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 "httpd.h"
  33 #include "timebase.h"
  34
  35 class QOpenGLContext;
  36 class QSurface;
  37
  38 #define CHECK_VASTATUS(va_status, func)                                 \
  39     if (va_status != VA_STATUS_SUCCESS) {                               \
  40         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  41         exit(1);                                                        \
  42     }
  43
  44 //#include "loadsurface.h"
  45
  46 #define NAL_REF_IDC_NONE        0
  47 #define NAL_REF_IDC_LOW         1
  48 #define NAL_REF_IDC_MEDIUM      2
  49 #define NAL_REF_IDC_HIGH        3
  50
  51 #define NAL_NON_IDR             1
  52 #define NAL_IDR                 5
  53 #define NAL_SPS                 7
  54 #define NAL_PPS                 8
  55 #define NAL_SEI                 6
  56
  57 #define SLICE_TYPE_P            0
  58 #define SLICE_TYPE_B            1
  59 #define SLICE_TYPE_I            2
  60 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  61 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  62 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  63
  64
  65 #define ENTROPY_MODE_CAVLC      0
  66 #define ENTROPY_MODE_CABAC      1
  67
  68 #define PROFILE_IDC_BASELINE    66
  69 #define PROFILE_IDC_MAIN        77
  70 #define PROFILE_IDC_HIGH        100
  71
  72 #define BITSTREAM_ALLOCATE_STEPPING     4096
  73
  74 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  75 static  VADisplay va_dpy;
  76 static  VAProfile h264_profile = (VAProfile)~0;
  77 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  78 static  int config_attrib_num = 0, enc_packed_header_idx;
  79
  80 struct GLSurface {
  81         VASurfaceID src_surface, ref_surface;
  82         VABufferID coded_buf;
  83
  84         VAImage surface_image;
  85         GLuint y_tex, cbcr_tex;
  86         EGLImage y_egl_image, cbcr_egl_image;
  87 };
  88 GLSurface gl_surfaces[SURFACE_NUM];
  89
  90 static  VAConfigID config_id;
  91 static  VAContextID context_id;
  92 static  VAEncSequenceParameterBufferH264 seq_param;
  93 static  VAEncPictureParameterBufferH264 pic_param;
  94 static  VAEncSliceParameterBufferH264 slice_param;
  95 static  VAPictureH264 CurrentCurrPic;
  96 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
  97
  98 static  unsigned int MaxFrameNum = (2<<16);
  99 static  unsigned int MaxPicOrderCntLsb = (2<<8);
 100 static  unsigned int Log2MaxFrameNum = 16;
 101 static  unsigned int Log2MaxPicOrderCntLsb = 8;
 102
 103 static  unsigned int num_ref_frames = 2;
 104 static  unsigned int numShortTerm = 0;
 105 static  int constraint_set_flag = 0;
 106 static  int h264_packedheader = 0; /* support pack header? */
 107 static  int h264_maxref = (1<<16|1);
 108 static  int h264_entropy_mode = 1; /* cabac */
 109
 110 static  char *coded_fn = NULL;
 111
 112 static  int frame_width = 176;
 113 static  int frame_height = 144;
 114 static  int frame_width_mbaligned;
 115 static  int frame_height_mbaligned;
 116 static  int frame_rate = 60;
 117 static  unsigned int frame_bitrate = 0;
 118 static  unsigned int frame_slices = 1;
 119 static  double frame_size = 0;
 120 static  int initial_qp = 15;
 121 //static  int initial_qp = 28;
 122 static  int minimal_qp = 0;
 123 static  int intra_period = 30;
 124 static  int intra_idr_period = 60;
 125 static  int ip_period = 3;
 126 static  int rc_mode = -1;
 127 static  int rc_default_modes[] = {
 128     VA_RC_VBR,
 129     VA_RC_CQP,
 130     VA_RC_VBR_CONSTRAINED,
 131     VA_RC_CBR,
 132     VA_RC_VCM,
 133     VA_RC_NONE,
 134 };
 135 static  unsigned long long current_frame_encoding = 0;
 136 static  unsigned long long current_frame_display = 0;
 137 static  unsigned long long current_IDR_display = 0;
 138 static  unsigned int current_frame_num = 0;
 139 static  int current_frame_type;
 140
 141 static  int misc_priv_type = 0;
 142 static  int misc_priv_value = 0;
 143
 144 /* thread to save coded data */
 145 #define SRC_SURFACE_FREE        0
 146 #define SRC_SURFACE_IN_ENCODING 1
 147
 148 struct __bitstream {
 149     unsigned int *buffer;
 150     int bit_offset;
 151     int max_size_in_dword;
 152 };
 153 typedef struct __bitstream bitstream;
 154
 155 using namespace std;
 156
 157 static unsigned int
 158 va_swap32(unsigned int val)
 159 {
 160     unsigned char *pval = (unsigned char *)&val;
 161
 162     return ((pval[0] << 24)     |
 163             (pval[1] << 16)     |
 164             (pval[2] << 8)      |
 165             (pval[3] << 0));
 166 }
 167
 168 static void
 169 bitstream_start(bitstream *bs)
 170 {
 171     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 172     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 173     bs->bit_offset = 0;
 174 }
 175
 176 static void
 177 bitstream_end(bitstream *bs)
 178 {
 179     int pos = (bs->bit_offset >> 5);
 180     int bit_offset = (bs->bit_offset & 0x1f);
 181     int bit_left = 32 - bit_offset;
 182
 183     if (bit_offset) {
 184         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 185     }
 186 }
 187
 188 static void
 189 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 190 {
 191     int pos = (bs->bit_offset >> 5);
 192     int bit_offset = (bs->bit_offset & 0x1f);
 193     int bit_left = 32 - bit_offset;
 194
 195     if (!size_in_bits)
 196         return;
 197
 198     bs->bit_offset += size_in_bits;
 199
 200     if (bit_left > size_in_bits) {
 201         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 202     } else {
 203         size_in_bits -= bit_left;
 204         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 205         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 206
 207         if (pos + 1 == bs->max_size_in_dword) {
 208             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 209             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 210         }
 211
 212         bs->buffer[pos + 1] = val;
 213     }
 214 }
 215
 216 static void
 217 bitstream_put_ue(bitstream *bs, unsigned int val)
 218 {
 219     int size_in_bits = 0;
 220     int tmp_val = ++val;
 221
 222     while (tmp_val) {
 223         tmp_val >>= 1;
 224         size_in_bits++;
 225     }
 226
 227     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 228     bitstream_put_ui(bs, val, size_in_bits);
 229 }
 230
 231 static void
 232 bitstream_put_se(bitstream *bs, int val)
 233 {
 234     unsigned int new_val;
 235
 236     if (val <= 0)
 237         new_val = -2 * val;
 238     else
 239         new_val = 2 * val - 1;
 240
 241     bitstream_put_ue(bs, new_val);
 242 }
 243
 244 static void
 245 bitstream_byte_aligning(bitstream *bs, int bit)
 246 {
 247     int bit_offset = (bs->bit_offset & 0x7);
 248     int bit_left = 8 - bit_offset;
 249     int new_val;
 250
 251     if (!bit_offset)
 252         return;
 253
 254     assert(bit == 0 || bit == 1);
 255
 256     if (bit)
 257         new_val = (1 << bit_left) - 1;
 258     else
 259         new_val = 0;
 260
 261     bitstream_put_ui(bs, new_val, bit_left);
 262 }
 263
 264 static void
 265 rbsp_trailing_bits(bitstream *bs)
 266 {
 267     bitstream_put_ui(bs, 1, 1);
 268     bitstream_byte_aligning(bs, 0);
 269 }
 270
 271 static void nal_start_code_prefix(bitstream *bs)
 272 {
 273     bitstream_put_ui(bs, 0x00000001, 32);
 274 }
 275
 276 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 277 {
 278     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 279     bitstream_put_ui(bs, nal_ref_idc, 2);
 280     bitstream_put_ui(bs, nal_unit_type, 5);
 281 }
 282
 283 static void sps_rbsp(bitstream *bs)
 284 {
 285     int profile_idc = PROFILE_IDC_BASELINE;
 286
 287     if (h264_profile  == VAProfileH264High)
 288         profile_idc = PROFILE_IDC_HIGH;
 289     else if (h264_profile  == VAProfileH264Main)
 290         profile_idc = PROFILE_IDC_MAIN;
 291
 292     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 293     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 294     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 295     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 296     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 297     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 298     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 299     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 300
 301     if ( profile_idc == PROFILE_IDC_HIGH) {
 302         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 303         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 304         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 305         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 306         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 307     }
 308
 309     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 310     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 311
 312     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 313         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 314     else {
 315         assert(0);
 316     }
 317
 318     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 319     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 320
 321     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 322     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 323     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 324
 325     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 326         assert(0);
 327     }
 328
 329     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 330     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 331
 332     if (seq_param.frame_cropping_flag) {
 333         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 334         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 335         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 336         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 337     }
 338
 339     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 340     if ( false ) {
 341         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 342     } else {
 343         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 344         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 345         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 346         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 347         {
 348             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 349             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 350             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 351             {
 352                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 353                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 354                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 355             }
 356         }
 357         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 358         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 359         {
 360             bitstream_put_ui(bs, 1, 32);  // FPS
 361             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 362             bitstream_put_ui(bs, 1, 1);
 363         }
 364         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 365         {
 366             // hrd_parameters
 367             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 368             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 369             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 370
 371             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 372             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 373             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 374
 375             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 376             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 377             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 378             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 379         }
 380         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 381         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 382
 383         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 384         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 385     }
 386
 387     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 388 }
 389
 390
 391 static void pps_rbsp(bitstream *bs)
 392 {
 393     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 394     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 395
 396     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 397
 398     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 399
 400     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 401
 402     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 403     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 404
 405     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 406     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 407
 408     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 409     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 410     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 411
 412     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 413     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 414     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 415
 416     /* more_rbsp_data */
 417     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 418     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 419     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 420
 421     rbsp_trailing_bits(bs);
 422 }
 423
 424 static void slice_header(bitstream *bs)
 425 {
 426     int first_mb_in_slice = slice_param.macroblock_address;
 427
 428     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 429     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 430     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 431     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 432
 433     /* frame_mbs_only_flag == 1 */
 434     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 435         /* FIXME: */
 436         assert(0);
 437     }
 438
 439     if (pic_param.pic_fields.bits.idr_pic_flag)
 440         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 441
 442     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 443         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 444         /* pic_order_present_flag == 0 */
 445     } else {
 446         /* FIXME: */
 447         assert(0);
 448     }
 449
 450     /* redundant_pic_cnt_present_flag == 0 */
 451     /* slice type */
 452     if (IS_P_SLICE(slice_param.slice_type)) {
 453         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 454
 455         if (slice_param.num_ref_idx_active_override_flag)
 456             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 457
 458         /* ref_pic_list_reordering */
 459         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 460     } else if (IS_B_SLICE(slice_param.slice_type)) {
 461         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 462
 463         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 464
 465         if (slice_param.num_ref_idx_active_override_flag) {
 466             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 467             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 468         }
 469
 470         /* ref_pic_list_reordering */
 471         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 472         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 473     }
 474
 475     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 476          IS_P_SLICE(slice_param.slice_type)) ||
 477         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 478          IS_B_SLICE(slice_param.slice_type))) {
 479         /* FIXME: fill weight/offset table */
 480         assert(0);
 481     }
 482
 483     /* dec_ref_pic_marking */
 484     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 485         unsigned char no_output_of_prior_pics_flag = 0;
 486         unsigned char long_term_reference_flag = 0;
 487         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 488
 489         if (pic_param.pic_fields.bits.idr_pic_flag) {
 490             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 491             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 492         } else {
 493             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 494         }
 495     }
 496
 497     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 498         !IS_I_SLICE(slice_param.slice_type))
 499         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 500
 501     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 502
 503     /* ignore for SP/SI */
 504
 505     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 506         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 507
 508         if (slice_param.disable_deblocking_filter_idc != 1) {
 509             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 510             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 511         }
 512     }
 513
 514     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 515         bitstream_byte_aligning(bs, 1);
 516     }
 517 }
 518
 519 static int
 520 build_packed_pic_buffer(unsigned char **header_buffer)
 521 {
 522     bitstream bs;
 523
 524     bitstream_start(&bs);
 525     nal_start_code_prefix(&bs);
 526     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 527     pps_rbsp(&bs);
 528     bitstream_end(&bs);
 529
 530     *header_buffer = (unsigned char *)bs.buffer;
 531     return bs.bit_offset;
 532 }
 533
 534 static int
 535 build_packed_seq_buffer(unsigned char **header_buffer)
 536 {
 537     bitstream bs;
 538
 539     bitstream_start(&bs);
 540     nal_start_code_prefix(&bs);
 541     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 542     sps_rbsp(&bs);
 543     bitstream_end(&bs);
 544
 545     *header_buffer = (unsigned char *)bs.buffer;
 546     return bs.bit_offset;
 547 }
 548
 549 static int build_packed_slice_buffer(unsigned char **header_buffer)
 550 {
 551     bitstream bs;
 552     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 553     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 554
 555     bitstream_start(&bs);
 556     nal_start_code_prefix(&bs);
 557
 558     if (IS_I_SLICE(slice_param.slice_type)) {
 559         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 560     } else if (IS_P_SLICE(slice_param.slice_type)) {
 561         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 562     } else {
 563         assert(IS_B_SLICE(slice_param.slice_type));
 564         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 565     }
 566
 567     slice_header(&bs);
 568     bitstream_end(&bs);
 569
 570     *header_buffer = (unsigned char *)bs.buffer;
 571     return bs.bit_offset;
 572 }
 573
 574
 575 /*
 576   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 577   1) period between Frame #X and Frame #N = #X - #N
 578   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 579   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 580   4) intra_period and intra_idr_period take precedence over ip_period
 581   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 582      of I/IDR frames, see bellow examples
 583   -------------------------------------------------------------------
 584   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 585   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 586   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 587   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 588   1            1                ignored    IDR IDR IDR IDR...
 589   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 590   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 591   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 592                                               (PBB)(IBB)(PBB)(IBB)...
 593   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 594                                            IDRPPPPP IPPPPP IPPPPP...
 595   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 596                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 597                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 598                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 599                                            {IDR(PBB)(PBB)}                     (6/6/3)
 600                                            {IDR(PBB)(PBB)}.
 601 */
 602
 603 // General pts/dts strategy:
 604 //
 605 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 606 // bit tricky. We assume first of all that the frame rate never goes _above_
 607 // <frame_rate>, which gives us a frame period N. The decoder can always decode
 608 // in at least this speed, as long at dts <= pts (the frame is not attempted
 609 // presented before it is decoded). Furthermore, we never have longer chains of
 610 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 611 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 612 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 613 // frame, with an order O <= C.)
 614 //
 615 // Many strategies are possible, but we establish these rules:
 616 //
 617 //  - Tip frames have dts = pts - (C-O)*N.
 618 //  - Non-tip frames have dts = dts_last + N.
 619 //
 620 // An example, with C=2 and N=10 and the data flow showed with arrows:
 621 //
 622 //        I  B  P  B  B  P
 623 //   pts: 30 40 50 60 70 80
 624 //        ↓  ↓     ↓
 625 //   dts: 10 30 20 60 50←40
 626 //         |  |  ↑        ↑
 627 //         `--|--'        |
 628 //             `----------'
 629 //
 630 // To show that this works fine also with irregular spacings, let's say that
 631 // the third frame is delayed a bit (something earlier was dropped). Now the
 632 // situation looks like this:
 633 //
 634 //        I  B  P  B  B   P
 635 //   pts: 30 40 80 90 100 110
 636 //        ↓  ↓     ↓
 637 //   dts: 10 30 20 90 50←40
 638 //         |  |  ↑        ↑
 639 //         `--|--'        |
 640 //             `----------'
 641 //
 642 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 643 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 644 //
 645 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 646 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 647 // a tip frame and should be given a dts based on the previous one.
 648 #define FRAME_P 0
 649 #define FRAME_B 1
 650 #define FRAME_I 2
 651 #define FRAME_IDR 7
 652 void encoding2display_order(
 653     unsigned long long encoding_order, int intra_period,
 654     int intra_idr_period, int ip_period,
 655     unsigned long long *displaying_order,
 656     int *frame_type, int *pts_lag)
 657 {
 658     int encoding_order_gop = 0;
 659
 660     *pts_lag = 0;
 661
 662     if (intra_period == 1) { /* all are I/IDR frames */
 663         *displaying_order = encoding_order;
 664         if (intra_idr_period == 0)
 665             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 666         else
 667             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 668         return;
 669     }
 670
 671     if (intra_period == 0)
 672         intra_idr_period = 0;
 673
 674     if (ip_period == 1) {
 675         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 676         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 677         *displaying_order = encoding_order;
 678
 679         if (encoding_order_gop == 0) { /* the first frame */
 680             *frame_type = FRAME_IDR;
 681         } else if (intra_period != 0 && /* have I frames */
 682                    encoding_order_gop >= 2 &&
 683                    (encoding_order_gop % intra_period == 0)) {
 684             *frame_type = FRAME_I;
 685         } else {
 686             *frame_type = FRAME_P;
 687         }
 688         return;
 689     }
 690
 691     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 692     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 693     *pts_lag = -1;  // Most frames are not tip frames.
 694
 695     if (encoding_order_gop == 0) { /* the first frame */
 696         *frame_type = FRAME_IDR;
 697         *displaying_order = encoding_order;
 698         // IDR frames are a special case; I honestly can't find the logic behind
 699         // why this is the right thing, but it seems to line up nicely in practice :-)
 700         *pts_lag = TIMEBASE / frame_rate;
 701     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 702         *frame_type = FRAME_B;
 703         *displaying_order = encoding_order - 1;
 704         if ((encoding_order_gop % ip_period) == 0) {
 705             *pts_lag = 0;  // Last B-frame.
 706         }
 707     } else if (intra_period != 0 && /* have I frames */
 708                encoding_order_gop >= 2 &&
 709                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 710         *frame_type = FRAME_I;
 711         *displaying_order = encoding_order + ip_period - 1;
 712     } else {
 713         *frame_type = FRAME_P;
 714         *displaying_order = encoding_order + ip_period - 1;
 715     }
 716 }
 717
 718
 719 static const char *rc_to_string(int rcmode)
 720 {
 721     switch (rc_mode) {
 722     case VA_RC_NONE:
 723         return "NONE";
 724     case VA_RC_CBR:
 725         return "CBR";
 726     case VA_RC_VBR:
 727         return "VBR";
 728     case VA_RC_VCM:
 729         return "VCM";
 730     case VA_RC_CQP:
 731         return "CQP";
 732     case VA_RC_VBR_CONSTRAINED:
 733         return "VBR_CONSTRAINED";
 734     default:
 735         return "Unknown";
 736     }
 737 }
 738
 739 #if 0
 740 static int process_cmdline(int argc, char *argv[])
 741 {
 742     char c;
 743     const struct option long_opts[] = {
 744         {"help", no_argument, NULL, 0 },
 745         {"bitrate", required_argument, NULL, 1 },
 746         {"minqp", required_argument, NULL, 2 },
 747         {"initialqp", required_argument, NULL, 3 },
 748         {"intra_period", required_argument, NULL, 4 },
 749         {"idr_period", required_argument, NULL, 5 },
 750         {"ip_period", required_argument, NULL, 6 },
 751         {"rcmode", required_argument, NULL, 7 },
 752         {"srcyuv", required_argument, NULL, 9 },
 753         {"recyuv", required_argument, NULL, 10 },
 754         {"fourcc", required_argument, NULL, 11 },
 755         {"syncmode", no_argument, NULL, 12 },
 756         {"enablePSNR", no_argument, NULL, 13 },
 757         {"prit", required_argument, NULL, 14 },
 758         {"priv", required_argument, NULL, 15 },
 759         {"framecount", required_argument, NULL, 16 },
 760         {"entropy", required_argument, NULL, 17 },
 761         {"profile", required_argument, NULL, 18 },
 762         {NULL, no_argument, NULL, 0 }};
 763     int long_index;
 764
 765     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 766         switch (c) {
 767         case 'w':
 768             frame_width = atoi(optarg);
 769             break;
 770         case 'h':
 771             frame_height = atoi(optarg);
 772             break;
 773         case 'n':
 774         case 'f':
 775             frame_rate = atoi(optarg);
 776             break;
 777         case 'o':
 778             coded_fn = strdup(optarg);
 779             break;
 780         case 0:
 781             print_help();
 782             exit(0);
 783         case 1:
 784             frame_bitrate = atoi(optarg);
 785             break;
 786         case 2:
 787             minimal_qp = atoi(optarg);
 788             break;
 789         case 3:
 790             initial_qp = atoi(optarg);
 791             break;
 792         case 4:
 793             intra_period = atoi(optarg);
 794             break;
 795         case 5:
 796             intra_idr_period = atoi(optarg);
 797             break;
 798         case 6:
 799             ip_period = atoi(optarg);
 800             break;
 801         case 7:
 802             rc_mode = string_to_rc(optarg);
 803             if (rc_mode < 0) {
 804                 print_help();
 805                 exit(1);
 806             }
 807             break;
 808         case 9:
 809             srcyuv_fn = strdup(optarg);
 810             break;
 811         case 11:
 812             srcyuv_fourcc = string_to_fourcc(optarg);
 813             if (srcyuv_fourcc <= 0) {
 814                 print_help();
 815                 exit(1);
 816             }
 817             break;
 818         case 13:
 819             calc_psnr = 1;
 820             break;
 821         case 14:
 822             misc_priv_type = strtol(optarg, NULL, 0);
 823             break;
 824         case 15:
 825             misc_priv_value = strtol(optarg, NULL, 0);
 826             break;
 827         case 17:
 828             h264_entropy_mode = atoi(optarg) ? 1: 0;
 829             break;
 830         case 18:
 831             if (strncmp(optarg, "BP", 2) == 0)
 832                 h264_profile = VAProfileH264Baseline;
 833             else if (strncmp(optarg, "MP", 2) == 0)
 834                 h264_profile = VAProfileH264Main;
 835             else if (strncmp(optarg, "HP", 2) == 0)
 836                 h264_profile = VAProfileH264High;
 837             else
 838                 h264_profile = (VAProfile)0;
 839             break;
 840         case ':':
 841         case '?':
 842             print_help();
 843             exit(0);
 844         }
 845     }
 846
 847     if (ip_period < 1) {
 848         printf(" ip_period must be greater than 0\n");
 849         exit(0);
 850     }
 851     if (intra_period != 1 && intra_period % ip_period != 0) {
 852         printf(" intra_period must be a multiplier of ip_period\n");
 853         exit(0);
 854     }
 855     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 856         printf(" intra_idr_period must be a multiplier of intra_period\n");
 857         exit(0);
 858     }
 859
 860     if (frame_bitrate == 0)
 861         frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
 862
 863     if (coded_fn == NULL) {
 864         struct stat buf;
 865         if (stat("/tmp", &buf) == 0)
 866             coded_fn = strdup("/tmp/test.264");
 867         else if (stat("/sdcard", &buf) == 0)
 868             coded_fn = strdup("/sdcard/test.264");
 869         else
 870             coded_fn = strdup("./test.264");
 871     }
 872
 873
 874     frame_width_mbaligned = (frame_width + 15) & (~15);
 875     frame_height_mbaligned = (frame_height + 15) & (~15);
 876     if (frame_width != frame_width_mbaligned ||
 877         frame_height != frame_height_mbaligned) {
 878         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 879                frame_width, frame_height,
 880                frame_width_mbaligned, frame_height_mbaligned
 881                );
 882     }
 883
 884     return 0;
 885 }
 886 #endif
 887
 888 static Display *x11_display;
 889 static Window   x11_window;
 890
 891 VADisplay
 892 va_open_display(void)
 893 {
 894     x11_display = XOpenDisplay(NULL);
 895     if (!x11_display) {
 896         fprintf(stderr, "error: can't connect to X server!\n");
 897         return NULL;
 898     }
 899     return vaGetDisplay(x11_display);
 900 }
 901
 902 void
 903 va_close_display(VADisplay va_dpy)
 904 {
 905     if (!x11_display)
 906         return;
 907
 908     if (x11_window) {
 909         XUnmapWindow(x11_display, x11_window);
 910         XDestroyWindow(x11_display, x11_window);
 911         x11_window = None;
 912     }
 913     XCloseDisplay(x11_display);
 914     x11_display = NULL;
 915 }
 916
 917 static int init_va(void)
 918 {
 919     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 920     VAEntrypoint *entrypoints;
 921     int num_entrypoints, slice_entrypoint;
 922     int support_encode = 0;
 923     int major_ver, minor_ver;
 924     VAStatus va_status;
 925     unsigned int i;
 926
 927     va_dpy = va_open_display();
 928     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 929     CHECK_VASTATUS(va_status, "vaInitialize");
 930
 931     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 932     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 933     if (!entrypoints) {
 934         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 935         exit(1);
 936     }
 937
 938     /* use the highest profile */
 939     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 940         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 941             continue;
 942
 943         h264_profile = profile_list[i];
 944         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 945         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 946             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 947                 support_encode = 1;
 948                 break;
 949             }
 950         }
 951         if (support_encode == 1)
 952             break;
 953     }
 954
 955     if (support_encode == 0) {
 956         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 957         exit(1);
 958     } else {
 959         switch (h264_profile) {
 960             case VAProfileH264Baseline:
 961                 printf("Use profile VAProfileH264Baseline\n");
 962                 ip_period = 1;
 963                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 964                 h264_entropy_mode = 0;
 965                 break;
 966             case VAProfileH264ConstrainedBaseline:
 967                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 968                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 969                 ip_period = 1;
 970                 break;
 971
 972             case VAProfileH264Main:
 973                 printf("Use profile VAProfileH264Main\n");
 974                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 975                 break;
 976
 977             case VAProfileH264High:
 978                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 979                 printf("Use profile VAProfileH264High\n");
 980                 break;
 981             default:
 982                 printf("unknow profile. Set to Baseline");
 983                 h264_profile = VAProfileH264Baseline;
 984                 ip_period = 1;
 985                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 986                 break;
 987         }
 988     }
 989
 990     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 991
 992     /* find out the format for the render target, and rate control mode */
 993     for (i = 0; i < VAConfigAttribTypeMax; i++)
 994         attrib[i].type = (VAConfigAttribType)i;
 995
 996     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 997                                       &attrib[0], VAConfigAttribTypeMax);
 998     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 999     /* check the interested configattrib */
1000     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1001         printf("Not find desired YUV420 RT format\n");
1002         exit(1);
1003     } else {
1004         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1005         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1006         config_attrib_num++;
1007     }
1008
1009     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1010         int tmp = attrib[VAConfigAttribRateControl].value;
1011
1012         printf("Support rate control mode (0x%x):", tmp);
1013
1014         if (tmp & VA_RC_NONE)
1015             printf("NONE ");
1016         if (tmp & VA_RC_CBR)
1017             printf("CBR ");
1018         if (tmp & VA_RC_VBR)
1019             printf("VBR ");
1020         if (tmp & VA_RC_VCM)
1021             printf("VCM ");
1022         if (tmp & VA_RC_CQP)
1023             printf("CQP ");
1024         if (tmp & VA_RC_VBR_CONSTRAINED)
1025             printf("VBR_CONSTRAINED ");
1026
1027         printf("\n");
1028
1029         if (rc_mode == -1 || !(rc_mode & tmp))  {
1030             if (rc_mode != -1) {
1031                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1032             }
1033
1034             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1035                 if (rc_default_modes[i] & tmp) {
1036                     rc_mode = rc_default_modes[i];
1037                     break;
1038                 }
1039             }
1040
1041             printf("RateControl mode: %s\n", rc_to_string(rc_mode));
1042         }
1043
1044         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1045         config_attrib[config_attrib_num].value = rc_mode;
1046         config_attrib_num++;
1047     }
1048
1049
1050     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1051         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1052
1053         printf("Support VAConfigAttribEncPackedHeaders\n");
1054
1055         h264_packedheader = 1;
1056         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1057         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1058
1059         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1060             printf("Support packed sequence headers\n");
1061             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1062         }
1063
1064         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1065             printf("Support packed picture headers\n");
1066             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1067         }
1068
1069         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1070             printf("Support packed slice headers\n");
1071             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1072         }
1073
1074         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1075             printf("Support packed misc headers\n");
1076             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1077         }
1078
1079         enc_packed_header_idx = config_attrib_num;
1080         config_attrib_num++;
1081     }
1082
1083     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1084         int tmp = attrib[VAConfigAttribEncInterlaced].value;
1085
1086         printf("Support VAConfigAttribEncInterlaced\n");
1087
1088         if (tmp & VA_ENC_INTERLACED_FRAME)
1089             printf("support VA_ENC_INTERLACED_FRAME\n");
1090         if (tmp & VA_ENC_INTERLACED_FIELD)
1091             printf("Support VA_ENC_INTERLACED_FIELD\n");
1092         if (tmp & VA_ENC_INTERLACED_MBAFF)
1093             printf("Support VA_ENC_INTERLACED_MBAFF\n");
1094         if (tmp & VA_ENC_INTERLACED_PAFF)
1095             printf("Support VA_ENC_INTERLACED_PAFF\n");
1096
1097         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1098         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1099         config_attrib_num++;
1100     }
1101
1102     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1103         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1104
1105         printf("Support %d RefPicList0 and %d RefPicList1\n",
1106                h264_maxref & 0xffff, (h264_maxref >> 16) & 0xffff );
1107     }
1108
1109     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1110         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1111
1112     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1113         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1114
1115         printf("Support VAConfigAttribEncSliceStructure\n");
1116
1117         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1118             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1119         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1120             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1121         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1122             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1123     }
1124     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1125         printf("Support VAConfigAttribEncMacroblockInfo\n");
1126     }
1127
1128     free(entrypoints);
1129     return 0;
1130 }
1131
1132 static int setup_encode()
1133 {
1134     VAStatus va_status;
1135     VASurfaceID *tmp_surfaceid;
1136     int codedbuf_size, i;
1137     static VASurfaceID src_surface[SURFACE_NUM];
1138     static VASurfaceID ref_surface[SURFACE_NUM];
1139
1140     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1141             &config_attrib[0], config_attrib_num, &config_id);
1142     CHECK_VASTATUS(va_status, "vaCreateConfig");
1143
1144     /* create source surfaces */
1145     va_status = vaCreateSurfaces(va_dpy,
1146                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1147                                  &src_surface[0], SURFACE_NUM,
1148                                  NULL, 0);
1149     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1150
1151     /* create reference surfaces */
1152     va_status = vaCreateSurfaces(va_dpy,
1153                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1154                                  &ref_surface[0], SURFACE_NUM,
1155                                  NULL, 0);
1156     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1157
1158     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1159     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1160     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1161
1162     /* Create a context for this encode pipe */
1163     va_status = vaCreateContext(va_dpy, config_id,
1164                                 frame_width_mbaligned, frame_height_mbaligned,
1165                                 VA_PROGRESSIVE,
1166                                 tmp_surfaceid, 2 * SURFACE_NUM,
1167                                 &context_id);
1168     CHECK_VASTATUS(va_status, "vaCreateContext");
1169     free(tmp_surfaceid);
1170
1171     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1172
1173     for (i = 0; i < SURFACE_NUM; i++) {
1174         /* create coded buffer once for all
1175          * other VA buffers which won't be used again after vaRenderPicture.
1176          * so APP can always vaCreateBuffer for every frame
1177          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1178          * so VA won't maintain the coded buffer
1179          */
1180         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1181                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1182         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1183     }
1184
1185     /* create OpenGL objects */
1186     //glGenFramebuffers(SURFACE_NUM, fbos);
1187
1188     for (i = 0; i < SURFACE_NUM; i++) {
1189         glGenTextures(1, &gl_surfaces[i].y_tex);
1190         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1191     }
1192
1193     for (i = 0; i < SURFACE_NUM; i++) {
1194         gl_surfaces[i].src_surface = src_surface[i];
1195         gl_surfaces[i].ref_surface = ref_surface[i];
1196     }
1197
1198     return 0;
1199 }
1200
1201
1202
1203 #define partition(ref, field, key, ascending)   \
1204     while (i <= j) {                            \
1205         if (ascending) {                        \
1206             while (ref[i].field < key)          \
1207                 i++;                            \
1208             while (ref[j].field > key)          \
1209                 j--;                            \
1210         } else {                                \
1211             while (ref[i].field > key)          \
1212                 i++;                            \
1213             while (ref[j].field < key)          \
1214                 j--;                            \
1215         }                                       \
1216         if (i <= j) {                           \
1217             tmp = ref[i];                       \
1218             ref[i] = ref[j];                    \
1219             ref[j] = tmp;                       \
1220             i++;                                \
1221             j--;                                \
1222         }                                       \
1223     }                                           \
1224
1225 static void sort_one(VAPictureH264 ref[], int left, int right,
1226                      int ascending, int frame_idx)
1227 {
1228     int i = left, j = right;
1229     unsigned int key;
1230     VAPictureH264 tmp;
1231
1232     if (frame_idx) {
1233         key = ref[(left + right) / 2].frame_idx;
1234         partition(ref, frame_idx, key, ascending);
1235     } else {
1236         key = ref[(left + right) / 2].TopFieldOrderCnt;
1237         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1238     }
1239
1240     /* recursion */
1241     if (left < j)
1242         sort_one(ref, left, j, ascending, frame_idx);
1243
1244     if (i < right)
1245         sort_one(ref, i, right, ascending, frame_idx);
1246 }
1247
1248 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1249                      int partition_ascending, int list0_ascending, int list1_ascending)
1250 {
1251     int i = left, j = right;
1252     VAPictureH264 tmp;
1253
1254     if (frame_idx) {
1255         partition(ref, frame_idx, key, partition_ascending);
1256     } else {
1257         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1258     }
1259
1260
1261     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1262     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1263 }
1264
1265 static int update_ReferenceFrames(void)
1266 {
1267     int i;
1268
1269     if (current_frame_type == FRAME_B)
1270         return 0;
1271
1272     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1273     numShortTerm++;
1274     if (numShortTerm > num_ref_frames)
1275         numShortTerm = num_ref_frames;
1276     for (i=numShortTerm-1; i>0; i--)
1277         ReferenceFrames[i] = ReferenceFrames[i-1];
1278     ReferenceFrames[0] = CurrentCurrPic;
1279
1280     if (current_frame_type != FRAME_B)
1281         current_frame_num++;
1282     if (current_frame_num > MaxFrameNum)
1283         current_frame_num = 0;
1284
1285     return 0;
1286 }
1287
1288
1289 static int update_RefPicList(void)
1290 {
1291     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1292
1293     if (current_frame_type == FRAME_P) {
1294         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1295         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1296     }
1297
1298     if (current_frame_type == FRAME_B) {
1299         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1300         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1301                  1, 0, 1);
1302
1303         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1304         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1305                  0, 1, 0);
1306     }
1307
1308     return 0;
1309 }
1310
1311
1312 static int render_sequence(void)
1313 {
1314     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1315     VAStatus va_status;
1316     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1317     VAEncMiscParameterRateControl *misc_rate_ctrl;
1318
1319     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1320     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1321     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1322     seq_param.bits_per_second = frame_bitrate;
1323
1324     seq_param.intra_period = intra_period;
1325     seq_param.intra_idr_period = intra_idr_period;
1326     seq_param.ip_period = ip_period;
1327
1328     seq_param.max_num_ref_frames = num_ref_frames;
1329     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1330     seq_param.time_scale = TIMEBASE * 2;
1331     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1332     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1333     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1334     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1335     seq_param.seq_fields.bits.chroma_format_idc = 1;
1336     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1337
1338     if (frame_width != frame_width_mbaligned ||
1339         frame_height != frame_height_mbaligned) {
1340         seq_param.frame_cropping_flag = 1;
1341         seq_param.frame_crop_left_offset = 0;
1342         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1343         seq_param.frame_crop_top_offset = 0;
1344         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1345     }
1346
1347     va_status = vaCreateBuffer(va_dpy, context_id,
1348                                VAEncSequenceParameterBufferType,
1349                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1350     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1351
1352     va_status = vaCreateBuffer(va_dpy, context_id,
1353                                VAEncMiscParameterBufferType,
1354                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1355                                1, NULL, &rc_param_buf);
1356     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1357
1358     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1359     misc_param->type = VAEncMiscParameterTypeRateControl;
1360     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1361     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1362     misc_rate_ctrl->bits_per_second = frame_bitrate;
1363     misc_rate_ctrl->target_percentage = 66;
1364     misc_rate_ctrl->window_size = 1000;
1365     misc_rate_ctrl->initial_qp = initial_qp;
1366     misc_rate_ctrl->min_qp = minimal_qp;
1367     misc_rate_ctrl->basic_unit_size = 0;
1368     vaUnmapBuffer(va_dpy, rc_param_buf);
1369
1370     render_id[0] = seq_param_buf;
1371     render_id[1] = rc_param_buf;
1372
1373     va_status = vaRenderPicture(va_dpy, context_id, &render_id[0], 2);
1374     CHECK_VASTATUS(va_status, "vaRenderPicture");;
1375
1376     if (misc_priv_type != 0) {
1377         va_status = vaCreateBuffer(va_dpy, context_id,
1378                                    VAEncMiscParameterBufferType,
1379                                    sizeof(VAEncMiscParameterBuffer),
1380                                    1, NULL, &misc_param_tmpbuf);
1381         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1382         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1383         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1384         misc_param_tmp->data[0] = misc_priv_value;
1385         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1386
1387         va_status = vaRenderPicture(va_dpy, context_id, &misc_param_tmpbuf, 1);
1388     }
1389
1390     return 0;
1391 }
1392
1393 static int calc_poc(int pic_order_cnt_lsb)
1394 {
1395     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1396     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1397     int PicOrderCntMsb, TopFieldOrderCnt;
1398
1399     if (current_frame_type == FRAME_IDR)
1400         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1401     else {
1402         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1403         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1404     }
1405
1406     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1407         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1408         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1409     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1410              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1411         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1412     else
1413         PicOrderCntMsb = prevPicOrderCntMsb;
1414
1415     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1416
1417     if (current_frame_type != FRAME_B) {
1418         PicOrderCntMsb_ref = PicOrderCntMsb;
1419         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1420     }
1421
1422     return TopFieldOrderCnt;
1423 }
1424
1425 static int render_picture(void)
1426 {
1427     VABufferID pic_param_buf;
1428     VAStatus va_status;
1429     int i = 0;
1430
1431     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1432     pic_param.CurrPic.frame_idx = current_frame_num;
1433     pic_param.CurrPic.flags = 0;
1434     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1435     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1436     CurrentCurrPic = pic_param.CurrPic;
1437
1438     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1439         update_RefPicList(); /* calc RefPicList */
1440         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1441         if (current_frame_type == FRAME_P) {
1442             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1443         } else if (current_frame_type == FRAME_B) {
1444             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1445             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1446         }
1447     } else {
1448         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1449         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1450             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1451             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1452         }
1453     }
1454
1455     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1456     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1457     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1458     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1459     pic_param.frame_num = current_frame_num;
1460     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1461     pic_param.last_picture = false;  // FIXME
1462     pic_param.pic_init_qp = initial_qp;
1463
1464     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1465                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1466     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1467
1468     va_status = vaRenderPicture(va_dpy, context_id, &pic_param_buf, 1);
1469     CHECK_VASTATUS(va_status, "vaRenderPicture");
1470
1471     return 0;
1472 }
1473
1474 static int render_packedsequence(void)
1475 {
1476     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1477     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1478     unsigned int length_in_bits;
1479     unsigned char *packedseq_buffer = NULL;
1480     VAStatus va_status;
1481
1482     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1483
1484     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1485
1486     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1487     packedheader_param_buffer.has_emulation_bytes = 0;
1488     va_status = vaCreateBuffer(va_dpy,
1489                                context_id,
1490                                VAEncPackedHeaderParameterBufferType,
1491                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1492                                &packedseq_para_bufid);
1493     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1494
1495     va_status = vaCreateBuffer(va_dpy,
1496                                context_id,
1497                                VAEncPackedHeaderDataBufferType,
1498                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1499                                &packedseq_data_bufid);
1500     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1501
1502     render_id[0] = packedseq_para_bufid;
1503     render_id[1] = packedseq_data_bufid;
1504     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1505     CHECK_VASTATUS(va_status, "vaRenderPicture");
1506
1507     free(packedseq_buffer);
1508
1509     return 0;
1510 }
1511
1512
1513 static int render_packedpicture(void)
1514 {
1515     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1516     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1517     unsigned int length_in_bits;
1518     unsigned char *packedpic_buffer = NULL;
1519     VAStatus va_status;
1520
1521     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1522     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1523     packedheader_param_buffer.bit_length = length_in_bits;
1524     packedheader_param_buffer.has_emulation_bytes = 0;
1525
1526     va_status = vaCreateBuffer(va_dpy,
1527                                context_id,
1528                                VAEncPackedHeaderParameterBufferType,
1529                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1530                                &packedpic_para_bufid);
1531     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1532
1533     va_status = vaCreateBuffer(va_dpy,
1534                                context_id,
1535                                VAEncPackedHeaderDataBufferType,
1536                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1537                                &packedpic_data_bufid);
1538     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1539
1540     render_id[0] = packedpic_para_bufid;
1541     render_id[1] = packedpic_data_bufid;
1542     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1543     CHECK_VASTATUS(va_status, "vaRenderPicture");
1544
1545     free(packedpic_buffer);
1546
1547     return 0;
1548 }
1549
1550 static void render_packedslice()
1551 {
1552     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1553     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1554     unsigned int length_in_bits;
1555     unsigned char *packedslice_buffer = NULL;
1556     VAStatus va_status;
1557
1558     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1559     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1560     packedheader_param_buffer.bit_length = length_in_bits;
1561     packedheader_param_buffer.has_emulation_bytes = 0;
1562
1563     va_status = vaCreateBuffer(va_dpy,
1564                                context_id,
1565                                VAEncPackedHeaderParameterBufferType,
1566                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1567                                &packedslice_para_bufid);
1568     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1569
1570     va_status = vaCreateBuffer(va_dpy,
1571                                context_id,
1572                                VAEncPackedHeaderDataBufferType,
1573                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1574                                &packedslice_data_bufid);
1575     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1576
1577     render_id[0] = packedslice_para_bufid;
1578     render_id[1] = packedslice_data_bufid;
1579     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1580     CHECK_VASTATUS(va_status, "vaRenderPicture");
1581
1582     free(packedslice_buffer);
1583 }
1584
1585 static int render_slice(void)
1586 {
1587     VABufferID slice_param_buf;
1588     VAStatus va_status;
1589     int i;
1590
1591     update_RefPicList();
1592
1593     /* one frame, one slice */
1594     slice_param.macroblock_address = 0;
1595     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1596     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1597     if (current_frame_type == FRAME_IDR) {
1598         if (current_frame_encoding != 0)
1599             ++slice_param.idr_pic_id;
1600     } else if (current_frame_type == FRAME_P) {
1601         int refpiclist0_max = h264_maxref & 0xffff;
1602         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1603
1604         for (i = refpiclist0_max; i < 32; i++) {
1605             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1606             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1607         }
1608     } else if (current_frame_type == FRAME_B) {
1609         int refpiclist0_max = h264_maxref & 0xffff;
1610         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1611
1612         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1613         for (i = refpiclist0_max; i < 32; i++) {
1614             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1615             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1616         }
1617
1618         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1619         for (i = refpiclist1_max; i < 32; i++) {
1620             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1621             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1622         }
1623     }
1624
1625     slice_param.slice_alpha_c0_offset_div2 = 0;
1626     slice_param.slice_beta_offset_div2 = 0;
1627     slice_param.direct_spatial_mv_pred_flag = 1;
1628     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1629
1630
1631     if (h264_packedheader &&
1632         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1633         render_packedslice();
1634
1635     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1636                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1637     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1638
1639     va_status = vaRenderPicture(va_dpy, context_id, &slice_param_buf, 1);
1640     CHECK_VASTATUS(va_status, "vaRenderPicture");
1641
1642     return 0;
1643 }
1644
1645
1646
1647 int H264Encoder::save_codeddata(storage_task task)
1648 {
1649     VACodedBufferSegment *buf_list = NULL;
1650     VAStatus va_status;
1651     unsigned int coded_size = 0;
1652
1653     string data;
1654
1655     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / frame_rate);  // So we never get negative dts.
1656
1657     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1658     CHECK_VASTATUS(va_status, "vaMapBuffer");
1659     while (buf_list != NULL) {
1660         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1661         buf_list = (VACodedBufferSegment *) buf_list->next;
1662
1663         frame_size += coded_size;
1664     }
1665     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1666
1667     {
1668         // Add video.
1669         AVPacket pkt;
1670         memset(&pkt, 0, sizeof(pkt));
1671         pkt.buf = nullptr;
1672         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1673         pkt.size = data.size();
1674         pkt.stream_index = 0;
1675         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1676             pkt.flags = AV_PKT_FLAG_KEY;
1677         } else {
1678             pkt.flags = 0;
1679         }
1680         //pkt.duration = 1;
1681         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1682     }
1683     // Encode and add all audio frames up to and including the pts of this video frame.
1684     // (They can never be queued to us after the video frame they belong to, only before.)
1685     for ( ;; ) {
1686         int64_t audio_pts;
1687         std::vector<float> audio;
1688         {
1689              unique_lock<mutex> lock(frame_queue_mutex);
1690              if (pending_audio_frames.empty()) break;
1691              auto it = pending_audio_frames.begin();
1692              if (it->first > task.pts) break;
1693              audio_pts = it->first;
1694              audio = move(it->second);
1695              pending_audio_frames.erase(it);
1696         }
1697
1698         AVFrame *frame = avcodec_alloc_frame();
1699         frame->nb_samples = audio.size() / 2;
1700         frame->format = AV_SAMPLE_FMT_FLT;
1701         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1702
1703         unique_ptr<float[]> planar_samples(new float[audio.size()]);
1704         avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_FLTP, (const uint8_t*)planar_samples.get(), audio.size() * sizeof(float), 0);
1705         for (int i = 0; i < frame->nb_samples; ++i) {
1706             planar_samples[i] = audio[i * 2 + 0];
1707             planar_samples[i + frame->nb_samples] = audio[i * 2 + 1];
1708         }
1709
1710         AVPacket pkt;
1711         av_init_packet(&pkt);
1712         pkt.data = nullptr;
1713         pkt.size = 0;
1714         int got_output;
1715         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1716         if (got_output) {
1717             pkt.stream_index = 1;
1718             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1719         }
1720         // TODO: Delayed frames.
1721         avcodec_free_frame(&frame);
1722     }
1723
1724 #if 0
1725     printf("\r      "); /* return back to startpoint */
1726     switch (encode_order % 4) {
1727         case 0:
1728             printf("|");
1729             break;
1730         case 1:
1731             printf("/");
1732             break;
1733         case 2:
1734             printf("-");
1735             break;
1736         case 3:
1737             printf("\\");
1738             break;
1739     }
1740     printf("%08lld", encode_order);
1741     printf("(%06d bytes coded)", coded_size);
1742 #endif
1743
1744     return 0;
1745 }
1746
1747
1748 // this is weird. but it seems to put a new frame onto the queue
1749 void H264Encoder::storage_task_enqueue(storage_task task)
1750 {
1751         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1752         storage_task_queue.push(move(task));
1753         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1754         storage_task_queue_changed.notify_all();
1755 }
1756
1757 void H264Encoder::storage_task_thread()
1758 {
1759         for ( ;; ) {
1760                 storage_task current;
1761                 {
1762                         // wait until there's an encoded frame
1763                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1764                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1765                         if (storage_thread_should_quit) return;
1766                         current = move(storage_task_queue.front());
1767                         storage_task_queue.pop();
1768                 }
1769
1770                 VAStatus va_status;
1771
1772                 // waits for data, then saves it to disk.
1773                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1774                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1775                 save_codeddata(move(current));
1776
1777                 {
1778                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1779                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1780                         storage_task_queue_changed.notify_all();
1781                 }
1782         }
1783 }
1784
1785 static int release_encode()
1786 {
1787     int i;
1788
1789     for (i = 0; i < SURFACE_NUM; i++) {
1790         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1791         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1792         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1793     }
1794
1795     vaDestroyContext(va_dpy, context_id);
1796     vaDestroyConfig(va_dpy, config_id);
1797
1798     return 0;
1799 }
1800
1801 static int deinit_va()
1802 {
1803     vaTerminate(va_dpy);
1804
1805     va_close_display(va_dpy);
1806
1807     return 0;
1808 }
1809
1810
1811 static int print_input()
1812 {
1813     printf("\n\nINPUT:Try to encode H264...\n");
1814     if (rc_mode != -1)
1815         printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1816     printf("INPUT: Resolution   : %dx%dframes\n", frame_width, frame_height);
1817     printf("INPUT: FrameRate    : %d\n", frame_rate);
1818     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
1819     printf("INPUT: Slieces      : %d\n", frame_slices);
1820     printf("INPUT: IntraPeriod  : %d\n", intra_period);
1821     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
1822     printf("INPUT: IpPeriod     : %d\n", ip_period);
1823     printf("INPUT: Initial QP   : %d\n", initial_qp);
1824     printf("INPUT: Min QP       : %d\n", minimal_qp);
1825     printf("INPUT: Coded Clip   : %s\n", coded_fn);
1826
1827     printf("\n\n"); /* return back to startpoint */
1828
1829     return 0;
1830 }
1831
1832 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
1833         : current_storage_frame(0), surface(surface), httpd(httpd)
1834 {
1835         AVCodec *codec_audio = avcodec_find_encoder(AV_CODEC_ID_MP3);
1836         context_audio = avcodec_alloc_context3(codec_audio);
1837         context_audio->bit_rate = 256000;
1838         context_audio->sample_rate = 48000;
1839         context_audio->sample_fmt = AV_SAMPLE_FMT_FLTP;
1840         context_audio->channels = 2;
1841         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1842         context_audio->time_base = AVRational{1, TIMEBASE};
1843         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1844                 fprintf(stderr, "Could not open codec\n");
1845                 exit(1);
1846         }
1847
1848         frame_width = width;
1849         frame_height = height;
1850         frame_width_mbaligned = (frame_width + 15) & (~15);
1851         frame_height_mbaligned = (frame_height + 15) & (~15);
1852         frame_bitrate = 15000000;  // / 60;
1853         current_frame_encoding = 0;
1854
1855         print_input();
1856
1857         init_va();
1858         setup_encode();
1859
1860         // No frames are ready yet.
1861         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1862
1863         memset(&seq_param, 0, sizeof(seq_param));
1864         memset(&pic_param, 0, sizeof(pic_param));
1865         memset(&slice_param, 0, sizeof(slice_param));
1866
1867         storage_thread = std::thread(&H264Encoder::storage_task_thread, this);
1868
1869         copy_thread = std::thread([this]{
1870                 //SDL_GL_MakeCurrent(window, context);
1871                 QOpenGLContext *context = create_context();
1872                 eglBindAPI(EGL_OPENGL_API);
1873                 if (!make_current(context, this->surface)) {
1874                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1875                                 eglGetError());
1876                         exit(1);
1877                 }
1878                 copy_thread_func();
1879         });
1880 }
1881
1882 H264Encoder::~H264Encoder()
1883 {
1884         {
1885                 unique_lock<mutex> lock(storage_task_queue_mutex);
1886                 storage_thread_should_quit = true;
1887                 storage_task_queue_changed.notify_all();
1888         }
1889         {
1890                 unique_lock<mutex> lock(frame_queue_mutex);
1891                 copy_thread_should_quit = true;
1892                 frame_queue_nonempty.notify_one();
1893         }
1894         storage_thread.join();
1895         copy_thread.join();
1896
1897         release_encode();
1898         deinit_va();
1899 }
1900
1901 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1902 {
1903         {
1904                 // Wait until this frame slot is done encoding.
1905                 std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1906                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1907                 if (storage_thread_should_quit) return false;
1908         }
1909
1910         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1911         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1912         *y_tex = surf->y_tex;
1913         *cbcr_tex = surf->cbcr_tex;
1914
1915         VASurfaceID surface = surf->src_surface;
1916         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1917         CHECK_VASTATUS(va_status, "vaDeriveImage");
1918
1919         VABufferInfo buf_info;
1920         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1921         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1922         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1923
1924         // Create Y image.
1925         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1926         EGLint y_attribs[] = {
1927                 EGL_WIDTH, frame_width,
1928                 EGL_HEIGHT, frame_height,
1929                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1930                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1931                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1932                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1933                 EGL_NONE
1934         };
1935
1936         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1937         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1938
1939         // Associate Y image to a texture.
1940         glBindTexture(GL_TEXTURE_2D, *y_tex);
1941         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1942
1943         // Create CbCr image.
1944         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1945         EGLint cbcr_attribs[] = {
1946                 EGL_WIDTH, frame_width,
1947                 EGL_HEIGHT, frame_height,
1948                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1949                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1950                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1951                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1952                 EGL_NONE
1953         };
1954
1955         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1956         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1957
1958         // Associate CbCr image to a texture.
1959         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1960         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1961
1962         return true;
1963 }
1964
1965 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
1966 {
1967         {
1968                 unique_lock<mutex> lock(frame_queue_mutex);
1969                 pending_audio_frames[pts] = move(audio);
1970         }
1971         frame_queue_nonempty.notify_one();
1972 }
1973
1974
1975 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
1976 {
1977         {
1978                 unique_lock<mutex> lock(frame_queue_mutex);
1979                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1980                 ++current_storage_frame;
1981         }
1982         frame_queue_nonempty.notify_one();
1983 }
1984
1985 void H264Encoder::copy_thread_func()
1986 {
1987         int64_t last_dts = -1;
1988         for ( ;; ) {
1989                 PendingFrame frame;
1990                 int pts_lag;
1991                 encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
1992                                        &current_frame_display, &current_frame_type, &pts_lag);
1993                 if (current_frame_type == FRAME_IDR) {
1994                         numShortTerm = 0;
1995                         current_frame_num = 0;
1996                         current_IDR_display = current_frame_display;
1997                 }
1998
1999                 {
2000                         unique_lock<mutex> lock(frame_queue_mutex);
2001                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; });
2002                         if (copy_thread_should_quit) return;
2003                         frame = move(pending_video_frames[current_frame_display]);
2004                         pending_video_frames.erase(current_frame_display);
2005                 }
2006
2007                 // Wait for the GPU to be done with the frame.
2008                 glClientWaitSync(frame.fence.get(), 0, 0);
2009
2010                 // Release back any input frames we needed to render this frame.
2011                 frame.input_frames.clear();
2012
2013                 // Unmap the image.
2014                 GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
2015                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2016                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2017                 VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2018                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2019                 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2020                 CHECK_VASTATUS(va_status, "vaDestroyImage");
2021
2022                 VASurfaceID surface = surf->src_surface;
2023
2024                 // Schedule the frame for encoding.
2025                 va_status = vaBeginPicture(va_dpy, context_id, surface);
2026                 CHECK_VASTATUS(va_status, "vaBeginPicture");
2027
2028                 if (current_frame_type == FRAME_IDR) {
2029                         render_sequence();
2030                         render_picture();
2031                         if (h264_packedheader) {
2032                                 render_packedsequence();
2033                                 render_packedpicture();
2034                         }
2035                 } else {
2036                         //render_sequence();
2037                         render_picture();
2038                 }
2039                 render_slice();
2040
2041                 va_status = vaEndPicture(va_dpy, context_id);
2042                 CHECK_VASTATUS(va_status, "vaEndPicture");
2043
2044                 // Determine the pts and dts of this frame.
2045                 int64_t pts = frame.pts;
2046                 int64_t dts;
2047                 if (pts_lag == -1) {
2048                         assert(last_dts != -1);
2049                         dts = last_dts + (TIMEBASE / frame_rate);
2050                 } else {
2051                         dts = pts - pts_lag;
2052                 }
2053                 last_dts = dts;
2054
2055                 // so now the data is done encoding (well, async job kicked off)...
2056                 // we send that to the storage thread
2057                 storage_task tmp;
2058                 tmp.display_order = current_frame_display;
2059                 tmp.frame_type = current_frame_type;
2060                 tmp.pts = pts;
2061                 tmp.dts = dts;
2062                 storage_task_enqueue(move(tmp));
2063
2064                 update_ReferenceFrames();
2065                 ++current_frame_encoding;
2066         }
2067 }