
1638
1639     string data;
1640
1641     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / frame_rate);  // So we never get negative dts.
1642
1643     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1644     CHECK_VASTATUS(va_status, "vaMapBuffer");
1645     while (buf_list != NULL) {
1646         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1647         buf_list = (VACodedBufferSegment *) buf_list->next;
1648
1649         frame_size += coded_size;
1650     }
1651     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1652
1653     {
1654         // Add video.
1655         AVPacket pkt;
1656         memset(&pkt, 0, sizeof(pkt));
1657         pkt.buf = nullptr;
1658         pkt.pts = av_rescale_q(task.pts + global_delay, AVRational{1, TIMEBASE}, avstream_video->time_base);
1659         pkt.dts = av_rescale_q(task.dts + global_delay, AVRational{1, TIMEBASE}, avstream_video->time_base);
1660         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1661         pkt.size = data.size();
1662         pkt.stream_index = 0;
1663         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1664             pkt.flags = AV_PKT_FLAG_KEY;
1665         } else {
1666             pkt.flags = 0;
1667         }
1668         //pkt.duration = 1;
1669         av_interleaved_write_frame(avctx, &pkt);
1670     }
1671     // Encode and add all audio frames up to and including the pts of this video frame.
1672     // (They can never be queued to us after the video frame they belong to, only before.)
1673     for ( ;; ) {
1674         int64_t audio_pts;
1675         std::vector<float> audio;
1676         {
1677              unique_lock<mutex> lock(frame_queue_mutex);
1678              if (pending_audio_frames.empty()) break;
1679              auto it = pending_audio_frames.begin();
1680              if (it->first > task.pts) break;
1681              audio_pts = it->first;
1682              audio = move(it->second);
1683              pending_audio_frames.erase(it);
1684         }
1685         AVFrame *frame = avcodec_alloc_frame();
1686         frame->nb_samples = audio.size() / 2;
1687         frame->format = AV_SAMPLE_FMT_FLT;
1688         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1689
1690         unique_ptr<float[]> planar_samples(new float[audio.size()]);
1691         avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_FLTP, (const uint8_t*)planar_samples.get(), audio.size() * sizeof(float), 0);
1692         for (int i = 0; i < frame->nb_samples; ++i) {
1693             planar_samples[i] = audio[i * 2 + 0];
1694             planar_samples[i + frame->nb_samples] = audio[i * 2 + 1];
1695         }
1696
1697         AVPacket pkt;
1698         av_init_packet(&pkt);
1699         pkt.data = nullptr;
1700         pkt.size = 0;
1701         int got_output;
1702         avcodec_encode_audio2(avstream_audio->codec, &pkt, frame, &got_output);
1703         if (got_output) {
1704             pkt.pts = av_rescale_q(audio_pts + global_delay, AVRational{1, TIMEBASE}, avstream_audio->time_base);
1705             pkt.dts = pkt.pts;
1706             pkt.stream_index = 1;
1707             av_interleaved_write_frame(avctx, &pkt);
1708         }
1709         // TODO: Delayed frames.
1710         avcodec_free_frame(&frame);
1711     }
1712
1713 #if 0
1714     printf("\r      "); /* return back to startpoint */
1715     switch (encode_order % 4) {
1716         case 0:
1717             printf("|");
1718             break;
1719         case 1:
1720             printf("/");
1721             break;
1722         case 2:
1723             printf("-");
1724             break;
1725         case 3:
1726             printf("\\");
1727             break;
1728     }
1729     printf("%08lld", encode_order);
1730     printf("(%06d bytes coded)", coded_size);
1731 #endif
1732
1733     return 0;
1734 }
1735
1736
1737 // this is weird. but it seems to put a new frame onto the queue
1738 void H264Encoder::storage_task_enqueue(storage_task task)
1739 {
1740         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1741         storage_task_queue.push(move(task));
1742         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1743         storage_task_queue_changed.notify_all();
1744 }
1745
1746 void H264Encoder::storage_task_thread()
1747 {
1748         for ( ;; ) {
1749                 storage_task current;
1750                 {
1751                         // wait until there's an encoded frame
1752                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1753                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1754                         if (storage_thread_should_quit) return;
1755                         current = move(storage_task_queue.front());
1756                         storage_task_queue.pop();
1757                 }
1758
1759                 VAStatus va_status;
1760
1761                 // waits for data, then saves it to disk.
1762                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1763                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1764                 save_codeddata(move(current));
1765
1766                 {
1767                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1768                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1769                         storage_task_queue_changed.notify_all();
1770                 }
1771         }
1772 }
1773
1774 static int release_encode()
1775 {
1776     int i;
1777
1778     for (i = 0; i < SURFACE_NUM; i++) {
1779         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1780         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1781         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1782     }
1783
1784     vaDestroyContext(va_dpy, context_id);
1785     vaDestroyConfig(va_dpy, config_id);
1786
1787     return 0;
1788 }
1789
1790 static int deinit_va()
1791 {
1792     vaTerminate(va_dpy);
1793
1794     va_close_display(va_dpy);
1795
1796     return 0;
1797 }
1798
1799
1800 static int print_input()
1801 {
1802     printf("\n\nINPUT:Try to encode H264...\n");
1803     if (rc_mode != -1)
1804         printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1805     printf("INPUT: Resolution   : %dx%dframes\n", frame_width, frame_height);
1806     printf("INPUT: FrameRate    : %d\n", frame_rate);
1807     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
1808     printf("INPUT: Slieces      : %d\n", frame_slices);
1809     printf("INPUT: IntraPeriod  : %d\n", intra_period);
1810     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
1811     printf("INPUT: IpPeriod     : %d\n", ip_period);
1812     printf("INPUT: Initial QP   : %d\n", initial_qp);
1813     printf("INPUT: Min QP       : %d\n", minimal_qp);
1814     printf("INPUT: Coded Clip   : %s\n", coded_fn);
1815
1816     printf("\n\n"); /* return back to startpoint */
1817
1818     return 0;
1819 }
1820
1821
1822 //H264Encoder::H264Encoder(SDL_Window *window, SDL_GLContext context, int width, int height, const char *output_filename)
1823 H264Encoder::H264Encoder(QSurface *surface, int width, int height, const char *output_filename)
1824         : current_storage_frame(0), surface(surface)
1825         //: width(width), height(height), current_encoding_frame(0)
1826 {
1827         av_register_all();
1828         avctx = avformat_alloc_context();
1829         avctx->oformat = av_guess_format(NULL, output_filename, NULL);
1830         strcpy(avctx->filename, output_filename);
1831         if (avio_open2(&avctx->pb, output_filename, AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL) < 0) {
1832                 fprintf(stderr, "%s: avio_open2() failed\n", output_filename);
1833                 exit(1);
1834         }
1835         AVCodec *codec_video = avcodec_find_encoder(AV_CODEC_ID_H264);
1836         avstream_video = avformat_new_stream(avctx, codec_video);
1837         if (avstream_video == nullptr) {
1838                 fprintf(stderr, "%s: avformat_new_stream() failed\n", output_filename);
1839                 exit(1);
1840         }
1841         avstream_video->time_base = AVRational{1, TIMEBASE};
1842         avstream_video->codec->width = width;
1843         avstream_video->codec->height = height;
1844         avstream_video->codec->time_base = AVRational{1, TIMEBASE};
1845         avstream_video->codec->ticks_per_frame = 1;  // or 2?
1846
1847         AVCodec *codec_audio = avcodec_find_encoder(AV_CODEC_ID_MP3);
1848         avstream_audio = avformat_new_stream(avctx, codec_audio);
1849         if (avstream_audio == nullptr) {
1850                 fprintf(stderr, "%s: avformat_new_stream() failed\n", output_filename);
1851                 exit(1);
1852         }
1853         avstream_audio->time_base = AVRational{1, TIMEBASE};
1854         avstream_audio->codec->bit_rate = 256000;
1855         avstream_audio->codec->sample_rate = 48000;
1856         avstream_audio->codec->sample_fmt = AV_SAMPLE_FMT_FLTP;
1857         avstream_audio->codec->channels = 2;
1858         avstream_audio->codec->channel_layout = AV_CH_LAYOUT_STEREO;
1859         avstream_audio->codec->time_base = AVRational{1, TIMEBASE};
1860
1861         /* open it */
1862         if (avcodec_open2(avstream_audio->codec, codec_audio, NULL) < 0) {
1863                 fprintf(stderr, "Could not open codec\n");
1864                 exit(1);
1865         }
1866
1867         if (avformat_write_header(avctx, NULL) < 0) {
1868                 fprintf(stderr, "%s: avformat_write_header() failed\n", output_filename);
1869                 exit(1);
1870         }
1871
1872         frame_width = width;
1873         frame_height = height;
1874         frame_width_mbaligned = (frame_width + 15) & (~15);
1875         frame_height_mbaligned = (frame_height + 15) & (~15);
1876         frame_bitrate = 15000000;  // / 60;
1877         current_frame_encoding = 0;
1878
1879         print_input();
1880
1881         init_va();
1882         setup_encode();
1883
1884         // No frames are ready yet.
1885         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1886
1887         memset(&seq_param, 0, sizeof(seq_param));
1888         memset(&pic_param, 0, sizeof(pic_param));
1889         memset(&slice_param, 0, sizeof(slice_param));
1890
1891         storage_thread = std::thread(&H264Encoder::storage_task_thread, this);
1892
1893         copy_thread = std::thread([this]{
1894                 //SDL_GL_MakeCurrent(window, context);
1895                 QOpenGLContext *context = create_context();
1896                 eglBindAPI(EGL_OPENGL_API);
1897                 if (!make_current(context, this->surface)) {
1898                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1899                                 eglGetError());
1900                         exit(1);
1901                 }
1902                 copy_thread_func();
1903         });
1904 }
1905
1906 H264Encoder::~H264Encoder()
1907 {
1908         {
1909                 unique_lock<mutex> lock(storage_task_queue_mutex);
1910                 storage_thread_should_quit = true;
1911                 storage_task_queue_changed.notify_all();
1912         }
1913         {
1914                 unique_lock<mutex> lock(frame_queue_mutex);
1915                 copy_thread_should_quit = true;
1916                 frame_queue_nonempty.notify_one();
1917         }
1918         storage_thread.join();
1919         copy_thread.join();
1920
1921         release_encode();
1922         deinit_va();
1923
1924         av_write_trailer(avctx);
1925         avformat_free_context(avctx);
1926 }
1927
1928 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1929 {
1930         {
1931                 // Wait until this frame slot is done encoding.
1932                 std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1933                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1934                 if (storage_thread_should_quit) return false;
1935         }
1936
1937         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1938         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1939         *y_tex = surf->y_tex;
1940         *cbcr_tex = surf->cbcr_tex;
1941
1942         VASurfaceID surface = surf->src_surface;
1943         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1944         CHECK_VASTATUS(va_status, "vaDeriveImage");
1945
1946         VABufferInfo buf_info;
1947         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1948         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1949         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1950
1951         // Create Y image.
1952         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1953         EGLint y_attribs[] = {
1954                 EGL_WIDTH, frame_width,
1955                 EGL_HEIGHT, frame_height,
1956                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1957                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1958                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1959                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1960                 EGL_NONE
1961         };
1962
1963         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1964         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1965
1966         // Associate Y image to a texture.
1967         glBindTexture(GL_TEXTURE_2D, *y_tex);
1968         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1969
1970         // Create CbCr image.
1971         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1972         EGLint cbcr_attribs[] = {
1973                 EGL_WIDTH, frame_width,
1974                 EGL_HEIGHT, frame_height,
1975                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1976                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1977                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1978                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1979                 EGL_NONE
1980         };
1981
1982         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1983         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1984
1985         // Associate CbCr image to a texture.
1986         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1987         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1988
1989         return true;
1990 }
1991
1992 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
1993 {
1994         {
1995                 unique_lock<mutex> lock(frame_queue_mutex);
1996                 pending_audio_frames[pts] = move(audio);
1997         }
1998         frame_queue_nonempty.notify_one();
1999 }
2000
2001
2002 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
2003 {
2004         {
2005                 unique_lock<mutex> lock(frame_queue_mutex);
2006                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
2007                 ++current_storage_frame;
2008         }
2009         frame_queue_nonempty.notify_one();
2010 }
2011
2012 void H264Encoder::copy_thread_func()
2013 {
2014         int64_t last_dts = -1;
2015         for ( ;; ) {
2016                 PendingFrame frame;
2017                 int pts_lag;
2018                 encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
2019                                        &current_frame_display, &current_frame_type, &pts_lag);
2020                 if (current_frame_type == FRAME_IDR) {
2021                         numShortTerm = 0;
2022                         current_frame_num = 0;
2023                         current_IDR_display = current_frame_display;
2024                 }
2025
2026                 {
2027                         unique_lock<mutex> lock(frame_queue_mutex);
2028                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; });
2029                         if (copy_thread_should_quit) return;
2030                         frame = move(pending_video_frames[current_frame_display]);
2031                         pending_video_frames.erase(current_frame_display);
2032                 }
2033
2034                 // Wait for the GPU to be done with the frame.
2035                 glClientWaitSync(frame.fence.get(), 0, 0);
2036
2037                 // Release back any input frames we needed to render this frame.
2038                 frame.input_frames.clear();
2039
2040                 // Unmap the image.
2041                 GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
2042                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2043                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2044                 VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2045                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2046                 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2047                 CHECK_VASTATUS(va_status, "vaDestroyImage");
2048
2049                 VASurfaceID surface = surf->src_surface;
2050
2051                 // Schedule the frame for encoding.
2052                 va_status = vaBeginPicture(va_dpy, context_id, surface);
2053                 CHECK_VASTATUS(va_status, "vaBeginPicture");
2054
2055                 if (current_frame_type == FRAME_IDR) {
2056                         render_sequence();
2057                         render_picture();
2058                         if (h264_packedheader) {
2059                                 render_packedsequence();
2060                                 render_packedpicture();
2061                         }
2062                 } else {
2063                         //render_sequence();
2064                         render_picture();
2065                 }
2066                 render_slice();
2067
2068                 va_status = vaEndPicture(va_dpy, context_id);
2069                 CHECK_VASTATUS(va_status, "vaEndPicture");
2070
2071                 // Determine the pts and dts of this frame.
2072                 int64_t pts = frame.pts;
2073                 int64_t dts;
2074                 if (pts_lag == -1) {
2075                         assert(last_dts != -1);
2076                         dts = last_dts + (TIMEBASE / frame_rate);
2077                 } else {
2078                         dts = pts - pts_lag;
2079                 }
2080                 last_dts = dts;
2081
2082                 // so now the data is done encoding (well, async job kicked off)...
2083                 // we send that to the storage thread
2084                 storage_task tmp;
2085                 tmp.display_order = current_frame_display;
2086                 tmp.frame_type = current_frame_type;
2087                 tmp.pts = pts;
2088                 tmp.dts = dts;
2089                 storage_task_enqueue(move(tmp));
2090
2091                 update_ReferenceFrames();
2092                 ++current_frame_encoding;
2093         }
2094 }