git.sesse.net Git - nageru/blob - h264encode.cpp

   1 //#include "sysdeps.h"
   2 #include "h264encode.h"
   3
   4 #include <EGL/eglplatform.h>
   5 #include <X11/X.h>
   6 #include <X11/Xlib.h>
   7 #include <assert.h>
   8 #include <epoxy/egl.h>
   9 #include <libavcodec/avcodec.h>
  10 #include <libavutil/channel_layout.h>
  11 #include <libavutil/frame.h>
  12 #include <libavutil/rational.h>
  13 #include <libavutil/samplefmt.h>
  14 #include <libdrm/drm_fourcc.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <string.h>
  18 #include <va/va.h>
  19 #include <va/va_drmcommon.h>
  20 #include <va/va_enc_h264.h>
  21 #include <va/va_x11.h>
  22 #include <condition_variable>
  23 #include <cstdint>
  24 #include <memory>
  25 #include <mutex>
  26 #include <queue>
  27 #include <string>
  28 #include <thread>
  29 #include <utility>
  30
  31 #include "context.h"
  32 #include "defs.h"
  33 #include "httpd.h"
  34 #include "timebase.h"
  35
  36 class QOpenGLContext;
  37 class QSurface;
  38
  39 #define CHECK_VASTATUS(va_status, func)                                 \
  40     if (va_status != VA_STATUS_SUCCESS) {                               \
  41         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  42         exit(1);                                                        \
  43     }
  44
  45 //#include "loadsurface.h"
  46
  47 #define NAL_REF_IDC_NONE        0
  48 #define NAL_REF_IDC_LOW         1
  49 #define NAL_REF_IDC_MEDIUM      2
  50 #define NAL_REF_IDC_HIGH        3
  51
  52 #define NAL_NON_IDR             1
  53 #define NAL_IDR                 5
  54 #define NAL_SPS                 7
  55 #define NAL_PPS                 8
  56 #define NAL_SEI                 6
  57
  58 #define SLICE_TYPE_P            0
  59 #define SLICE_TYPE_B            1
  60 #define SLICE_TYPE_I            2
  61 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  62 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  63 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  64
  65
  66 #define ENTROPY_MODE_CAVLC      0
  67 #define ENTROPY_MODE_CABAC      1
  68
  69 #define PROFILE_IDC_BASELINE    66
  70 #define PROFILE_IDC_MAIN        77
  71 #define PROFILE_IDC_HIGH        100
  72
  73 #define BITSTREAM_ALLOCATE_STEPPING     4096
  74
  75 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  76 static  VADisplay va_dpy;
  77 static  VAProfile h264_profile = (VAProfile)~0;
  78 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  79 static  int config_attrib_num = 0, enc_packed_header_idx;
  80
  81 struct GLSurface {
  82         VASurfaceID src_surface, ref_surface;
  83         VABufferID coded_buf;
  84
  85         VAImage surface_image;
  86         GLuint y_tex, cbcr_tex;
  87         EGLImage y_egl_image, cbcr_egl_image;
  88 };
  89 GLSurface gl_surfaces[SURFACE_NUM];
  90
  91 static  VAConfigID config_id;
  92 static  VAContextID context_id;
  93 static  VAEncSequenceParameterBufferH264 seq_param;
  94 static  VAEncPictureParameterBufferH264 pic_param;
  95 static  VAEncSliceParameterBufferH264 slice_param;
  96 static  VAPictureH264 CurrentCurrPic;
  97 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
  98
  99 static  unsigned int MaxFrameNum = (2<<16);
 100 static  unsigned int MaxPicOrderCntLsb = (2<<8);
 101 static  unsigned int Log2MaxFrameNum = 16;
 102 static  unsigned int Log2MaxPicOrderCntLsb = 8;
 103
 104 static  unsigned int num_ref_frames = 2;
 105 static  unsigned int numShortTerm = 0;
 106 static  int constraint_set_flag = 0;
 107 static  int h264_packedheader = 0; /* support pack header? */
 108 static  int h264_maxref = (1<<16|1);
 109 static  int h264_entropy_mode = 1; /* cabac */
 110
 111 static  char *coded_fn = NULL;
 112
 113 static  int frame_width = 176;
 114 static  int frame_height = 144;
 115 static  int frame_width_mbaligned;
 116 static  int frame_height_mbaligned;
 117 static  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 = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 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 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 158 // but if we don't delete it here, we get leaks. The GStreamer implementation
 159 // does the same.
 160 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 161 {
 162     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 163     CHECK_VASTATUS(va_status, "vaRenderPicture");
 164
 165     for (int i = 0; i < num_buffers; ++i) {
 166         va_status = vaDestroyBuffer(va_dpy, buffers[i]);
 167         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 168     }
 169 }
 170
 171 static unsigned int
 172 va_swap32(unsigned int val)
 173 {
 174     unsigned char *pval = (unsigned char *)&val;
 175
 176     return ((pval[0] << 24)     |
 177             (pval[1] << 16)     |
 178             (pval[2] << 8)      |
 179             (pval[3] << 0));
 180 }
 181
 182 static void
 183 bitstream_start(bitstream *bs)
 184 {
 185     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 186     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 187     bs->bit_offset = 0;
 188 }
 189
 190 static void
 191 bitstream_end(bitstream *bs)
 192 {
 193     int pos = (bs->bit_offset >> 5);
 194     int bit_offset = (bs->bit_offset & 0x1f);
 195     int bit_left = 32 - bit_offset;
 196
 197     if (bit_offset) {
 198         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 199     }
 200 }
 201
 202 static void
 203 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 204 {
 205     int pos = (bs->bit_offset >> 5);
 206     int bit_offset = (bs->bit_offset & 0x1f);
 207     int bit_left = 32 - bit_offset;
 208
 209     if (!size_in_bits)
 210         return;
 211
 212     bs->bit_offset += size_in_bits;
 213
 214     if (bit_left > size_in_bits) {
 215         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 216     } else {
 217         size_in_bits -= bit_left;
 218         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 219         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 220
 221         if (pos + 1 == bs->max_size_in_dword) {
 222             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 223             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 224         }
 225
 226         bs->buffer[pos + 1] = val;
 227     }
 228 }
 229
 230 static void
 231 bitstream_put_ue(bitstream *bs, unsigned int val)
 232 {
 233     int size_in_bits = 0;
 234     int tmp_val = ++val;
 235
 236     while (tmp_val) {
 237         tmp_val >>= 1;
 238         size_in_bits++;
 239     }
 240
 241     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 242     bitstream_put_ui(bs, val, size_in_bits);
 243 }
 244
 245 static void
 246 bitstream_put_se(bitstream *bs, int val)
 247 {
 248     unsigned int new_val;
 249
 250     if (val <= 0)
 251         new_val = -2 * val;
 252     else
 253         new_val = 2 * val - 1;
 254
 255     bitstream_put_ue(bs, new_val);
 256 }
 257
 258 static void
 259 bitstream_byte_aligning(bitstream *bs, int bit)
 260 {
 261     int bit_offset = (bs->bit_offset & 0x7);
 262     int bit_left = 8 - bit_offset;
 263     int new_val;
 264
 265     if (!bit_offset)
 266         return;
 267
 268     assert(bit == 0 || bit == 1);
 269
 270     if (bit)
 271         new_val = (1 << bit_left) - 1;
 272     else
 273         new_val = 0;
 274
 275     bitstream_put_ui(bs, new_val, bit_left);
 276 }
 277
 278 static void
 279 rbsp_trailing_bits(bitstream *bs)
 280 {
 281     bitstream_put_ui(bs, 1, 1);
 282     bitstream_byte_aligning(bs, 0);
 283 }
 284
 285 static void nal_start_code_prefix(bitstream *bs)
 286 {
 287     bitstream_put_ui(bs, 0x00000001, 32);
 288 }
 289
 290 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 291 {
 292     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 293     bitstream_put_ui(bs, nal_ref_idc, 2);
 294     bitstream_put_ui(bs, nal_unit_type, 5);
 295 }
 296
 297 static void sps_rbsp(bitstream *bs)
 298 {
 299     int profile_idc = PROFILE_IDC_BASELINE;
 300
 301     if (h264_profile  == VAProfileH264High)
 302         profile_idc = PROFILE_IDC_HIGH;
 303     else if (h264_profile  == VAProfileH264Main)
 304         profile_idc = PROFILE_IDC_MAIN;
 305
 306     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 307     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 308     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 309     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 310     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 311     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 312     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 313     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 314
 315     if ( profile_idc == PROFILE_IDC_HIGH) {
 316         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 317         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 318         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 319         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 320         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 321     }
 322
 323     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 324     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 325
 326     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 327         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 328     else {
 329         assert(0);
 330     }
 331
 332     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 333     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 334
 335     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 336     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 337     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 338
 339     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 340         assert(0);
 341     }
 342
 343     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 344     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 345
 346     if (seq_param.frame_cropping_flag) {
 347         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 348         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 349         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 350         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 351     }
 352
 353     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 354     if ( false ) {
 355         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 356     } else {
 357         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 358         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 359         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 360         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 361         {
 362             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 363             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 364             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 365             {
 366                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 367                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 368                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 369             }
 370         }
 371         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 372         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 373         {
 374             bitstream_put_ui(bs, 1, 32);  // FPS
 375             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 376             bitstream_put_ui(bs, 1, 1);
 377         }
 378         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 379         {
 380             // hrd_parameters
 381             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 382             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 383             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 384
 385             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 386             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 387             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 388
 389             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 390             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 391             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 392             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 393         }
 394         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 395         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 396
 397         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 398         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 399     }
 400
 401     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 402 }
 403
 404
 405 static void pps_rbsp(bitstream *bs)
 406 {
 407     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 408     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 409
 410     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 411
 412     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 413
 414     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 415
 416     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 417     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 418
 419     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 420     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 421
 422     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 423     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 424     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 425
 426     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 427     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 428     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 429
 430     /* more_rbsp_data */
 431     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 432     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 433     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 434
 435     rbsp_trailing_bits(bs);
 436 }
 437
 438 static void slice_header(bitstream *bs)
 439 {
 440     int first_mb_in_slice = slice_param.macroblock_address;
 441
 442     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 443     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 444     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 445     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 446
 447     /* frame_mbs_only_flag == 1 */
 448     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 449         /* FIXME: */
 450         assert(0);
 451     }
 452
 453     if (pic_param.pic_fields.bits.idr_pic_flag)
 454         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 455
 456     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 457         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 458         /* pic_order_present_flag == 0 */
 459     } else {
 460         /* FIXME: */
 461         assert(0);
 462     }
 463
 464     /* redundant_pic_cnt_present_flag == 0 */
 465     /* slice type */
 466     if (IS_P_SLICE(slice_param.slice_type)) {
 467         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 468
 469         if (slice_param.num_ref_idx_active_override_flag)
 470             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 471
 472         /* ref_pic_list_reordering */
 473         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 474     } else if (IS_B_SLICE(slice_param.slice_type)) {
 475         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 476
 477         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 478
 479         if (slice_param.num_ref_idx_active_override_flag) {
 480             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 481             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 482         }
 483
 484         /* ref_pic_list_reordering */
 485         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 486         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 487     }
 488
 489     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 490          IS_P_SLICE(slice_param.slice_type)) ||
 491         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 492          IS_B_SLICE(slice_param.slice_type))) {
 493         /* FIXME: fill weight/offset table */
 494         assert(0);
 495     }
 496
 497     /* dec_ref_pic_marking */
 498     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 499         unsigned char no_output_of_prior_pics_flag = 0;
 500         unsigned char long_term_reference_flag = 0;
 501         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 502
 503         if (pic_param.pic_fields.bits.idr_pic_flag) {
 504             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 505             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 506         } else {
 507             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 508         }
 509     }
 510
 511     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 512         !IS_I_SLICE(slice_param.slice_type))
 513         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 514
 515     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 516
 517     /* ignore for SP/SI */
 518
 519     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 520         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 521
 522         if (slice_param.disable_deblocking_filter_idc != 1) {
 523             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 524             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 525         }
 526     }
 527
 528     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 529         bitstream_byte_aligning(bs, 1);
 530     }
 531 }
 532
 533 static int
 534 build_packed_pic_buffer(unsigned char **header_buffer)
 535 {
 536     bitstream bs;
 537
 538     bitstream_start(&bs);
 539     nal_start_code_prefix(&bs);
 540     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 541     pps_rbsp(&bs);
 542     bitstream_end(&bs);
 543
 544     *header_buffer = (unsigned char *)bs.buffer;
 545     return bs.bit_offset;
 546 }
 547
 548 static int
 549 build_packed_seq_buffer(unsigned char **header_buffer)
 550 {
 551     bitstream bs;
 552
 553     bitstream_start(&bs);
 554     nal_start_code_prefix(&bs);
 555     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 556     sps_rbsp(&bs);
 557     bitstream_end(&bs);
 558
 559     *header_buffer = (unsigned char *)bs.buffer;
 560     return bs.bit_offset;
 561 }
 562
 563 static int build_packed_slice_buffer(unsigned char **header_buffer)
 564 {
 565     bitstream bs;
 566     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 567     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 568
 569     bitstream_start(&bs);
 570     nal_start_code_prefix(&bs);
 571
 572     if (IS_I_SLICE(slice_param.slice_type)) {
 573         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 574     } else if (IS_P_SLICE(slice_param.slice_type)) {
 575         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 576     } else {
 577         assert(IS_B_SLICE(slice_param.slice_type));
 578         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 579     }
 580
 581     slice_header(&bs);
 582     bitstream_end(&bs);
 583
 584     *header_buffer = (unsigned char *)bs.buffer;
 585     return bs.bit_offset;
 586 }
 587
 588
 589 /*
 590   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 591   1) period between Frame #X and Frame #N = #X - #N
 592   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 593   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 594   4) intra_period and intra_idr_period take precedence over ip_period
 595   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 596      of I/IDR frames, see bellow examples
 597   -------------------------------------------------------------------
 598   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 599   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 600   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 601   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 602   1            1                ignored    IDR IDR IDR IDR...
 603   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 604   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 605   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 606                                               (PBB)(IBB)(PBB)(IBB)...
 607   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 608                                            IDRPPPPP IPPPPP IPPPPP...
 609   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 610                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 611                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 612                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 613                                            {IDR(PBB)(PBB)}                     (6/6/3)
 614                                            {IDR(PBB)(PBB)}.
 615 */
 616
 617 // General pts/dts strategy:
 618 //
 619 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 620 // bit tricky. We assume first of all that the frame rate never goes _above_
 621 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 622 // in at least this speed, as long at dts <= pts (the frame is not attempted
 623 // presented before it is decoded). Furthermore, we never have longer chains of
 624 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 625 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 626 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 627 // frame, with an order O <= C.)
 628 //
 629 // Many strategies are possible, but we establish these rules:
 630 //
 631 //  - Tip frames have dts = pts - (C-O)*N.
 632 //  - Non-tip frames have dts = dts_last + N.
 633 //
 634 // An example, with C=2 and N=10 and the data flow showed with arrows:
 635 //
 636 //        I  B  P  B  B  P
 637 //   pts: 30 40 50 60 70 80
 638 //        ↓  ↓     ↓
 639 //   dts: 10 30 20 60 50←40
 640 //         |  |  ↑        ↑
 641 //         `--|--'        |
 642 //             `----------'
 643 //
 644 // To show that this works fine also with irregular spacings, let's say that
 645 // the third frame is delayed a bit (something earlier was dropped). Now the
 646 // situation looks like this:
 647 //
 648 //        I  B  P  B  B   P
 649 //   pts: 30 40 80 90 100 110
 650 //        ↓  ↓     ↓
 651 //   dts: 10 30 20 90 50←40
 652 //         |  |  ↑        ↑
 653 //         `--|--'        |
 654 //             `----------'
 655 //
 656 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 657 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 658 //
 659 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 660 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 661 // a tip frame and should be given a dts based on the previous one.
 662 #define FRAME_P 0
 663 #define FRAME_B 1
 664 #define FRAME_I 2
 665 #define FRAME_IDR 7
 666 void encoding2display_order(
 667     unsigned long long encoding_order, int intra_period,
 668     int intra_idr_period, int ip_period,
 669     unsigned long long *displaying_order,
 670     int *frame_type, int *pts_lag)
 671 {
 672     int encoding_order_gop = 0;
 673
 674     *pts_lag = 0;
 675
 676     if (intra_period == 1) { /* all are I/IDR frames */
 677         *displaying_order = encoding_order;
 678         if (intra_idr_period == 0)
 679             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 680         else
 681             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 682         return;
 683     }
 684
 685     if (intra_period == 0)
 686         intra_idr_period = 0;
 687
 688     if (ip_period == 1) {
 689         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 690         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 691         *displaying_order = encoding_order;
 692
 693         if (encoding_order_gop == 0) { /* the first frame */
 694             *frame_type = FRAME_IDR;
 695         } else if (intra_period != 0 && /* have I frames */
 696                    encoding_order_gop >= 2 &&
 697                    (encoding_order_gop % intra_period == 0)) {
 698             *frame_type = FRAME_I;
 699         } else {
 700             *frame_type = FRAME_P;
 701         }
 702         return;
 703     }
 704
 705     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 706     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 707     *pts_lag = -1;  // Most frames are not tip frames.
 708
 709     if (encoding_order_gop == 0) { /* the first frame */
 710         *frame_type = FRAME_IDR;
 711         *displaying_order = encoding_order;
 712         // IDR frames are a special case; I honestly can't find the logic behind
 713         // why this is the right thing, but it seems to line up nicely in practice :-)
 714         *pts_lag = TIMEBASE / MAX_FPS;
 715     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 716         *frame_type = FRAME_B;
 717         *displaying_order = encoding_order - 1;
 718         if ((encoding_order_gop % ip_period) == 0) {
 719             *pts_lag = 0;  // Last B-frame.
 720         }
 721     } else if (intra_period != 0 && /* have I frames */
 722                encoding_order_gop >= 2 &&
 723                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 724         *frame_type = FRAME_I;
 725         *displaying_order = encoding_order + ip_period - 1;
 726     } else {
 727         *frame_type = FRAME_P;
 728         *displaying_order = encoding_order + ip_period - 1;
 729     }
 730 }
 731
 732
 733 static const char *rc_to_string(int rcmode)
 734 {
 735     switch (rc_mode) {
 736     case VA_RC_NONE:
 737         return "NONE";
 738     case VA_RC_CBR:
 739         return "CBR";
 740     case VA_RC_VBR:
 741         return "VBR";
 742     case VA_RC_VCM:
 743         return "VCM";
 744     case VA_RC_CQP:
 745         return "CQP";
 746     case VA_RC_VBR_CONSTRAINED:
 747         return "VBR_CONSTRAINED";
 748     default:
 749         return "Unknown";
 750     }
 751 }
 752
 753 #if 0
 754 static int process_cmdline(int argc, char *argv[])
 755 {
 756     char c;
 757     const struct option long_opts[] = {
 758         {"help", no_argument, NULL, 0 },
 759         {"bitrate", required_argument, NULL, 1 },
 760         {"minqp", required_argument, NULL, 2 },
 761         {"initialqp", required_argument, NULL, 3 },
 762         {"intra_period", required_argument, NULL, 4 },
 763         {"idr_period", required_argument, NULL, 5 },
 764         {"ip_period", required_argument, NULL, 6 },
 765         {"rcmode", required_argument, NULL, 7 },
 766         {"srcyuv", required_argument, NULL, 9 },
 767         {"recyuv", required_argument, NULL, 10 },
 768         {"fourcc", required_argument, NULL, 11 },
 769         {"syncmode", no_argument, NULL, 12 },
 770         {"enablePSNR", no_argument, NULL, 13 },
 771         {"prit", required_argument, NULL, 14 },
 772         {"priv", required_argument, NULL, 15 },
 773         {"framecount", required_argument, NULL, 16 },
 774         {"entropy", required_argument, NULL, 17 },
 775         {"profile", required_argument, NULL, 18 },
 776         {NULL, no_argument, NULL, 0 }};
 777     int long_index;
 778
 779     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 780         switch (c) {
 781         case 'w':
 782             frame_width = atoi(optarg);
 783             break;
 784         case 'h':
 785             frame_height = atoi(optarg);
 786             break;
 787         case 'n':
 788         case 'f':
 789             frame_rate = atoi(optarg);
 790             break;
 791         case 'o':
 792             coded_fn = strdup(optarg);
 793             break;
 794         case 0:
 795             print_help();
 796             exit(0);
 797         case 1:
 798             frame_bitrate = atoi(optarg);
 799             break;
 800         case 2:
 801             minimal_qp = atoi(optarg);
 802             break;
 803         case 3:
 804             initial_qp = atoi(optarg);
 805             break;
 806         case 4:
 807             intra_period = atoi(optarg);
 808             break;
 809         case 5:
 810             intra_idr_period = atoi(optarg);
 811             break;
 812         case 6:
 813             ip_period = atoi(optarg);
 814             break;
 815         case 7:
 816             rc_mode = string_to_rc(optarg);
 817             if (rc_mode < 0) {
 818                 print_help();
 819                 exit(1);
 820             }
 821             break;
 822         case 9:
 823             srcyuv_fn = strdup(optarg);
 824             break;
 825         case 11:
 826             srcyuv_fourcc = string_to_fourcc(optarg);
 827             if (srcyuv_fourcc <= 0) {
 828                 print_help();
 829                 exit(1);
 830             }
 831             break;
 832         case 13:
 833             calc_psnr = 1;
 834             break;
 835         case 14:
 836             misc_priv_type = strtol(optarg, NULL, 0);
 837             break;
 838         case 15:
 839             misc_priv_value = strtol(optarg, NULL, 0);
 840             break;
 841         case 17:
 842             h264_entropy_mode = atoi(optarg) ? 1: 0;
 843             break;
 844         case 18:
 845             if (strncmp(optarg, "BP", 2) == 0)
 846                 h264_profile = VAProfileH264Baseline;
 847             else if (strncmp(optarg, "MP", 2) == 0)
 848                 h264_profile = VAProfileH264Main;
 849             else if (strncmp(optarg, "HP", 2) == 0)
 850                 h264_profile = VAProfileH264High;
 851             else
 852                 h264_profile = (VAProfile)0;
 853             break;
 854         case ':':
 855         case '?':
 856             print_help();
 857             exit(0);
 858         }
 859     }
 860
 861     if (ip_period < 1) {
 862         printf(" ip_period must be greater than 0\n");
 863         exit(0);
 864     }
 865     if (intra_period != 1 && intra_period % ip_period != 0) {
 866         printf(" intra_period must be a multiplier of ip_period\n");
 867         exit(0);
 868     }
 869     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 870         printf(" intra_idr_period must be a multiplier of intra_period\n");
 871         exit(0);
 872     }
 873
 874     if (frame_bitrate == 0)
 875         frame_bitrate = frame_width * frame_height * 12 * MAX_FPS / 50;
 876
 877     if (coded_fn == NULL) {
 878         struct stat buf;
 879         if (stat("/tmp", &buf) == 0)
 880             coded_fn = strdup("/tmp/test.264");
 881         else if (stat("/sdcard", &buf) == 0)
 882             coded_fn = strdup("/sdcard/test.264");
 883         else
 884             coded_fn = strdup("./test.264");
 885     }
 886
 887
 888     frame_width_mbaligned = (frame_width + 15) & (~15);
 889     frame_height_mbaligned = (frame_height + 15) & (~15);
 890     if (frame_width != frame_width_mbaligned ||
 891         frame_height != frame_height_mbaligned) {
 892         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 893                frame_width, frame_height,
 894                frame_width_mbaligned, frame_height_mbaligned
 895                );
 896     }
 897
 898     return 0;
 899 }
 900 #endif
 901
 902 static Display *x11_display;
 903 static Window   x11_window;
 904
 905 VADisplay
 906 va_open_display(void)
 907 {
 908     x11_display = XOpenDisplay(NULL);
 909     if (!x11_display) {
 910         fprintf(stderr, "error: can't connect to X server!\n");
 911         return NULL;
 912     }
 913     return vaGetDisplay(x11_display);
 914 }
 915
 916 void
 917 va_close_display(VADisplay va_dpy)
 918 {
 919     if (!x11_display)
 920         return;
 921
 922     if (x11_window) {
 923         XUnmapWindow(x11_display, x11_window);
 924         XDestroyWindow(x11_display, x11_window);
 925         x11_window = None;
 926     }
 927     XCloseDisplay(x11_display);
 928     x11_display = NULL;
 929 }
 930
 931 static int init_va(void)
 932 {
 933     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 934     VAEntrypoint *entrypoints;
 935     int num_entrypoints, slice_entrypoint;
 936     int support_encode = 0;
 937     int major_ver, minor_ver;
 938     VAStatus va_status;
 939     unsigned int i;
 940
 941     va_dpy = va_open_display();
 942     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 943     CHECK_VASTATUS(va_status, "vaInitialize");
 944
 945     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 946     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 947     if (!entrypoints) {
 948         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 949         exit(1);
 950     }
 951
 952     /* use the highest profile */
 953     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 954         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 955             continue;
 956
 957         h264_profile = profile_list[i];
 958         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 959         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 960             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 961                 support_encode = 1;
 962                 break;
 963             }
 964         }
 965         if (support_encode == 1)
 966             break;
 967     }
 968
 969     if (support_encode == 0) {
 970         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 971         exit(1);
 972     } else {
 973         switch (h264_profile) {
 974             case VAProfileH264Baseline:
 975                 printf("Use profile VAProfileH264Baseline\n");
 976                 ip_period = 1;
 977                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 978                 h264_entropy_mode = 0;
 979                 break;
 980             case VAProfileH264ConstrainedBaseline:
 981                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 982                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 983                 ip_period = 1;
 984                 break;
 985
 986             case VAProfileH264Main:
 987                 printf("Use profile VAProfileH264Main\n");
 988                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 989                 break;
 990
 991             case VAProfileH264High:
 992                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 993                 printf("Use profile VAProfileH264High\n");
 994                 break;
 995             default:
 996                 printf("unknow profile. Set to Baseline");
 997                 h264_profile = VAProfileH264Baseline;
 998                 ip_period = 1;
 999                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1000                 break;
1001         }
1002     }
1003
1004     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1005
1006     /* find out the format for the render target, and rate control mode */
1007     for (i = 0; i < VAConfigAttribTypeMax; i++)
1008         attrib[i].type = (VAConfigAttribType)i;
1009
1010     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1011                                       &attrib[0], VAConfigAttribTypeMax);
1012     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1013     /* check the interested configattrib */
1014     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1015         printf("Not find desired YUV420 RT format\n");
1016         exit(1);
1017     } else {
1018         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1019         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1020         config_attrib_num++;
1021     }
1022
1023     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1024         int tmp = attrib[VAConfigAttribRateControl].value;
1025
1026         printf("Support rate control mode (0x%x):", tmp);
1027
1028         if (tmp & VA_RC_NONE)
1029             printf("NONE ");
1030         if (tmp & VA_RC_CBR)
1031             printf("CBR ");
1032         if (tmp & VA_RC_VBR)
1033             printf("VBR ");
1034         if (tmp & VA_RC_VCM)
1035             printf("VCM ");
1036         if (tmp & VA_RC_CQP)
1037             printf("CQP ");
1038         if (tmp & VA_RC_VBR_CONSTRAINED)
1039             printf("VBR_CONSTRAINED ");
1040
1041         printf("\n");
1042
1043         if (rc_mode == -1 || !(rc_mode & tmp))  {
1044             if (rc_mode != -1) {
1045                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1046             }
1047
1048             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1049                 if (rc_default_modes[i] & tmp) {
1050                     rc_mode = rc_default_modes[i];
1051                     break;
1052                 }
1053             }
1054
1055             printf("RateControl mode: %s\n", rc_to_string(rc_mode));
1056         }
1057
1058         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1059         config_attrib[config_attrib_num].value = rc_mode;
1060         config_attrib_num++;
1061     }
1062
1063
1064     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1065         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1066
1067         printf("Support VAConfigAttribEncPackedHeaders\n");
1068
1069         h264_packedheader = 1;
1070         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1071         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1072
1073         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1074             printf("Support packed sequence headers\n");
1075             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1076         }
1077
1078         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1079             printf("Support packed picture headers\n");
1080             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1081         }
1082
1083         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1084             printf("Support packed slice headers\n");
1085             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1086         }
1087
1088         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1089             printf("Support packed misc headers\n");
1090             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1091         }
1092
1093         enc_packed_header_idx = config_attrib_num;
1094         config_attrib_num++;
1095     }
1096
1097     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1098         int tmp = attrib[VAConfigAttribEncInterlaced].value;
1099
1100         printf("Support VAConfigAttribEncInterlaced\n");
1101
1102         if (tmp & VA_ENC_INTERLACED_FRAME)
1103             printf("support VA_ENC_INTERLACED_FRAME\n");
1104         if (tmp & VA_ENC_INTERLACED_FIELD)
1105             printf("Support VA_ENC_INTERLACED_FIELD\n");
1106         if (tmp & VA_ENC_INTERLACED_MBAFF)
1107             printf("Support VA_ENC_INTERLACED_MBAFF\n");
1108         if (tmp & VA_ENC_INTERLACED_PAFF)
1109             printf("Support VA_ENC_INTERLACED_PAFF\n");
1110
1111         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1112         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1113         config_attrib_num++;
1114     }
1115
1116     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1117         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1118
1119         printf("Support %d RefPicList0 and %d RefPicList1\n",
1120                h264_maxref & 0xffff, (h264_maxref >> 16) & 0xffff );
1121     }
1122
1123     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1124         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1125
1126     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1127         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1128
1129         printf("Support VAConfigAttribEncSliceStructure\n");
1130
1131         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1132             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1133         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1134             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1135         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1136             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1137     }
1138     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1139         printf("Support VAConfigAttribEncMacroblockInfo\n");
1140     }
1141
1142     free(entrypoints);
1143     return 0;
1144 }
1145
1146 static int setup_encode()
1147 {
1148     VAStatus va_status;
1149     VASurfaceID *tmp_surfaceid;
1150     int codedbuf_size, i;
1151     static VASurfaceID src_surface[SURFACE_NUM];
1152     static VASurfaceID ref_surface[SURFACE_NUM];
1153
1154     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1155             &config_attrib[0], config_attrib_num, &config_id);
1156     CHECK_VASTATUS(va_status, "vaCreateConfig");
1157
1158     /* create source surfaces */
1159     va_status = vaCreateSurfaces(va_dpy,
1160                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1161                                  &src_surface[0], SURFACE_NUM,
1162                                  NULL, 0);
1163     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1164
1165     /* create reference surfaces */
1166     va_status = vaCreateSurfaces(va_dpy,
1167                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1168                                  &ref_surface[0], SURFACE_NUM,
1169                                  NULL, 0);
1170     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1171
1172     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1173     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1174     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1175
1176     /* Create a context for this encode pipe */
1177     va_status = vaCreateContext(va_dpy, config_id,
1178                                 frame_width_mbaligned, frame_height_mbaligned,
1179                                 VA_PROGRESSIVE,
1180                                 tmp_surfaceid, 2 * SURFACE_NUM,
1181                                 &context_id);
1182     CHECK_VASTATUS(va_status, "vaCreateContext");
1183     free(tmp_surfaceid);
1184
1185     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1186
1187     for (i = 0; i < SURFACE_NUM; i++) {
1188         /* create coded buffer once for all
1189          * other VA buffers which won't be used again after vaRenderPicture.
1190          * so APP can always vaCreateBuffer for every frame
1191          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1192          * so VA won't maintain the coded buffer
1193          */
1194         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1195                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1196         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1197     }
1198
1199     /* create OpenGL objects */
1200     //glGenFramebuffers(SURFACE_NUM, fbos);
1201
1202     for (i = 0; i < SURFACE_NUM; i++) {
1203         glGenTextures(1, &gl_surfaces[i].y_tex);
1204         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1205     }
1206
1207     for (i = 0; i < SURFACE_NUM; i++) {
1208         gl_surfaces[i].src_surface = src_surface[i];
1209         gl_surfaces[i].ref_surface = ref_surface[i];
1210     }
1211
1212     return 0;
1213 }
1214
1215
1216
1217 #define partition(ref, field, key, ascending)   \
1218     while (i <= j) {                            \
1219         if (ascending) {                        \
1220             while (ref[i].field < key)          \
1221                 i++;                            \
1222             while (ref[j].field > key)          \
1223                 j--;                            \
1224         } else {                                \
1225             while (ref[i].field > key)          \
1226                 i++;                            \
1227             while (ref[j].field < key)          \
1228                 j--;                            \
1229         }                                       \
1230         if (i <= j) {                           \
1231             tmp = ref[i];                       \
1232             ref[i] = ref[j];                    \
1233             ref[j] = tmp;                       \
1234             i++;                                \
1235             j--;                                \
1236         }                                       \
1237     }                                           \
1238
1239 static void sort_one(VAPictureH264 ref[], int left, int right,
1240                      int ascending, int frame_idx)
1241 {
1242     int i = left, j = right;
1243     unsigned int key;
1244     VAPictureH264 tmp;
1245
1246     if (frame_idx) {
1247         key = ref[(left + right) / 2].frame_idx;
1248         partition(ref, frame_idx, key, ascending);
1249     } else {
1250         key = ref[(left + right) / 2].TopFieldOrderCnt;
1251         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1252     }
1253
1254     /* recursion */
1255     if (left < j)
1256         sort_one(ref, left, j, ascending, frame_idx);
1257
1258     if (i < right)
1259         sort_one(ref, i, right, ascending, frame_idx);
1260 }
1261
1262 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1263                      int partition_ascending, int list0_ascending, int list1_ascending)
1264 {
1265     int i = left, j = right;
1266     VAPictureH264 tmp;
1267
1268     if (frame_idx) {
1269         partition(ref, frame_idx, key, partition_ascending);
1270     } else {
1271         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1272     }
1273
1274
1275     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1276     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1277 }
1278
1279 static int update_ReferenceFrames(void)
1280 {
1281     int i;
1282
1283     if (current_frame_type == FRAME_B)
1284         return 0;
1285
1286     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1287     numShortTerm++;
1288     if (numShortTerm > num_ref_frames)
1289         numShortTerm = num_ref_frames;
1290     for (i=numShortTerm-1; i>0; i--)
1291         ReferenceFrames[i] = ReferenceFrames[i-1];
1292     ReferenceFrames[0] = CurrentCurrPic;
1293
1294     if (current_frame_type != FRAME_B)
1295         current_frame_num++;
1296     if (current_frame_num > MaxFrameNum)
1297         current_frame_num = 0;
1298
1299     return 0;
1300 }
1301
1302
1303 static int update_RefPicList(void)
1304 {
1305     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1306
1307     if (current_frame_type == FRAME_P) {
1308         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1309         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1310     }
1311
1312     if (current_frame_type == FRAME_B) {
1313         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1314         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1315                  1, 0, 1);
1316
1317         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1318         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1319                  0, 1, 0);
1320     }
1321
1322     return 0;
1323 }
1324
1325
1326 static int render_sequence(void)
1327 {
1328     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1329     VAStatus va_status;
1330     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1331     VAEncMiscParameterRateControl *misc_rate_ctrl;
1332
1333     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1334     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1335     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1336     seq_param.bits_per_second = frame_bitrate;
1337
1338     seq_param.intra_period = intra_period;
1339     seq_param.intra_idr_period = intra_idr_period;
1340     seq_param.ip_period = ip_period;
1341
1342     seq_param.max_num_ref_frames = num_ref_frames;
1343     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1344     seq_param.time_scale = TIMEBASE * 2;
1345     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1346     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1347     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1348     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1349     seq_param.seq_fields.bits.chroma_format_idc = 1;
1350     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1351
1352     if (frame_width != frame_width_mbaligned ||
1353         frame_height != frame_height_mbaligned) {
1354         seq_param.frame_cropping_flag = 1;
1355         seq_param.frame_crop_left_offset = 0;
1356         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1357         seq_param.frame_crop_top_offset = 0;
1358         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1359     }
1360
1361     va_status = vaCreateBuffer(va_dpy, context_id,
1362                                VAEncSequenceParameterBufferType,
1363                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1364     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1365
1366     va_status = vaCreateBuffer(va_dpy, context_id,
1367                                VAEncMiscParameterBufferType,
1368                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1369                                1, NULL, &rc_param_buf);
1370     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1371
1372     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1373     misc_param->type = VAEncMiscParameterTypeRateControl;
1374     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1375     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1376     misc_rate_ctrl->bits_per_second = frame_bitrate;
1377     misc_rate_ctrl->target_percentage = 66;
1378     misc_rate_ctrl->window_size = 1000;
1379     misc_rate_ctrl->initial_qp = initial_qp;
1380     misc_rate_ctrl->min_qp = minimal_qp;
1381     misc_rate_ctrl->basic_unit_size = 0;
1382     vaUnmapBuffer(va_dpy, rc_param_buf);
1383
1384     render_id[0] = seq_param_buf;
1385     render_id[1] = rc_param_buf;
1386
1387     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1388
1389     if (misc_priv_type != 0) {
1390         va_status = vaCreateBuffer(va_dpy, context_id,
1391                                    VAEncMiscParameterBufferType,
1392                                    sizeof(VAEncMiscParameterBuffer),
1393                                    1, NULL, &misc_param_tmpbuf);
1394         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1395         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1396         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1397         misc_param_tmp->data[0] = misc_priv_value;
1398         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1399
1400         render_picture_and_delete(va_dpy, context_id, &misc_param_tmpbuf, 1);
1401     }
1402
1403     return 0;
1404 }
1405
1406 static int calc_poc(int pic_order_cnt_lsb)
1407 {
1408     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1409     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1410     int PicOrderCntMsb, TopFieldOrderCnt;
1411
1412     if (current_frame_type == FRAME_IDR)
1413         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1414     else {
1415         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1416         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1417     }
1418
1419     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1420         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1421         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1422     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1423              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1424         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1425     else
1426         PicOrderCntMsb = prevPicOrderCntMsb;
1427
1428     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1429
1430     if (current_frame_type != FRAME_B) {
1431         PicOrderCntMsb_ref = PicOrderCntMsb;
1432         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1433     }
1434
1435     return TopFieldOrderCnt;
1436 }
1437
1438 static int render_picture(void)
1439 {
1440     VABufferID pic_param_buf;
1441     VAStatus va_status;
1442     int i = 0;
1443
1444     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1445     pic_param.CurrPic.frame_idx = current_frame_num;
1446     pic_param.CurrPic.flags = 0;
1447     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1448     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1449     CurrentCurrPic = pic_param.CurrPic;
1450
1451     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1452         update_RefPicList(); /* calc RefPicList */
1453         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1454         if (current_frame_type == FRAME_P) {
1455             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1456         } else if (current_frame_type == FRAME_B) {
1457             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1458             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1459         }
1460     } else {
1461         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1462         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1463             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1464             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1465         }
1466     }
1467
1468     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1469     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1470     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1471     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1472     pic_param.frame_num = current_frame_num;
1473     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1474     pic_param.last_picture = false;  // FIXME
1475     pic_param.pic_init_qp = initial_qp;
1476
1477     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1478                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1479     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1480
1481     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1482
1483     return 0;
1484 }
1485
1486 static int render_packedsequence(void)
1487 {
1488     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1489     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1490     unsigned int length_in_bits;
1491     unsigned char *packedseq_buffer = NULL;
1492     VAStatus va_status;
1493
1494     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1495
1496     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1497
1498     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1499     packedheader_param_buffer.has_emulation_bytes = 0;
1500     va_status = vaCreateBuffer(va_dpy,
1501                                context_id,
1502                                VAEncPackedHeaderParameterBufferType,
1503                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1504                                &packedseq_para_bufid);
1505     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1506
1507     va_status = vaCreateBuffer(va_dpy,
1508                                context_id,
1509                                VAEncPackedHeaderDataBufferType,
1510                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1511                                &packedseq_data_bufid);
1512     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1513
1514     render_id[0] = packedseq_para_bufid;
1515     render_id[1] = packedseq_data_bufid;
1516     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1517
1518     free(packedseq_buffer);
1519
1520     return 0;
1521 }
1522
1523
1524 static int render_packedpicture(void)
1525 {
1526     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1527     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1528     unsigned int length_in_bits;
1529     unsigned char *packedpic_buffer = NULL;
1530     VAStatus va_status;
1531
1532     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1533     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1534     packedheader_param_buffer.bit_length = length_in_bits;
1535     packedheader_param_buffer.has_emulation_bytes = 0;
1536
1537     va_status = vaCreateBuffer(va_dpy,
1538                                context_id,
1539                                VAEncPackedHeaderParameterBufferType,
1540                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1541                                &packedpic_para_bufid);
1542     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1543
1544     va_status = vaCreateBuffer(va_dpy,
1545                                context_id,
1546                                VAEncPackedHeaderDataBufferType,
1547                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1548                                &packedpic_data_bufid);
1549     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1550
1551     render_id[0] = packedpic_para_bufid;
1552     render_id[1] = packedpic_data_bufid;
1553     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1554
1555     free(packedpic_buffer);
1556
1557     return 0;
1558 }
1559
1560 static void render_packedslice()
1561 {
1562     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1563     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1564     unsigned int length_in_bits;
1565     unsigned char *packedslice_buffer = NULL;
1566     VAStatus va_status;
1567
1568     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1569     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1570     packedheader_param_buffer.bit_length = length_in_bits;
1571     packedheader_param_buffer.has_emulation_bytes = 0;
1572
1573     va_status = vaCreateBuffer(va_dpy,
1574                                context_id,
1575                                VAEncPackedHeaderParameterBufferType,
1576                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1577                                &packedslice_para_bufid);
1578     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1579
1580     va_status = vaCreateBuffer(va_dpy,
1581                                context_id,
1582                                VAEncPackedHeaderDataBufferType,
1583                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1584                                &packedslice_data_bufid);
1585     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1586
1587     render_id[0] = packedslice_para_bufid;
1588     render_id[1] = packedslice_data_bufid;
1589     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1590
1591     free(packedslice_buffer);
1592 }
1593
1594 static int render_slice(void)
1595 {
1596     VABufferID slice_param_buf;
1597     VAStatus va_status;
1598     int i;
1599
1600     update_RefPicList();
1601
1602     /* one frame, one slice */
1603     slice_param.macroblock_address = 0;
1604     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1605     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1606     if (current_frame_type == FRAME_IDR) {
1607         if (current_frame_encoding != 0)
1608             ++slice_param.idr_pic_id;
1609     } else if (current_frame_type == FRAME_P) {
1610         int refpiclist0_max = h264_maxref & 0xffff;
1611         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1612
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     } else if (current_frame_type == FRAME_B) {
1618         int refpiclist0_max = h264_maxref & 0xffff;
1619         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1620
1621         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1622         for (i = refpiclist0_max; i < 32; i++) {
1623             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1624             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1625         }
1626
1627         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1628         for (i = refpiclist1_max; i < 32; i++) {
1629             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1630             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1631         }
1632     }
1633
1634     slice_param.slice_alpha_c0_offset_div2 = 0;
1635     slice_param.slice_beta_offset_div2 = 0;
1636     slice_param.direct_spatial_mv_pred_flag = 1;
1637     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1638
1639
1640     if (h264_packedheader &&
1641         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1642         render_packedslice();
1643
1644     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1645                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1646     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1647
1648     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1649
1650     return 0;
1651 }
1652
1653
1654
1655 void H264Encoder::save_codeddata(storage_task task)
1656 {
1657     VACodedBufferSegment *buf_list = NULL;
1658     VAStatus va_status;
1659     unsigned int coded_size = 0;
1660
1661     string data;
1662
1663     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1664
1665     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1666     CHECK_VASTATUS(va_status, "vaMapBuffer");
1667     while (buf_list != NULL) {
1668         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1669         buf_list = (VACodedBufferSegment *) buf_list->next;
1670
1671         frame_size += coded_size;
1672     }
1673     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1674
1675     {
1676         // Add video.
1677         AVPacket pkt;
1678         memset(&pkt, 0, sizeof(pkt));
1679         pkt.buf = nullptr;
1680         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1681         pkt.size = data.size();
1682         pkt.stream_index = 0;
1683         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1684             pkt.flags = AV_PKT_FLAG_KEY;
1685         } else {
1686             pkt.flags = 0;
1687         }
1688         //pkt.duration = 1;
1689         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1690     }
1691     // Encode and add all audio frames up to and including the pts of this video frame.
1692     for ( ;; ) {
1693         int64_t audio_pts;
1694         std::vector<float> audio;
1695         {
1696              unique_lock<mutex> lock(frame_queue_mutex);
1697              frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || !pending_audio_frames.empty(); });
1698              if (copy_thread_should_quit) return;
1699              auto it = pending_audio_frames.begin();
1700              if (it->first > task.pts) break;
1701              audio_pts = it->first;
1702              audio = move(it->second);
1703              pending_audio_frames.erase(it);
1704         }
1705
1706         AVFrame *frame = avcodec_alloc_frame();
1707         frame->nb_samples = audio.size() / 2;
1708         frame->format = AV_SAMPLE_FMT_S32;
1709         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1710
1711         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1712         avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 0);
1713         for (int i = 0; i < frame->nb_samples * 2; ++i) {
1714             if (audio[i] >= 1.0f) {
1715                 int_samples[i] = 2147483647;
1716             } else if (audio[i] <= -1.0f) {
1717                 int_samples[i] = -2147483647;
1718             } else {
1719                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1720             }
1721         }
1722
1723         AVPacket pkt;
1724         av_init_packet(&pkt);
1725         pkt.data = nullptr;
1726         pkt.size = 0;
1727         int got_output;
1728         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1729         if (got_output) {
1730             pkt.stream_index = 1;
1731             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1732         }
1733         // TODO: Delayed frames.
1734         avcodec_free_frame(&frame);
1735         av_free_packet(&pkt);
1736         if (audio_pts == task.pts) break;
1737     }
1738
1739 #if 0
1740     printf("\r      "); /* return back to startpoint */
1741     switch (encode_order % 4) {
1742         case 0:
1743             printf("|");
1744             break;
1745         case 1:
1746             printf("/");
1747             break;
1748         case 2:
1749             printf("-");
1750             break;
1751         case 3:
1752             printf("\\");
1753             break;
1754     }
1755     printf("%08lld", encode_order);
1756     printf("(%06d bytes coded)", coded_size);
1757 #endif
1758 }
1759
1760
1761 // this is weird. but it seems to put a new frame onto the queue
1762 void H264Encoder::storage_task_enqueue(storage_task task)
1763 {
1764         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1765         storage_task_queue.push(move(task));
1766         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1767         storage_task_queue_changed.notify_all();
1768 }
1769
1770 void H264Encoder::storage_task_thread()
1771 {
1772         for ( ;; ) {
1773                 storage_task current;
1774                 {
1775                         // wait until there's an encoded frame
1776                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1777                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1778                         if (storage_thread_should_quit) return;
1779                         current = move(storage_task_queue.front());
1780                         storage_task_queue.pop();
1781                 }
1782
1783                 VAStatus va_status;
1784
1785                 // waits for data, then saves it to disk.
1786                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1787                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1788                 save_codeddata(move(current));
1789
1790                 {
1791                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1792                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1793                         storage_task_queue_changed.notify_all();
1794                 }
1795         }
1796 }
1797
1798 static int release_encode()
1799 {
1800     int i;
1801
1802     for (i = 0; i < SURFACE_NUM; i++) {
1803         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1804         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1805         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1806     }
1807
1808     vaDestroyContext(va_dpy, context_id);
1809     vaDestroyConfig(va_dpy, config_id);
1810
1811     return 0;
1812 }
1813
1814 static int deinit_va()
1815 {
1816     vaTerminate(va_dpy);
1817
1818     va_close_display(va_dpy);
1819
1820     return 0;
1821 }
1822
1823
1824 static int print_input()
1825 {
1826     printf("\n\nINPUT:Try to encode H264...\n");
1827     if (rc_mode != -1)
1828         printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1829     printf("INPUT: Resolution   : %dx%dframes\n", frame_width, frame_height);
1830     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
1831     printf("INPUT: Slieces      : %d\n", frame_slices);
1832     printf("INPUT: IntraPeriod  : %d\n", intra_period);
1833     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
1834     printf("INPUT: IpPeriod     : %d\n", ip_period);
1835     printf("INPUT: Initial QP   : %d\n", initial_qp);
1836     printf("INPUT: Min QP       : %d\n", minimal_qp);
1837     printf("INPUT: Coded Clip   : %s\n", coded_fn);
1838
1839     printf("\n\n"); /* return back to startpoint */
1840
1841     return 0;
1842 }
1843
1844 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
1845         : current_storage_frame(0), surface(surface), httpd(httpd)
1846 {
1847         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1848         context_audio = avcodec_alloc_context3(codec_audio);
1849         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1850         context_audio->sample_rate = OUTPUT_FREQUENCY;
1851         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1852         context_audio->channels = 2;
1853         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1854         context_audio->time_base = AVRational{1, TIMEBASE};
1855         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1856                 fprintf(stderr, "Could not open codec\n");
1857                 exit(1);
1858         }
1859
1860         frame_width = width;
1861         frame_height = height;
1862         frame_width_mbaligned = (frame_width + 15) & (~15);
1863         frame_height_mbaligned = (frame_height + 15) & (~15);
1864         frame_bitrate = 15000000;  // / 60;
1865         current_frame_encoding = 0;
1866
1867         print_input();
1868
1869         init_va();
1870         setup_encode();
1871
1872         // No frames are ready yet.
1873         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1874
1875         memset(&seq_param, 0, sizeof(seq_param));
1876         memset(&pic_param, 0, sizeof(pic_param));
1877         memset(&slice_param, 0, sizeof(slice_param));
1878
1879         storage_thread = std::thread(&H264Encoder::storage_task_thread, this);
1880
1881         copy_thread = std::thread([this]{
1882                 //SDL_GL_MakeCurrent(window, context);
1883                 QOpenGLContext *context = create_context(this->surface);
1884                 eglBindAPI(EGL_OPENGL_API);
1885                 if (!make_current(context, this->surface)) {
1886                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1887                                 eglGetError());
1888                         exit(1);
1889                 }
1890                 copy_thread_func();
1891         });
1892 }
1893
1894 H264Encoder::~H264Encoder()
1895 {
1896         {
1897                 unique_lock<mutex> lock(storage_task_queue_mutex);
1898                 storage_thread_should_quit = true;
1899                 storage_task_queue_changed.notify_all();
1900         }
1901         {
1902                 unique_lock<mutex> lock(frame_queue_mutex);
1903                 copy_thread_should_quit = true;
1904                 frame_queue_nonempty.notify_all();
1905         }
1906         storage_thread.join();
1907         copy_thread.join();
1908
1909         release_encode();
1910         deinit_va();
1911 }
1912
1913 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1914 {
1915         {
1916                 // Wait until this frame slot is done encoding.
1917                 std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1918                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1919                 if (storage_thread_should_quit) return false;
1920         }
1921
1922         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1923         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1924         *y_tex = surf->y_tex;
1925         *cbcr_tex = surf->cbcr_tex;
1926
1927         VASurfaceID surface = surf->src_surface;
1928         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1929         CHECK_VASTATUS(va_status, "vaDeriveImage");
1930
1931         VABufferInfo buf_info;
1932         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1933         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1934         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1935
1936         // Create Y image.
1937         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1938         EGLint y_attribs[] = {
1939                 EGL_WIDTH, frame_width,
1940                 EGL_HEIGHT, frame_height,
1941                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1942                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1943                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1944                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1945                 EGL_NONE
1946         };
1947
1948         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1949         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1950
1951         // Associate Y image to a texture.
1952         glBindTexture(GL_TEXTURE_2D, *y_tex);
1953         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1954
1955         // Create CbCr image.
1956         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1957         EGLint cbcr_attribs[] = {
1958                 EGL_WIDTH, frame_width,
1959                 EGL_HEIGHT, frame_height,
1960                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1961                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1962                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1963                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1964                 EGL_NONE
1965         };
1966
1967         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1968         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1969
1970         // Associate CbCr image to a texture.
1971         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1972         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1973
1974         return true;
1975 }
1976
1977 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
1978 {
1979         {
1980                 unique_lock<mutex> lock(frame_queue_mutex);
1981                 pending_audio_frames[pts] = move(audio);
1982         }
1983         frame_queue_nonempty.notify_all();
1984 }
1985
1986 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
1987 {
1988         {
1989                 unique_lock<mutex> lock(frame_queue_mutex);
1990                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1991                 ++current_storage_frame;
1992         }
1993         frame_queue_nonempty.notify_all();
1994 }
1995
1996 void H264Encoder::copy_thread_func()
1997 {
1998         int64_t last_dts = -1;
1999         for ( ;; ) {
2000                 PendingFrame frame;
2001                 int pts_lag;
2002                 encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
2003                                        &current_frame_display, &current_frame_type, &pts_lag);
2004                 if (current_frame_type == FRAME_IDR) {
2005                         numShortTerm = 0;
2006                         current_frame_num = 0;
2007                         current_IDR_display = current_frame_display;
2008                 }
2009
2010                 {
2011                         unique_lock<mutex> lock(frame_queue_mutex);
2012                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; });
2013                         if (copy_thread_should_quit) return;
2014                         frame = move(pending_video_frames[current_frame_display]);
2015                         pending_video_frames.erase(current_frame_display);
2016                 }
2017
2018                 // Wait for the GPU to be done with the frame.
2019                 glClientWaitSync(frame.fence.get(), 0, 0);
2020
2021                 // Release back any input frames we needed to render this frame.
2022                 frame.input_frames.clear();
2023
2024                 // Unmap the image.
2025                 GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
2026                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2027                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2028                 VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2029                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2030                 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2031                 CHECK_VASTATUS(va_status, "vaDestroyImage");
2032
2033                 VASurfaceID surface = surf->src_surface;
2034
2035                 // Schedule the frame for encoding.
2036                 va_status = vaBeginPicture(va_dpy, context_id, surface);
2037                 CHECK_VASTATUS(va_status, "vaBeginPicture");
2038
2039                 if (current_frame_type == FRAME_IDR) {
2040                         render_sequence();
2041                         render_picture();
2042                         if (h264_packedheader) {
2043                                 render_packedsequence();
2044                                 render_packedpicture();
2045                         }
2046                 } else {
2047                         //render_sequence();
2048                         render_picture();
2049                 }
2050                 render_slice();
2051
2052                 va_status = vaEndPicture(va_dpy, context_id);
2053                 CHECK_VASTATUS(va_status, "vaEndPicture");
2054
2055                 // Determine the pts and dts of this frame.
2056                 int64_t pts = frame.pts;
2057                 int64_t dts;
2058                 if (pts_lag == -1) {
2059                         assert(last_dts != -1);
2060                         dts = last_dts + (TIMEBASE / MAX_FPS);
2061                 } else {
2062                         dts = pts - pts_lag;
2063                 }
2064                 last_dts = dts;
2065
2066                 // so now the data is done encoding (well, async job kicked off)...
2067                 // we send that to the storage thread
2068                 storage_task tmp;
2069                 tmp.display_order = current_frame_display;
2070                 tmp.frame_type = current_frame_type;
2071                 tmp.pts = pts;
2072                 tmp.dts = dts;
2073                 storage_task_enqueue(move(tmp));
2074
2075                 update_ReferenceFrames();
2076                 ++current_frame_encoding;
2077         }
2078 }