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 /*
 590  * Return displaying order with specified periods and encoding order
 591  * displaying_order: displaying order
 592  * frame_type: frame type
 593  */
 594 #define FRAME_P 0
 595 #define FRAME_B 1
 596 #define FRAME_I 2
 597 #define FRAME_IDR 7
 598 void encoding2display_order(
 599     unsigned long long encoding_order, int intra_period,
 600     int intra_idr_period, int ip_period,
 601     unsigned long long *displaying_order,
 602     int *frame_type)
 603 {
 604     int encoding_order_gop = 0;
 605
 606     if (intra_period == 1) { /* all are I/IDR frames */
 607         *displaying_order = encoding_order;
 608         if (intra_idr_period == 0)
 609             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 610         else
 611             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 612         return;
 613     }
 614
 615     if (intra_period == 0)
 616         intra_idr_period = 0;
 617
 618     /* new sequence like
 619      * IDR PPPPP IPPPPP
 620      * IDR (PBB)(PBB)(IBB)(PBB)
 621      */
 622     encoding_order_gop = (intra_idr_period == 0)? encoding_order:
 623         (encoding_order % (intra_idr_period + ((ip_period == 1)?0:1)));
 624
 625     if (encoding_order_gop == 0) { /* the first frame */
 626         *frame_type = FRAME_IDR;
 627         *displaying_order = encoding_order;
 628     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 629         *frame_type = FRAME_B;
 630         *displaying_order = encoding_order - 1;
 631     } else if ((intra_period != 0) && /* have I frames */
 632                (encoding_order_gop >= 2) &&
 633                ((ip_period == 1 && encoding_order_gop % intra_period == 0) || /* for IDR PPPPP IPPPP */
 634                 /* for IDR (PBB)(PBB)(IBB) */
 635                 (ip_period >= 2 && ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0))) {
 636         *frame_type = FRAME_I;
 637         *displaying_order = encoding_order + ip_period - 1;
 638     } else {
 639         *frame_type = FRAME_P;
 640         *displaying_order = encoding_order + ip_period - 1;
 641     }
 642 }
 643
 644
 645 static const char *rc_to_string(int rcmode)
 646 {
 647     switch (rc_mode) {
 648     case VA_RC_NONE:
 649         return "NONE";
 650     case VA_RC_CBR:
 651         return "CBR";
 652     case VA_RC_VBR:
 653         return "VBR";
 654     case VA_RC_VCM:
 655         return "VCM";
 656     case VA_RC_CQP:
 657         return "CQP";
 658     case VA_RC_VBR_CONSTRAINED:
 659         return "VBR_CONSTRAINED";
 660     default:
 661         return "Unknown";
 662     }
 663 }
 664
 665 #if 0
 666 static int process_cmdline(int argc, char *argv[])
 667 {
 668     char c;
 669     const struct option long_opts[] = {
 670         {"help", no_argument, NULL, 0 },
 671         {"bitrate", required_argument, NULL, 1 },
 672         {"minqp", required_argument, NULL, 2 },
 673         {"initialqp", required_argument, NULL, 3 },
 674         {"intra_period", required_argument, NULL, 4 },
 675         {"idr_period", required_argument, NULL, 5 },
 676         {"ip_period", required_argument, NULL, 6 },
 677         {"rcmode", required_argument, NULL, 7 },
 678         {"srcyuv", required_argument, NULL, 9 },
 679         {"recyuv", required_argument, NULL, 10 },
 680         {"fourcc", required_argument, NULL, 11 },
 681         {"syncmode", no_argument, NULL, 12 },
 682         {"enablePSNR", no_argument, NULL, 13 },
 683         {"prit", required_argument, NULL, 14 },
 684         {"priv", required_argument, NULL, 15 },
 685         {"framecount", required_argument, NULL, 16 },
 686         {"entropy", required_argument, NULL, 17 },
 687         {"profile", required_argument, NULL, 18 },
 688         {NULL, no_argument, NULL, 0 }};
 689     int long_index;
 690
 691     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 692         switch (c) {
 693         case 'w':
 694             frame_width = atoi(optarg);
 695             break;
 696         case 'h':
 697             frame_height = atoi(optarg);
 698             break;
 699         case 'n':
 700         case 'f':
 701             frame_rate = atoi(optarg);
 702             break;
 703         case 'o':
 704             coded_fn = strdup(optarg);
 705             break;
 706         case 0:
 707             print_help();
 708             exit(0);
 709         case 1:
 710             frame_bitrate = atoi(optarg);
 711             break;
 712         case 2:
 713             minimal_qp = atoi(optarg);
 714             break;
 715         case 3:
 716             initial_qp = atoi(optarg);
 717             break;
 718         case 4:
 719             intra_period = atoi(optarg);
 720             break;
 721         case 5:
 722             intra_idr_period = atoi(optarg);
 723             break;
 724         case 6:
 725             ip_period = atoi(optarg);
 726             break;
 727         case 7:
 728             rc_mode = string_to_rc(optarg);
 729             if (rc_mode < 0) {
 730                 print_help();
 731                 exit(1);
 732             }
 733             break;
 734         case 9:
 735             srcyuv_fn = strdup(optarg);
 736             break;
 737         case 11:
 738             srcyuv_fourcc = string_to_fourcc(optarg);
 739             if (srcyuv_fourcc <= 0) {
 740                 print_help();
 741                 exit(1);
 742             }
 743             break;
 744         case 13:
 745             calc_psnr = 1;
 746             break;
 747         case 14:
 748             misc_priv_type = strtol(optarg, NULL, 0);
 749             break;
 750         case 15:
 751             misc_priv_value = strtol(optarg, NULL, 0);
 752             break;
 753         case 17:
 754             h264_entropy_mode = atoi(optarg) ? 1: 0;
 755             break;
 756         case 18:
 757             if (strncmp(optarg, "BP", 2) == 0)
 758                 h264_profile = VAProfileH264Baseline;
 759             else if (strncmp(optarg, "MP", 2) == 0)
 760                 h264_profile = VAProfileH264Main;
 761             else if (strncmp(optarg, "HP", 2) == 0)
 762                 h264_profile = VAProfileH264High;
 763             else
 764                 h264_profile = (VAProfile)0;
 765             break;
 766         case ':':
 767         case '?':
 768             print_help();
 769             exit(0);
 770         }
 771     }
 772
 773     if (ip_period < 1) {
 774         printf(" ip_period must be greater than 0\n");
 775         exit(0);
 776     }
 777     if (intra_period != 1 && intra_period % ip_period != 0) {
 778         printf(" intra_period must be a multiplier of ip_period\n");
 779         exit(0);
 780     }
 781     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 782         printf(" intra_idr_period must be a multiplier of intra_period\n");
 783         exit(0);
 784     }
 785
 786     if (frame_bitrate == 0)
 787         frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
 788
 789     if (coded_fn == NULL) {
 790         struct stat buf;
 791         if (stat("/tmp", &buf) == 0)
 792             coded_fn = strdup("/tmp/test.264");
 793         else if (stat("/sdcard", &buf) == 0)
 794             coded_fn = strdup("/sdcard/test.264");
 795         else
 796             coded_fn = strdup("./test.264");
 797     }
 798
 799
 800     frame_width_mbaligned = (frame_width + 15) & (~15);
 801     frame_height_mbaligned = (frame_height + 15) & (~15);
 802     if (frame_width != frame_width_mbaligned ||
 803         frame_height != frame_height_mbaligned) {
 804         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 805                frame_width, frame_height,
 806                frame_width_mbaligned, frame_height_mbaligned
 807                );
 808     }
 809
 810     return 0;
 811 }
 812 #endif
 813
 814 static Display *x11_display;
 815 static Window   x11_window;
 816
 817 VADisplay
 818 va_open_display(void)
 819 {
 820     x11_display = XOpenDisplay(NULL);
 821     if (!x11_display) {
 822         fprintf(stderr, "error: can't connect to X server!\n");
 823         return NULL;
 824     }
 825     return vaGetDisplay(x11_display);
 826 }
 827
 828 void
 829 va_close_display(VADisplay va_dpy)
 830 {
 831     if (!x11_display)
 832         return;
 833
 834     if (x11_window) {
 835         XUnmapWindow(x11_display, x11_window);
 836         XDestroyWindow(x11_display, x11_window);
 837         x11_window = None;
 838     }
 839     XCloseDisplay(x11_display);
 840     x11_display = NULL;
 841 }
 842
 843 static int init_va(void)
 844 {
 845     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 846     VAEntrypoint *entrypoints;
 847     int num_entrypoints, slice_entrypoint;
 848     int support_encode = 0;
 849     int major_ver, minor_ver;
 850     VAStatus va_status;
 851     unsigned int i;
 852
 853     va_dpy = va_open_display();
 854     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 855     CHECK_VASTATUS(va_status, "vaInitialize");
 856
 857     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 858     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 859     if (!entrypoints) {
 860         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 861         exit(1);
 862     }
 863
 864     /* use the highest profile */
 865     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 866         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 867             continue;
 868
 869         h264_profile = profile_list[i];
 870         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 871         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 872             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 873                 support_encode = 1;
 874                 break;
 875             }
 876         }
 877         if (support_encode == 1)
 878             break;
 879     }
 880
 881     if (support_encode == 0) {
 882         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 883         exit(1);
 884     } else {
 885         switch (h264_profile) {
 886             case VAProfileH264Baseline:
 887                 printf("Use profile VAProfileH264Baseline\n");
 888                 ip_period = 1;
 889                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 890                 h264_entropy_mode = 0;
 891                 break;
 892             case VAProfileH264ConstrainedBaseline:
 893                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 894                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 895                 ip_period = 1;
 896                 break;
 897
 898             case VAProfileH264Main:
 899                 printf("Use profile VAProfileH264Main\n");
 900                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 901                 break;
 902
 903             case VAProfileH264High:
 904                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 905                 printf("Use profile VAProfileH264High\n");
 906                 break;
 907             default:
 908                 printf("unknow profile. Set to Baseline");
 909                 h264_profile = VAProfileH264Baseline;
 910                 ip_period = 1;
 911                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 912                 break;
 913         }
 914     }
 915
 916     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 917
 918     /* find out the format for the render target, and rate control mode */
 919     for (i = 0; i < VAConfigAttribTypeMax; i++)
 920         attrib[i].type = (VAConfigAttribType)i;
 921
 922     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 923                                       &attrib[0], VAConfigAttribTypeMax);
 924     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 925     /* check the interested configattrib */
 926     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 927         printf("Not find desired YUV420 RT format\n");
 928         exit(1);
 929     } else {
 930         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 931         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 932         config_attrib_num++;
 933     }
 934
 935     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 936         int tmp = attrib[VAConfigAttribRateControl].value;
 937
 938         printf("Support rate control mode (0x%x):", tmp);
 939
 940         if (tmp & VA_RC_NONE)
 941             printf("NONE ");
 942         if (tmp & VA_RC_CBR)
 943             printf("CBR ");
 944         if (tmp & VA_RC_VBR)
 945             printf("VBR ");
 946         if (tmp & VA_RC_VCM)
 947             printf("VCM ");
 948         if (tmp & VA_RC_CQP)
 949             printf("CQP ");
 950         if (tmp & VA_RC_VBR_CONSTRAINED)
 951             printf("VBR_CONSTRAINED ");
 952
 953         printf("\n");
 954
 955         if (rc_mode == -1 || !(rc_mode & tmp))  {
 956             if (rc_mode != -1) {
 957                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
 958             }
 959
 960             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
 961                 if (rc_default_modes[i] & tmp) {
 962                     rc_mode = rc_default_modes[i];
 963                     break;
 964                 }
 965             }
 966
 967             printf("RateControl mode: %s\n", rc_to_string(rc_mode));
 968         }
 969
 970         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 971         config_attrib[config_attrib_num].value = rc_mode;
 972         config_attrib_num++;
 973     }
 974
 975
 976     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 977         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
 978
 979         printf("Support VAConfigAttribEncPackedHeaders\n");
 980
 981         h264_packedheader = 1;
 982         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
 983         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
 984
 985         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
 986             printf("Support packed sequence headers\n");
 987             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
 988         }
 989
 990         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
 991             printf("Support packed picture headers\n");
 992             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
 993         }
 994
 995         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
 996             printf("Support packed slice headers\n");
 997             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
 998         }
 999
1000         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1001             printf("Support packed misc headers\n");
1002             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1003         }
1004
1005         enc_packed_header_idx = config_attrib_num;
1006         config_attrib_num++;
1007     }
1008
1009     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1010         int tmp = attrib[VAConfigAttribEncInterlaced].value;
1011
1012         printf("Support VAConfigAttribEncInterlaced\n");
1013
1014         if (tmp & VA_ENC_INTERLACED_FRAME)
1015             printf("support VA_ENC_INTERLACED_FRAME\n");
1016         if (tmp & VA_ENC_INTERLACED_FIELD)
1017             printf("Support VA_ENC_INTERLACED_FIELD\n");
1018         if (tmp & VA_ENC_INTERLACED_MBAFF)
1019             printf("Support VA_ENC_INTERLACED_MBAFF\n");
1020         if (tmp & VA_ENC_INTERLACED_PAFF)
1021             printf("Support VA_ENC_INTERLACED_PAFF\n");
1022
1023         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1024         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1025         config_attrib_num++;
1026     }
1027
1028     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1029         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1030
1031         printf("Support %d RefPicList0 and %d RefPicList1\n",
1032                h264_maxref & 0xffff, (h264_maxref >> 16) & 0xffff );
1033     }
1034
1035     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1036         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1037
1038     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1039         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1040
1041         printf("Support VAConfigAttribEncSliceStructure\n");
1042
1043         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1044             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1045         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1046             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1047         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1048             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1049     }
1050     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1051         printf("Support VAConfigAttribEncMacroblockInfo\n");
1052     }
1053
1054     free(entrypoints);
1055     return 0;
1056 }
1057
1058 static int setup_encode()
1059 {
1060     VAStatus va_status;
1061     VASurfaceID *tmp_surfaceid;
1062     int codedbuf_size, i;
1063     static VASurfaceID src_surface[SURFACE_NUM];
1064     static VASurfaceID ref_surface[SURFACE_NUM];
1065
1066     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1067             &config_attrib[0], config_attrib_num, &config_id);
1068     CHECK_VASTATUS(va_status, "vaCreateConfig");
1069
1070     /* create source surfaces */
1071     va_status = vaCreateSurfaces(va_dpy,
1072                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1073                                  &src_surface[0], SURFACE_NUM,
1074                                  NULL, 0);
1075     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1076
1077     /* create reference surfaces */
1078     va_status = vaCreateSurfaces(va_dpy,
1079                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1080                                  &ref_surface[0], SURFACE_NUM,
1081                                  NULL, 0);
1082     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1083
1084     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1085     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1086     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1087
1088     /* Create a context for this encode pipe */
1089     va_status = vaCreateContext(va_dpy, config_id,
1090                                 frame_width_mbaligned, frame_height_mbaligned,
1091                                 VA_PROGRESSIVE,
1092                                 tmp_surfaceid, 2 * SURFACE_NUM,
1093                                 &context_id);
1094     CHECK_VASTATUS(va_status, "vaCreateContext");
1095     free(tmp_surfaceid);
1096
1097     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1098
1099     for (i = 0; i < SURFACE_NUM; i++) {
1100         /* create coded buffer once for all
1101          * other VA buffers which won't be used again after vaRenderPicture.
1102          * so APP can always vaCreateBuffer for every frame
1103          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1104          * so VA won't maintain the coded buffer
1105          */
1106         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1107                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1108         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1109     }
1110
1111     /* create OpenGL objects */
1112     //glGenFramebuffers(SURFACE_NUM, fbos);
1113
1114     for (i = 0; i < SURFACE_NUM; i++) {
1115         glGenTextures(1, &gl_surfaces[i].y_tex);
1116         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1117     }
1118
1119     for (i = 0; i < SURFACE_NUM; i++) {
1120         gl_surfaces[i].src_surface = src_surface[i];
1121         gl_surfaces[i].ref_surface = ref_surface[i];
1122     }
1123
1124     return 0;
1125 }
1126
1127
1128
1129 #define partition(ref, field, key, ascending)   \
1130     while (i <= j) {                            \
1131         if (ascending) {                        \
1132             while (ref[i].field < key)          \
1133                 i++;                            \
1134             while (ref[j].field > key)          \
1135                 j--;                            \
1136         } else {                                \
1137             while (ref[i].field > key)          \
1138                 i++;                            \
1139             while (ref[j].field < key)          \
1140                 j--;                            \
1141         }                                       \
1142         if (i <= j) {                           \
1143             tmp = ref[i];                       \
1144             ref[i] = ref[j];                    \
1145             ref[j] = tmp;                       \
1146             i++;                                \
1147             j--;                                \
1148         }                                       \
1149     }                                           \
1150
1151 static void sort_one(VAPictureH264 ref[], int left, int right,
1152                      int ascending, int frame_idx)
1153 {
1154     int i = left, j = right;
1155     unsigned int key;
1156     VAPictureH264 tmp;
1157
1158     if (frame_idx) {
1159         key = ref[(left + right) / 2].frame_idx;
1160         partition(ref, frame_idx, key, ascending);
1161     } else {
1162         key = ref[(left + right) / 2].TopFieldOrderCnt;
1163         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1164     }
1165
1166     /* recursion */
1167     if (left < j)
1168         sort_one(ref, left, j, ascending, frame_idx);
1169
1170     if (i < right)
1171         sort_one(ref, i, right, ascending, frame_idx);
1172 }
1173
1174 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1175                      int partition_ascending, int list0_ascending, int list1_ascending)
1176 {
1177     int i = left, j = right;
1178     VAPictureH264 tmp;
1179
1180     if (frame_idx) {
1181         partition(ref, frame_idx, key, partition_ascending);
1182     } else {
1183         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1184     }
1185
1186
1187     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1188     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1189 }
1190
1191 static int update_ReferenceFrames(void)
1192 {
1193     int i;
1194
1195     if (current_frame_type == FRAME_B)
1196         return 0;
1197
1198     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1199     numShortTerm++;
1200     if (numShortTerm > num_ref_frames)
1201         numShortTerm = num_ref_frames;
1202     for (i=numShortTerm-1; i>0; i--)
1203         ReferenceFrames[i] = ReferenceFrames[i-1];
1204     ReferenceFrames[0] = CurrentCurrPic;
1205
1206     if (current_frame_type != FRAME_B)
1207         current_frame_num++;
1208     if (current_frame_num > MaxFrameNum)
1209         current_frame_num = 0;
1210
1211     return 0;
1212 }
1213
1214
1215 static int update_RefPicList(void)
1216 {
1217     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1218
1219     if (current_frame_type == FRAME_P) {
1220         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1221         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1222     }
1223
1224     if (current_frame_type == FRAME_B) {
1225         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1226         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1227                  1, 0, 1);
1228
1229         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1230         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1231                  0, 1, 0);
1232     }
1233
1234     return 0;
1235 }
1236
1237
1238 static int render_sequence(void)
1239 {
1240     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1241     VAStatus va_status;
1242     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1243     VAEncMiscParameterRateControl *misc_rate_ctrl;
1244
1245     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1246     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1247     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1248     seq_param.bits_per_second = frame_bitrate;
1249
1250     seq_param.intra_period = intra_period;
1251     seq_param.intra_idr_period = intra_idr_period;
1252     seq_param.ip_period = ip_period;
1253
1254     seq_param.max_num_ref_frames = num_ref_frames;
1255     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1256     seq_param.time_scale = TIMEBASE * 2;
1257     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1258     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1259     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1260     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1261     seq_param.seq_fields.bits.chroma_format_idc = 1;
1262     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1263
1264     if (frame_width != frame_width_mbaligned ||
1265         frame_height != frame_height_mbaligned) {
1266         seq_param.frame_cropping_flag = 1;
1267         seq_param.frame_crop_left_offset = 0;
1268         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1269         seq_param.frame_crop_top_offset = 0;
1270         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1271     }
1272
1273     va_status = vaCreateBuffer(va_dpy, context_id,
1274                                VAEncSequenceParameterBufferType,
1275                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1276     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1277
1278     va_status = vaCreateBuffer(va_dpy, context_id,
1279                                VAEncMiscParameterBufferType,
1280                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1281                                1, NULL, &rc_param_buf);
1282     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1283
1284     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1285     misc_param->type = VAEncMiscParameterTypeRateControl;
1286     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1287     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1288     misc_rate_ctrl->bits_per_second = frame_bitrate;
1289     misc_rate_ctrl->target_percentage = 66;
1290     misc_rate_ctrl->window_size = 1000;
1291     misc_rate_ctrl->initial_qp = initial_qp;
1292     misc_rate_ctrl->min_qp = minimal_qp;
1293     misc_rate_ctrl->basic_unit_size = 0;
1294     vaUnmapBuffer(va_dpy, rc_param_buf);
1295
1296     render_id[0] = seq_param_buf;
1297     render_id[1] = rc_param_buf;
1298
1299     va_status = vaRenderPicture(va_dpy, context_id, &render_id[0], 2);
1300     CHECK_VASTATUS(va_status, "vaRenderPicture");;
1301
1302     if (misc_priv_type != 0) {
1303         va_status = vaCreateBuffer(va_dpy, context_id,
1304                                    VAEncMiscParameterBufferType,
1305                                    sizeof(VAEncMiscParameterBuffer),
1306                                    1, NULL, &misc_param_tmpbuf);
1307         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1308         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1309         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1310         misc_param_tmp->data[0] = misc_priv_value;
1311         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1312
1313         va_status = vaRenderPicture(va_dpy, context_id, &misc_param_tmpbuf, 1);
1314     }
1315
1316     return 0;
1317 }
1318
1319 static int calc_poc(int pic_order_cnt_lsb)
1320 {
1321     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1322     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1323     int PicOrderCntMsb, TopFieldOrderCnt;
1324
1325     if (current_frame_type == FRAME_IDR)
1326         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1327     else {
1328         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1329         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1330     }
1331
1332     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1333         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1334         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1335     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1336              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1337         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1338     else
1339         PicOrderCntMsb = prevPicOrderCntMsb;
1340
1341     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1342
1343     if (current_frame_type != FRAME_B) {
1344         PicOrderCntMsb_ref = PicOrderCntMsb;
1345         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1346     }
1347
1348     return TopFieldOrderCnt;
1349 }
1350
1351 static int render_picture(void)
1352 {
1353     VABufferID pic_param_buf;
1354     VAStatus va_status;
1355     int i = 0;
1356
1357     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1358     pic_param.CurrPic.frame_idx = current_frame_num;
1359     pic_param.CurrPic.flags = 0;
1360     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1361     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1362     CurrentCurrPic = pic_param.CurrPic;
1363
1364     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1365         update_RefPicList(); /* calc RefPicList */
1366         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1367         if (current_frame_type == FRAME_P) {
1368             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1369         } else if (current_frame_type == FRAME_B) {
1370             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1371             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1372         }
1373     } else {
1374         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1375         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1376             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1377             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1378         }
1379     }
1380
1381     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1382     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1383     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1384     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1385     pic_param.frame_num = current_frame_num;
1386     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1387     pic_param.last_picture = false;  // FIXME
1388     pic_param.pic_init_qp = initial_qp;
1389
1390     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1391                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1392     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1393
1394     va_status = vaRenderPicture(va_dpy, context_id, &pic_param_buf, 1);
1395     CHECK_VASTATUS(va_status, "vaRenderPicture");
1396
1397     return 0;
1398 }
1399
1400 static int render_packedsequence(void)
1401 {
1402     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1403     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1404     unsigned int length_in_bits;
1405     unsigned char *packedseq_buffer = NULL;
1406     VAStatus va_status;
1407
1408     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1409
1410     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1411
1412     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1413     packedheader_param_buffer.has_emulation_bytes = 0;
1414     va_status = vaCreateBuffer(va_dpy,
1415                                context_id,
1416                                VAEncPackedHeaderParameterBufferType,
1417                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1418                                &packedseq_para_bufid);
1419     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1420
1421     va_status = vaCreateBuffer(va_dpy,
1422                                context_id,
1423                                VAEncPackedHeaderDataBufferType,
1424                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1425                                &packedseq_data_bufid);
1426     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1427
1428     render_id[0] = packedseq_para_bufid;
1429     render_id[1] = packedseq_data_bufid;
1430     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1431     CHECK_VASTATUS(va_status, "vaRenderPicture");
1432
1433     free(packedseq_buffer);
1434
1435     return 0;
1436 }
1437
1438
1439 static int render_packedpicture(void)
1440 {
1441     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1442     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1443     unsigned int length_in_bits;
1444     unsigned char *packedpic_buffer = NULL;
1445     VAStatus va_status;
1446
1447     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1448     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1449     packedheader_param_buffer.bit_length = length_in_bits;
1450     packedheader_param_buffer.has_emulation_bytes = 0;
1451
1452     va_status = vaCreateBuffer(va_dpy,
1453                                context_id,
1454                                VAEncPackedHeaderParameterBufferType,
1455                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1456                                &packedpic_para_bufid);
1457     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1458
1459     va_status = vaCreateBuffer(va_dpy,
1460                                context_id,
1461                                VAEncPackedHeaderDataBufferType,
1462                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1463                                &packedpic_data_bufid);
1464     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1465
1466     render_id[0] = packedpic_para_bufid;
1467     render_id[1] = packedpic_data_bufid;
1468     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1469     CHECK_VASTATUS(va_status, "vaRenderPicture");
1470
1471     free(packedpic_buffer);
1472
1473     return 0;
1474 }
1475
1476 static void render_packedslice()
1477 {
1478     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1479     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1480     unsigned int length_in_bits;
1481     unsigned char *packedslice_buffer = NULL;
1482     VAStatus va_status;
1483
1484     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1485     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1486     packedheader_param_buffer.bit_length = length_in_bits;
1487     packedheader_param_buffer.has_emulation_bytes = 0;
1488
1489     va_status = vaCreateBuffer(va_dpy,
1490                                context_id,
1491                                VAEncPackedHeaderParameterBufferType,
1492                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1493                                &packedslice_para_bufid);
1494     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1495
1496     va_status = vaCreateBuffer(va_dpy,
1497                                context_id,
1498                                VAEncPackedHeaderDataBufferType,
1499                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1500                                &packedslice_data_bufid);
1501     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1502
1503     render_id[0] = packedslice_para_bufid;
1504     render_id[1] = packedslice_data_bufid;
1505     va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
1506     CHECK_VASTATUS(va_status, "vaRenderPicture");
1507
1508     free(packedslice_buffer);
1509 }
1510
1511 static int render_slice(void)
1512 {
1513     VABufferID slice_param_buf;
1514     VAStatus va_status;
1515     int i;
1516
1517     update_RefPicList();
1518
1519     /* one frame, one slice */
1520     slice_param.macroblock_address = 0;
1521     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1522     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1523     if (current_frame_type == FRAME_IDR) {
1524         if (current_frame_encoding != 0)
1525             ++slice_param.idr_pic_id;
1526     } else if (current_frame_type == FRAME_P) {
1527         int refpiclist0_max = h264_maxref & 0xffff;
1528         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1529
1530         for (i = refpiclist0_max; i < 32; i++) {
1531             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1532             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1533         }
1534     } else if (current_frame_type == FRAME_B) {
1535         int refpiclist0_max = h264_maxref & 0xffff;
1536         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1537
1538         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1539         for (i = refpiclist0_max; i < 32; i++) {
1540             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1541             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1542         }
1543
1544         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1545         for (i = refpiclist1_max; i < 32; i++) {
1546             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1547             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1548         }
1549     }
1550
1551     slice_param.slice_alpha_c0_offset_div2 = 0;
1552     slice_param.slice_beta_offset_div2 = 0;
1553     slice_param.direct_spatial_mv_pred_flag = 1;
1554     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1555
1556
1557     if (h264_packedheader &&
1558         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1559         render_packedslice();
1560
1561     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1562                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1563     CHECK_VASTATUS(va_status, "vaCreateBuffer");;
1564
1565     va_status = vaRenderPicture(va_dpy, context_id, &slice_param_buf, 1);
1566     CHECK_VASTATUS(va_status, "vaRenderPicture");
1567
1568     return 0;
1569 }
1570
1571
1572
1573 int H264Encoder::save_codeddata(storage_task task)
1574 {
1575     VACodedBufferSegment *buf_list = NULL;
1576     VAStatus va_status;
1577     unsigned int coded_size = 0;
1578
1579     string data;
1580
1581     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1582     CHECK_VASTATUS(va_status, "vaMapBuffer");
1583     while (buf_list != NULL) {
1584         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1585         buf_list = (VACodedBufferSegment *) buf_list->next;
1586
1587         frame_size += coded_size;
1588     }
1589     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1590
1591     const int64_t pts_dts_delay = (ip_period - 1) * (TIMEBASE / frame_rate);  // FIXME: Wrong for variable frame rate.
1592     const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
1593     int64_t pts, dts;
1594     {
1595         {
1596              unique_lock<mutex> lock(frame_queue_mutex);
1597              assert(timestamps.count(task.display_order));
1598              assert(timestamps.count(task.encode_order));
1599              pts = timestamps[task.display_order];
1600              dts = timestamps[task.encode_order];
1601         }
1602         // Add video.
1603         AVPacket pkt;
1604         memset(&pkt, 0, sizeof(pkt));
1605         pkt.buf = nullptr;
1606         pkt.pts = av_rescale_q(pts + av_delay + pts_dts_delay, AVRational{1, TIMEBASE}, avstream_video->time_base);
1607         pkt.dts = av_rescale_q(dts + av_delay, AVRational{1, TIMEBASE}, avstream_video->time_base);
1608         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1609         pkt.size = data.size();
1610         pkt.stream_index = 0;
1611         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1612             pkt.flags = AV_PKT_FLAG_KEY;
1613         } else {
1614             pkt.flags = 0;
1615         }
1616         //pkt.duration = 1;
1617         av_interleaved_write_frame(avctx, &pkt);
1618     }
1619     // Encode and add all audio frames up to and including the pts of this video frame.
1620     // (They can never be queued to us after the video frame they belong to, only before.)
1621     for ( ;; ) {
1622         int64_t audio_pts;
1623         std::vector<float> audio;
1624         {
1625              unique_lock<mutex> lock(frame_queue_mutex);
1626              if (pending_audio_frames.empty()) break;
1627              auto it = pending_audio_frames.begin();
1628              if (it->first > int(pts)) break;
1629              audio_pts = it->first;
1630              audio = move(it->second);
1631              pending_audio_frames.erase(it);
1632         }
1633         AVFrame *frame = avcodec_alloc_frame();
1634         frame->nb_samples = audio.size() / 2;
1635         frame->format = AV_SAMPLE_FMT_FLT;
1636         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1637
1638         unique_ptr<float[]> planar_samples(new float[audio.size()]);
1639         avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_FLTP, (const uint8_t*)planar_samples.get(), audio.size() * sizeof(float), 0);
1640         for (int i = 0; i < frame->nb_samples; ++i) {
1641             planar_samples[i] = audio[i * 2 + 0];
1642             planar_samples[i + frame->nb_samples] = audio[i * 2 + 1];
1643         }
1644
1645         AVPacket pkt;
1646         av_init_packet(&pkt);
1647         pkt.data = nullptr;
1648         pkt.size = 0;
1649         int got_output;
1650         avcodec_encode_audio2(avstream_audio->codec, &pkt, frame, &got_output);
1651         if (got_output) {
1652             pkt.pts = av_rescale_q(audio_pts + pts_dts_delay, AVRational{1, TIMEBASE}, avstream_audio->time_base);
1653             pkt.dts = pkt.pts;
1654             pkt.stream_index = 1;
1655             av_interleaved_write_frame(avctx, &pkt);
1656         }
1657         // TODO: Delayed frames.
1658         avcodec_free_frame(&frame);
1659     }
1660     {
1661         unique_lock<mutex> lock(frame_queue_mutex);
1662         timestamps.erase(task.encode_order - (ip_period - 1));
1663     }
1664
1665 #if 0
1666     printf("\r      "); /* return back to startpoint */
1667     switch (encode_order % 4) {
1668         case 0:
1669             printf("|");
1670             break;
1671         case 1:
1672             printf("/");
1673             break;
1674         case 2:
1675             printf("-");
1676             break;
1677         case 3:
1678             printf("\\");
1679             break;
1680     }
1681     printf("%08lld", encode_order);
1682     printf("(%06d bytes coded)", coded_size);
1683 #endif
1684
1685     return 0;
1686 }
1687
1688
1689 // this is weird. but it seems to put a new frame onto the queue
1690 void H264Encoder::storage_task_enqueue(storage_task task)
1691 {
1692         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1693         storage_task_queue.push(move(task));
1694         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1695         storage_task_queue_changed.notify_all();
1696 }
1697
1698 void H264Encoder::storage_task_thread()
1699 {
1700         for ( ;; ) {
1701                 storage_task current;
1702                 {
1703                         // wait until there's an encoded frame
1704                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1705                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1706                         if (storage_thread_should_quit) return;
1707                         current = move(storage_task_queue.front());
1708                         storage_task_queue.pop();
1709                 }
1710
1711                 VAStatus va_status;
1712
1713                 // waits for data, then saves it to disk.
1714                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1715                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1716                 save_codeddata(move(current));
1717
1718                 {
1719                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1720                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1721                         storage_task_queue_changed.notify_all();
1722                 }
1723         }
1724 }
1725
1726 static int release_encode()
1727 {
1728     int i;
1729
1730     for (i = 0; i < SURFACE_NUM; i++) {
1731         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1732         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1733         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1734     }
1735
1736     vaDestroyContext(va_dpy, context_id);
1737     vaDestroyConfig(va_dpy, config_id);
1738
1739     return 0;
1740 }
1741
1742 static int deinit_va()
1743 {
1744     vaTerminate(va_dpy);
1745
1746     va_close_display(va_dpy);
1747
1748     return 0;
1749 }
1750
1751
1752 static int print_input()
1753 {
1754     printf("\n\nINPUT:Try to encode H264...\n");
1755     if (rc_mode != -1)
1756         printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1757     printf("INPUT: Resolution   : %dx%dframes\n", frame_width, frame_height);
1758     printf("INPUT: FrameRate    : %d\n", frame_rate);
1759     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
1760     printf("INPUT: Slieces      : %d\n", frame_slices);
1761     printf("INPUT: IntraPeriod  : %d\n", intra_period);
1762     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
1763     printf("INPUT: IpPeriod     : %d\n", ip_period);
1764     printf("INPUT: Initial QP   : %d\n", initial_qp);
1765     printf("INPUT: Min QP       : %d\n", minimal_qp);
1766     printf("INPUT: Coded Clip   : %s\n", coded_fn);
1767
1768     printf("\n\n"); /* return back to startpoint */
1769
1770     return 0;
1771 }
1772
1773
1774 //H264Encoder::H264Encoder(SDL_Window *window, SDL_GLContext context, int width, int height, const char *output_filename)
1775 H264Encoder::H264Encoder(QSurface *surface, int width, int height, const char *output_filename)
1776         : current_storage_frame(0), surface(surface)
1777         //: width(width), height(height), current_encoding_frame(0)
1778 {
1779         av_register_all();
1780         avctx = avformat_alloc_context();
1781         avctx->oformat = av_guess_format(NULL, output_filename, NULL);
1782         strcpy(avctx->filename, output_filename);
1783         if (avio_open2(&avctx->pb, output_filename, AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL) < 0) {
1784                 fprintf(stderr, "%s: avio_open2() failed\n", output_filename);
1785                 exit(1);
1786         }
1787         AVCodec *codec_video = avcodec_find_encoder(AV_CODEC_ID_H264);
1788         avstream_video = avformat_new_stream(avctx, codec_video);
1789         if (avstream_video == nullptr) {
1790                 fprintf(stderr, "%s: avformat_new_stream() failed\n", output_filename);
1791                 exit(1);
1792         }
1793         avstream_video->time_base = AVRational{1, TIMEBASE};
1794         avstream_video->codec->width = width;
1795         avstream_video->codec->height = height;
1796         avstream_video->codec->time_base = AVRational{1, TIMEBASE};
1797         avstream_video->codec->ticks_per_frame = 1;  // or 2?
1798
1799         AVCodec *codec_audio = avcodec_find_encoder(AV_CODEC_ID_MP3);
1800         avstream_audio = avformat_new_stream(avctx, codec_audio);
1801         if (avstream_audio == nullptr) {
1802                 fprintf(stderr, "%s: avformat_new_stream() failed\n", output_filename);
1803                 exit(1);
1804         }
1805         avstream_audio->time_base = AVRational{1, TIMEBASE};
1806         avstream_audio->codec->bit_rate = 256000;
1807         avstream_audio->codec->sample_rate = 48000;
1808         avstream_audio->codec->sample_fmt = AV_SAMPLE_FMT_FLTP;
1809         avstream_audio->codec->channels = 2;
1810         avstream_audio->codec->channel_layout = AV_CH_LAYOUT_STEREO;
1811         avstream_audio->codec->time_base = AVRational{1, TIMEBASE};
1812
1813         /* open it */
1814         if (avcodec_open2(avstream_audio->codec, codec_audio, NULL) < 0) {
1815                 fprintf(stderr, "Could not open codec\n");
1816                 exit(1);
1817         }
1818
1819         if (avformat_write_header(avctx, NULL) < 0) {
1820                 fprintf(stderr, "%s: avformat_write_header() failed\n", output_filename);
1821                 exit(1);
1822         }
1823
1824         frame_width = width;
1825         frame_height = height;
1826         frame_width_mbaligned = (frame_width + 15) & (~15);
1827         frame_height_mbaligned = (frame_height + 15) & (~15);
1828         frame_bitrate = 15000000;  // / 60;
1829         current_frame_encoding = 0;
1830
1831         print_input();
1832
1833         init_va();
1834         setup_encode();
1835
1836         // No frames are ready yet.
1837         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1838
1839         memset(&seq_param, 0, sizeof(seq_param));
1840         memset(&pic_param, 0, sizeof(pic_param));
1841         memset(&slice_param, 0, sizeof(slice_param));
1842
1843         storage_thread = std::thread(&H264Encoder::storage_task_thread, this);
1844
1845         copy_thread = std::thread([this]{
1846                 //SDL_GL_MakeCurrent(window, context);
1847                 QOpenGLContext *context = create_context();
1848                 eglBindAPI(EGL_OPENGL_API);
1849                 if (!make_current(context, this->surface)) {
1850                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1851                                 eglGetError());
1852                         exit(1);
1853                 }
1854                 copy_thread_func();
1855         });
1856 }
1857
1858 H264Encoder::~H264Encoder()
1859 {
1860         {
1861                 unique_lock<mutex> lock(storage_task_queue_mutex);
1862                 storage_thread_should_quit = true;
1863                 storage_task_queue_changed.notify_all();
1864         }
1865         {
1866                 unique_lock<mutex> lock(frame_queue_mutex);
1867                 copy_thread_should_quit = true;
1868                 frame_queue_nonempty.notify_one();
1869         }
1870         storage_thread.join();
1871         copy_thread.join();
1872
1873         release_encode();
1874         deinit_va();
1875
1876         av_write_trailer(avctx);
1877         avformat_free_context(avctx);
1878 }
1879
1880 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1881 {
1882         {
1883                 // Wait until this frame slot is done encoding.
1884                 std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1885                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1886                 if (storage_thread_should_quit) return false;
1887         }
1888
1889         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1890         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1891         *y_tex = surf->y_tex;
1892         *cbcr_tex = surf->cbcr_tex;
1893
1894         VASurfaceID surface = surf->src_surface;
1895         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1896         CHECK_VASTATUS(va_status, "vaDeriveImage");
1897
1898         VABufferInfo buf_info;
1899         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1900         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1901         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1902
1903         // Create Y image.
1904         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1905         EGLint y_attribs[] = {
1906                 EGL_WIDTH, frame_width,
1907                 EGL_HEIGHT, frame_height,
1908                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1909                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1910                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1911                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1912                 EGL_NONE
1913         };
1914
1915         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1916         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1917
1918         // Associate Y image to a texture.
1919         glBindTexture(GL_TEXTURE_2D, *y_tex);
1920         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1921
1922         // Create CbCr image.
1923         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1924         EGLint cbcr_attribs[] = {
1925                 EGL_WIDTH, frame_width,
1926                 EGL_HEIGHT, frame_height,
1927                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1928                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1929                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1930                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1931                 EGL_NONE
1932         };
1933
1934         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1935         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1936
1937         // Associate CbCr image to a texture.
1938         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1939         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1940
1941         return true;
1942 }
1943
1944 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
1945 {
1946         {
1947                 unique_lock<mutex> lock(frame_queue_mutex);
1948                 pending_audio_frames[pts] = move(audio);
1949         }
1950         frame_queue_nonempty.notify_one();
1951 }
1952
1953
1954 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
1955 {
1956         {
1957                 unique_lock<mutex> lock(frame_queue_mutex);
1958                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames };
1959                 timestamps[current_storage_frame] = pts;
1960                 ++current_storage_frame;
1961         }
1962         frame_queue_nonempty.notify_one();
1963 }
1964
1965 void H264Encoder::copy_thread_func()
1966 {
1967         for ( ;; ) {
1968                 PendingFrame frame;
1969                 encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
1970                                        &current_frame_display, &current_frame_type);
1971                 if (current_frame_type == FRAME_IDR) {
1972                         numShortTerm = 0;
1973                         current_frame_num = 0;
1974                         current_IDR_display = current_frame_display;
1975                 }
1976
1977                 {
1978                         unique_lock<mutex> lock(frame_queue_mutex);
1979                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; });
1980                         if (copy_thread_should_quit) return;
1981                         frame = move(pending_video_frames[current_frame_display]);
1982                         pending_video_frames.erase(current_frame_display);
1983                 }
1984
1985                 // Wait for the GPU to be done with the frame.
1986                 glClientWaitSync(frame.fence.get(), 0, 0);
1987
1988                 // Release back any input frames we needed to render this frame.
1989                 frame.input_frames.clear();
1990
1991                 // Unmap the image.
1992                 GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
1993                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1994                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1995                 VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1996                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1997                 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1998                 CHECK_VASTATUS(va_status, "vaDestroyImage");
1999
2000                 VASurfaceID surface = surf->src_surface;
2001
2002                 // Schedule the frame for encoding.
2003                 va_status = vaBeginPicture(va_dpy, context_id, surface);
2004                 CHECK_VASTATUS(va_status, "vaBeginPicture");
2005
2006                 if (current_frame_type == FRAME_IDR) {
2007                         render_sequence();
2008                         render_picture();
2009                         if (h264_packedheader) {
2010                                 render_packedsequence();
2011                                 render_packedpicture();
2012                         }
2013                 } else {
2014                         //render_sequence();
2015                         render_picture();
2016                 }
2017                 render_slice();
2018
2019                 va_status = vaEndPicture(va_dpy, context_id);
2020                 CHECK_VASTATUS(va_status, "vaEndPicture");
2021
2022                 // so now the data is done encoding (well, async job kicked off)...
2023                 // we send that to the storage thread
2024                 storage_task tmp;
2025                 tmp.display_order = current_frame_display;
2026                 tmp.encode_order = current_frame_encoding;
2027                 tmp.frame_type = current_frame_type;
2028                 storage_task_enqueue(move(tmp));
2029
2030                 update_ReferenceFrames();
2031                 ++current_frame_encoding;
2032         }
2033 }