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

   1 //#include "sysdeps.h"
   2 #include "h264encode.h"
   3
   4 #include <EGL/eglplatform.h>
   5 #include <X11/X.h>
   6 #include <X11/Xlib.h>
   7 #include <assert.h>
   8 #include <epoxy/egl.h>
   9 #include <libavcodec/avcodec.h>
  10 #include <libavutil/channel_layout.h>
  11 #include <libavutil/frame.h>
  12 #include <libavutil/rational.h>
  13 #include <libavutil/samplefmt.h>
  14 #include <libdrm/drm_fourcc.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <string.h>
  18 #include <va/va.h>
  19 #include <va/va_drmcommon.h>
  20 #include <va/va_enc_h264.h>
  21 #include <va/va_x11.h>
  22 #include <condition_variable>
  23 #include <cstdint>
  24 #include <memory>
  25 #include <mutex>
  26 #include <queue>
  27 #include <string>
  28 #include <thread>
  29 #include <utility>
  30
  31 #include "context.h"
  32 #include "defs.h"
  33 #include "httpd.h"
  34 #include "timebase.h"
  35
  36 class QOpenGLContext;
  37 class QSurface;
  38
  39 #define CHECK_VASTATUS(va_status, func)                                 \
  40     if (va_status != VA_STATUS_SUCCESS) {                               \
  41         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  42         exit(1);                                                        \
  43     }
  44
  45 //#include "loadsurface.h"
  46
  47 #define NAL_REF_IDC_NONE        0
  48 #define NAL_REF_IDC_LOW         1
  49 #define NAL_REF_IDC_MEDIUM      2
  50 #define NAL_REF_IDC_HIGH        3
  51
  52 #define NAL_NON_IDR             1
  53 #define NAL_IDR                 5
  54 #define NAL_SPS                 7
  55 #define NAL_PPS                 8
  56 #define NAL_SEI                 6
  57
  58 #define SLICE_TYPE_P            0
  59 #define SLICE_TYPE_B            1
  60 #define SLICE_TYPE_I            2
  61 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  62 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  63 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  64
  65
  66 #define ENTROPY_MODE_CAVLC      0
  67 #define ENTROPY_MODE_CABAC      1
  68
  69 #define PROFILE_IDC_BASELINE    66
  70 #define PROFILE_IDC_MAIN        77
  71 #define PROFILE_IDC_HIGH        100
  72
  73 #define BITSTREAM_ALLOCATE_STEPPING     4096
  74
  75 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  76 static  VADisplay va_dpy;
  77 static  VAProfile h264_profile = (VAProfile)~0;
  78 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  79 static  int config_attrib_num = 0, enc_packed_header_idx;
  80
  81 struct GLSurface {
  82         VASurfaceID src_surface, ref_surface;
  83         VABufferID coded_buf;
  84
  85         VAImage surface_image;
  86         GLuint y_tex, cbcr_tex;
  87         EGLImage y_egl_image, cbcr_egl_image;
  88 };
  89 GLSurface gl_surfaces[SURFACE_NUM];
  90
  91 static  VAConfigID config_id;
  92 static  VAContextID context_id;
  93 static  VAEncSequenceParameterBufferH264 seq_param;
  94 static  VAEncPictureParameterBufferH264 pic_param;
  95 static  VAEncSliceParameterBufferH264 slice_param;
  96 static  VAPictureH264 CurrentCurrPic;
  97 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
  98
  99 static  unsigned int MaxFrameNum = (2<<16);
 100 static  unsigned int MaxPicOrderCntLsb = (2<<8);
 101 static  unsigned int Log2MaxFrameNum = 16;
 102 static  unsigned int Log2MaxPicOrderCntLsb = 8;
 103
 104 static  unsigned int num_ref_frames = 2;
 105 static  unsigned int numShortTerm = 0;
 106 static  int constraint_set_flag = 0;
 107 static  int h264_packedheader = 0; /* support pack header? */
 108 static  int h264_maxref = (1<<16|1);
 109 static  int h264_entropy_mode = 1; /* cabac */
 110
 111 static  char *coded_fn = NULL;
 112
 113 static  int frame_width = 176;
 114 static  int frame_height = 144;
 115 static  int frame_width_mbaligned;
 116 static  int frame_height_mbaligned;
 117 static  int frame_rate = FPS;
 118 static  unsigned int frame_bitrate = 0;
 119 static  unsigned int frame_slices = 1;
 120 static  double frame_size = 0;
 121 static  int initial_qp = 15;
 122 //static  int initial_qp = 28;
 123 static  int minimal_qp = 0;
 124 static  int intra_period = 30;
 125 static  int intra_idr_period = FPS;
 126 static  int ip_period = 3;
 127 static  int rc_mode = -1;
 128 static  int rc_default_modes[] = {
 129     VA_RC_VBR,
 130     VA_RC_CQP,
 131     VA_RC_VBR_CONSTRAINED,
 132     VA_RC_CBR,
 133     VA_RC_VCM,
 134     VA_RC_NONE,
 135 };
 136 static  unsigned long long current_frame_encoding = 0;
 137 static  unsigned long long current_frame_display = 0;
 138 static  unsigned long long current_IDR_display = 0;
 139 static  unsigned int current_frame_num = 0;
 140 static  int current_frame_type;
 141
 142 static  int misc_priv_type = 0;
 143 static  int misc_priv_value = 0;
 144
 145 /* thread to save coded data */
 146 #define SRC_SURFACE_FREE        0
 147 #define SRC_SURFACE_IN_ENCODING 1
 148
 149 struct __bitstream {
 150     unsigned int *buffer;
 151     int bit_offset;
 152     int max_size_in_dword;
 153 };
 154 typedef struct __bitstream bitstream;
 155
 156 using namespace std;
 157
 158 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 159 // but if we don't delete it here, we get leaks. The GStreamer implementation
 160 // does the same.
 161 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 162 {
 163     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 164     CHECK_VASTATUS(va_status, "vaRenderPicture");
 165
 166     for (int i = 0; i < num_buffers; ++i) {
 167         va_status = vaDestroyBuffer(va_dpy, buffers[i]);
 168         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 169     }
 170 }
 171
 172 static unsigned int
 173 va_swap32(unsigned int val)
 174 {
 175     unsigned char *pval = (unsigned char *)&val;
 176
 177     return ((pval[0] << 24)     |
 178             (pval[1] << 16)     |
 179             (pval[2] << 8)      |
 180             (pval[3] << 0));
 181 }
 182
 183 static void
 184 bitstream_start(bitstream *bs)
 185 {
 186     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 187     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 188     bs->bit_offset = 0;
 189 }
 190
 191 static void
 192 bitstream_end(bitstream *bs)
 193 {
 194     int pos = (bs->bit_offset >> 5);
 195     int bit_offset = (bs->bit_offset & 0x1f);
 196     int bit_left = 32 - bit_offset;
 197
 198     if (bit_offset) {
 199         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 200     }
 201 }
 202
 203 static void
 204 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 205 {
 206     int pos = (bs->bit_offset >> 5);
 207     int bit_offset = (bs->bit_offset & 0x1f);
 208     int bit_left = 32 - bit_offset;
 209
 210     if (!size_in_bits)
 211         return;
 212
 213     bs->bit_offset += size_in_bits;
 214
 215     if (bit_left > size_in_bits) {
 216         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 217     } else {
 218         size_in_bits -= bit_left;
 219         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 220         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 221
 222         if (pos + 1 == bs->max_size_in_dword) {
 223             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 224             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 225         }
 226
 227         bs->buffer[pos + 1] = val;
 228     }
 229 }
 230
 231 static void
 232 bitstream_put_ue(bitstream *bs, unsigned int val)
 233 {
 234     int size_in_bits = 0;
 235     int tmp_val = ++val;
 236
 237     while (tmp_val) {
 238         tmp_val >>= 1;
 239         size_in_bits++;
 240     }
 241
 242     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 243     bitstream_put_ui(bs, val, size_in_bits);
 244 }
 245
 246 static void
 247 bitstream_put_se(bitstream *bs, int val)
 248 {
 249     unsigned int new_val;
 250
 251     if (val <= 0)
 252         new_val = -2 * val;
 253     else
 254         new_val = 2 * val - 1;
 255
 256     bitstream_put_ue(bs, new_val);
 257 }
 258
 259 static void
 260 bitstream_byte_aligning(bitstream *bs, int bit)
 261 {
 262     int bit_offset = (bs->bit_offset & 0x7);
 263     int bit_left = 8 - bit_offset;
 264     int new_val;
 265
 266     if (!bit_offset)
 267         return;
 268
 269     assert(bit == 0 || bit == 1);
 270
 271     if (bit)
 272         new_val = (1 << bit_left) - 1;
 273     else
 274         new_val = 0;
 275
 276     bitstream_put_ui(bs, new_val, bit_left);
 277 }
 278
 279 static void
 280 rbsp_trailing_bits(bitstream *bs)
 281 {
 282     bitstream_put_ui(bs, 1, 1);
 283     bitstream_byte_aligning(bs, 0);
 284 }
 285
 286 static void nal_start_code_prefix(bitstream *bs)
 287 {
 288     bitstream_put_ui(bs, 0x00000001, 32);
 289 }
 290
 291 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 292 {
 293     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 294     bitstream_put_ui(bs, nal_ref_idc, 2);
 295     bitstream_put_ui(bs, nal_unit_type, 5);
 296 }
 297
 298 static void sps_rbsp(bitstream *bs)
 299 {
 300     int profile_idc = PROFILE_IDC_BASELINE;
 301
 302     if (h264_profile  == VAProfileH264High)
 303         profile_idc = PROFILE_IDC_HIGH;
 304     else if (h264_profile  == VAProfileH264Main)
 305         profile_idc = PROFILE_IDC_MAIN;
 306
 307     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 308     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 309     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 310     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 311     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 312     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 313     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 314     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 315
 316     if ( profile_idc == PROFILE_IDC_HIGH) {
 317         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 318         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 319         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 320         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 321         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 322     }
 323
 324     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 325     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 326
 327     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 328         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 329     else {
 330         assert(0);
 331     }
 332
 333     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 334     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 335
 336     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 337     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 338     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 339
 340     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 341         assert(0);
 342     }
 343
 344     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 345     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 346
 347     if (seq_param.frame_cropping_flag) {
 348         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 349         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 350         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 351         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 352     }
 353
 354     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 355     if ( false ) {
 356         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 357     } else {
 358         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 359         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 360         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 361         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 362         {
 363             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 364             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 365             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 366             {
 367                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 368                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 369                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 370             }
 371         }
 372         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 373         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 374         {
 375             bitstream_put_ui(bs, 1, 32);  // FPS
 376             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 377             bitstream_put_ui(bs, 1, 1);
 378         }
 379         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 380         {
 381             // hrd_parameters
 382             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 383             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 384             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 385
 386             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 387             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 388             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 389
 390             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 391             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 392             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 393             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 394         }
 395         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 396         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 397
 398         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 399         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 400     }
 401
 402     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 403 }
 404
 405
 406 static void pps_rbsp(bitstream *bs)
 407 {
 408     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 409     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 410
 411     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 412
 413     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 414
 415     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 416
 417     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 418     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 419
 420     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 421     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 422
 423     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 424     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 425     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 426
 427     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 428     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 429     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 430
 431     /* more_rbsp_data */
 432     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 433     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 434     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 435
 436     rbsp_trailing_bits(bs);
 437 }
 438
 439 static void slice_header(bitstream *bs)
 440 {
 441     int first_mb_in_slice = slice_param.macroblock_address;
 442
 443     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 444     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 445     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 446     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 447
 448     /* frame_mbs_only_flag == 1 */
 449     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 450         /* FIXME: */
 451         assert(0);
 452     }
 453
 454     if (pic_param.pic_fields.bits.idr_pic_flag)
 455         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 456
 457     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 458         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 459         /* pic_order_present_flag == 0 */
 460     } else {
 461         /* FIXME: */
 462         assert(0);
 463     }
 464
 465     /* redundant_pic_cnt_present_flag == 0 */
 466     /* slice type */
 467     if (IS_P_SLICE(slice_param.slice_type)) {
 468         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 469
 470         if (slice_param.num_ref_idx_active_override_flag)
 471             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 472
 473         /* ref_pic_list_reordering */
 474         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 475     } else if (IS_B_SLICE(slice_param.slice_type)) {
 476         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 477
 478         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 479
 480         if (slice_param.num_ref_idx_active_override_flag) {
 481             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 482             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 483         }
 484
 485         /* ref_pic_list_reordering */
 486         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 487         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 488     }
 489
 490     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 491          IS_P_SLICE(slice_param.slice_type)) ||
 492         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 493          IS_B_SLICE(slice_param.slice_type))) {
 494         /* FIXME: fill weight/offset table */
 495         assert(0);
 496     }
 497
 498     /* dec_ref_pic_marking */
 499     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 500         unsigned char no_output_of_prior_pics_flag = 0;
 501         unsigned char long_term_reference_flag = 0;
 502         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 503
 504         if (pic_param.pic_fields.bits.idr_pic_flag) {
 505             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 506             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 507         } else {
 508             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 509         }
 510     }
 511
 512     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 513         !IS_I_SLICE(slice_param.slice_type))
 514         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 515
 516     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 517
 518     /* ignore for SP/SI */
 519
 520     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 521         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 522
 523         if (slice_param.disable_deblocking_filter_idc != 1) {
 524             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 525             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 526         }
 527     }
 528
 529     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 530         bitstream_byte_aligning(bs, 1);
 531     }
 532 }
 533
 534 static int
 535 build_packed_pic_buffer(unsigned char **header_buffer)
 536 {
 537     bitstream bs;
 538
 539     bitstream_start(&bs);
 540     nal_start_code_prefix(&bs);
 541     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 542     pps_rbsp(&bs);
 543     bitstream_end(&bs);
 544
 545     *header_buffer = (unsigned char *)bs.buffer;
 546     return bs.bit_offset;
 547 }
 548
 549 static int
 550 build_packed_seq_buffer(unsigned char **header_buffer)
 551 {
 552     bitstream bs;
 553
 554     bitstream_start(&bs);
 555     nal_start_code_prefix(&bs);
 556     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 557     sps_rbsp(&bs);
 558     bitstream_end(&bs);
 559
 560     *header_buffer = (unsigned char *)bs.buffer;
 561     return bs.bit_offset;
 562 }
 563
 564 static int build_packed_slice_buffer(unsigned char **header_buffer)
 565 {
 566     bitstream bs;
 567     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 568     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 569
 570     bitstream_start(&bs);
 571     nal_start_code_prefix(&bs);
 572
 573     if (IS_I_SLICE(slice_param.slice_type)) {
 574         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 575     } else if (IS_P_SLICE(slice_param.slice_type)) {
 576         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 577     } else {
 578         assert(IS_B_SLICE(slice_param.slice_type));
 579         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 580     }
 581
 582     slice_header(&bs);
 583     bitstream_end(&bs);
 584
 585     *header_buffer = (unsigned char *)bs.buffer;
 586     return bs.bit_offset;
 587 }
 588
 589
 590 /*
 591   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 592   1) period between Frame #X and Frame #N = #X - #N
 593   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 594   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 595   4) intra_period and intra_idr_period take precedence over ip_period
 596   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 597      of I/IDR frames, see bellow examples
 598   -------------------------------------------------------------------
 599   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 600   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 601   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 602   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 603   1            1                ignored    IDR IDR IDR IDR...
 604   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 605   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 606   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 607                                               (PBB)(IBB)(PBB)(IBB)...
 608   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 609                                            IDRPPPPP IPPPPP IPPPPP...
 610   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 611                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 612                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 613                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 614                                            {IDR(PBB)(PBB)}                     (6/6/3)
 615                                            {IDR(PBB)(PBB)}.
 616 */
 617
 618 // General pts/dts strategy:
 619 //
 620 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 621 // bit tricky. We assume first of all that the frame rate never goes _above_
 622 // <frame_rate>, which gives us a frame period N. The decoder can always decode
 623 // in at least this speed, as long at dts <= pts (the frame is not attempted
 624 // presented before it is decoded). Furthermore, we never have longer chains of
 625 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 626 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 627 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 628 // frame, with an order O <= C.)
 629 //
 630 // Many strategies are possible, but we establish these rules:
 631 //
 632 //  - Tip frames have dts = pts - (C-O)*N.
 633 //  - Non-tip frames have dts = dts_last + N.
 634 //
 635 // An example, with C=2 and N=10 and the data flow showed with arrows:
 636 //
 637 //        I  B  P  B  B  P
 638 //   pts: 30 40 50 60 70 80
 639 //        ↓  ↓     ↓
 640 //   dts: 10 30 20 60 50←40
 641 //         |  |  ↑        ↑
 642 //         `--|--'        |
 643 //             `----------'
 644 //
 645 // To show that this works fine also with irregular spacings, let's say that
 646 // the third frame is delayed a bit (something earlier was dropped). Now the
 647 // situation looks like this:
 648 //
 649 //        I  B  P  B  B   P
 650 //   pts: 30 40 80 90 100 110
 651 //        ↓  ↓     ↓
 652 //   dts: 10 30 20 90 50←40
 653 //         |  |  ↑        ↑
 654 //         `--|--'        |
 655 //             `----------'
 656 //
 657 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 658 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 659 //
 660 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 661 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 662 // a tip frame and should be given a dts based on the previous one.
 663 #define FRAME_P 0
 664 #define FRAME_B 1
 665 #define FRAME_I 2
 666 #define FRAME_IDR 7
 667 void encoding2display_order(
 668     unsigned long long encoding_order, int intra_period,
 669     int intra_idr_period, int ip_period,
 670     unsigned long long *displaying_order,
 671     int *frame_type, int *pts_lag)
 672 {
 673     int encoding_order_gop = 0;
 674
 675     *pts_lag = 0;
 676
 677     if (intra_period == 1) { /* all are I/IDR frames */
 678         *displaying_order = encoding_order;
 679         if (intra_idr_period == 0)
 680             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 681         else
 682             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 683         return;
 684     }
 685
 686     if (intra_period == 0)
 687         intra_idr_period = 0;
 688
 689     if (ip_period == 1) {
 690         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 691         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 692         *displaying_order = encoding_order;
 693
 694         if (encoding_order_gop == 0) { /* the first frame */
 695             *frame_type = FRAME_IDR;
 696         } else if (intra_period != 0 && /* have I frames */
 697                    encoding_order_gop >= 2 &&
 698                    (encoding_order_gop % intra_period == 0)) {
 699             *frame_type = FRAME_I;
 700         } else {
 701             *frame_type = FRAME_P;
 702         }
 703         return;
 704     }
 705
 706     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 707     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 708     *pts_lag = -1;  // Most frames are not tip frames.
 709
 710     if (encoding_order_gop == 0) { /* the first frame */
 711         *frame_type = FRAME_IDR;
 712         *displaying_order = encoding_order;
 713         // IDR frames are a special case; I honestly can't find the logic behind
 714         // why this is the right thing, but it seems to line up nicely in practice :-)
 715         *pts_lag = TIMEBASE / frame_rate;
 716     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 717         *frame_type = FRAME_B;
 718         *displaying_order = encoding_order - 1;
 719         if ((encoding_order_gop % ip_period) == 0) {
 720             *pts_lag = 0;  // Last B-frame.
 721         }
 722     } else if (intra_period != 0 && /* have I frames */
 723                encoding_order_gop >= 2 &&
 724                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 725         *frame_type = FRAME_I;
 726         *displaying_order = encoding_order + ip_period - 1;
 727     } else {
 728         *frame_type = FRAME_P;
 729         *displaying_order = encoding_order + ip_period - 1;
 730     }
 731 }
 732
 733
 734 static const char *rc_to_string(int rcmode)
 735 {
 736     switch (rc_mode) {
 737     case VA_RC_NONE:
 738         return "NONE";
 739     case VA_RC_CBR:
 740         return "CBR";
 741     case VA_RC_VBR:
 742         return "VBR";
 743     case VA_RC_VCM:
 744         return "VCM";
 745     case VA_RC_CQP:
 746         return "CQP";
 747     case VA_RC_VBR_CONSTRAINED:
 748         return "VBR_CONSTRAINED";
 749     default:
 750         return "Unknown";
 751     }
 752 }
 753
 754 #if 0
 755 static int process_cmdline(int argc, char *argv[])
 756 {
 757     char c;
 758     const struct option long_opts[] = {
 759         {"help", no_argument, NULL, 0 },
 760         {"bitrate", required_argument, NULL, 1 },
 761         {"minqp", required_argument, NULL, 2 },
 762         {"initialqp", required_argument, NULL, 3 },
 763         {"intra_period", required_argument, NULL, 4 },
 764         {"idr_period", required_argument, NULL, 5 },
 765         {"ip_period", required_argument, NULL, 6 },
 766         {"rcmode", required_argument, NULL, 7 },
 767         {"srcyuv", required_argument, NULL, 9 },
 768         {"recyuv", required_argument, NULL, 10 },
 769         {"fourcc", required_argument, NULL, 11 },
 770         {"syncmode", no_argument, NULL, 12 },
 771         {"enablePSNR", no_argument, NULL, 13 },
 772         {"prit", required_argument, NULL, 14 },
 773         {"priv", required_argument, NULL, 15 },
 774         {"framecount", required_argument, NULL, 16 },
 775         {"entropy", required_argument, NULL, 17 },
 776         {"profile", required_argument, NULL, 18 },
 777         {NULL, no_argument, NULL, 0 }};
 778     int long_index;
 779
 780     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 781         switch (c) {
 782         case 'w':
 783             frame_width = atoi(optarg);
 784             break;
 785         case 'h':
 786             frame_height = atoi(optarg);
 787             break;
 788         case 'n':
 789         case 'f':
 790             frame_rate = atoi(optarg);
 791             break;
 792         case 'o':
 793             coded_fn = strdup(optarg);
 794             break;
 795         case 0:
 796             print_help();
 797             exit(0);
 798         case 1:
 799             frame_bitrate = atoi(optarg);
 800             break;
 801         case 2:
 802             minimal_qp = atoi(optarg);
 803             break;
 804         case 3:
 805             initial_qp = atoi(optarg);
 806             break;
 807         case 4:
 808             intra_period = atoi(optarg);
 809             break;
 810         case 5:
 811             intra_idr_period = atoi(optarg);
 812             break;
 813         case 6:
 814             ip_period = atoi(optarg);
 815             break;
 816         case 7:
 817             rc_mode = string_to_rc(optarg);
 818             if (rc_mode < 0) {
 819                 print_help();
 820                 exit(1);
 821             }
 822             break;
 823         case 9:
 824             srcyuv_fn = strdup(optarg);
 825             break;
 826         case 11:
 827             srcyuv_fourcc = string_to_fourcc(optarg);
 828             if (srcyuv_fourcc <= 0) {
 829                 print_help();
 830                 exit(1);
 831             }
 832             break;
 833         case 13:
 834             calc_psnr = 1;
 835             break;
 836         case 14:
 837             misc_priv_type = strtol(optarg, NULL, 0);
 838             break;
 839         case 15:
 840             misc_priv_value = strtol(optarg, NULL, 0);
 841             break;
 842         case 17:
 843             h264_entropy_mode = atoi(optarg) ? 1: 0;
 844             break;
 845         case 18:
 846             if (strncmp(optarg, "BP", 2) == 0)
 847                 h264_profile = VAProfileH264Baseline;
 848             else if (strncmp(optarg, "MP", 2) == 0)
 849                 h264_profile = VAProfileH264Main;
 850             else if (strncmp(optarg, "HP", 2) == 0)
 851                 h264_profile = VAProfileH264High;
 852             else
 853                 h264_profile = (VAProfile)0;
 854             break;
 855         case ':':
 856         case '?':
 857             print_help();
 858             exit(0);
 859         }
 860     }
 861
 862     if (ip_period < 1) {
 863         printf(" ip_period must be greater than 0\n");
 864         exit(0);
 865     }
 866     if (intra_period != 1 && intra_period % ip_period != 0) {
 867         printf(" intra_period must be a multiplier of ip_period\n");
 868         exit(0);
 869     }
 870     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 871         printf(" intra_idr_period must be a multiplier of intra_period\n");
 872         exit(0);
 873     }
 874
 875     if (frame_bitrate == 0)
 876         frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
 877
 878     if (coded_fn == NULL) {
 879         struct stat buf;
 880         if (stat("/tmp", &buf) == 0)
 881             coded_fn = strdup("/tmp/test.264");
 882         else if (stat("/sdcard", &buf) == 0)
 883             coded_fn = strdup("/sdcard/test.264");
 884         else
 885             coded_fn = strdup("./test.264");
 886     }
 887
 888
 889     frame_width_mbaligned = (frame_width + 15) & (~15);
 890     frame_height_mbaligned = (frame_height + 15) & (~15);
 891     if (frame_width != frame_width_mbaligned ||
 892         frame_height != frame_height_mbaligned) {
 893         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 894                frame_width, frame_height,
 895                frame_width_mbaligned, frame_height_mbaligned
 896                );
 897     }
 898
 899     return 0;
 900 }
 901 #endif
 902
 903 static Display *x11_display;
 904 static Window   x11_window;
 905
 906 VADisplay
 907 va_open_display(void)
 908 {
 909     x11_display = XOpenDisplay(NULL);
 910     if (!x11_display) {
 911         fprintf(stderr, "error: can't connect to X server!\n");
 912         return NULL;
 913     }
 914     return vaGetDisplay(x11_display);
 915 }
 916
 917 void
 918 va_close_display(VADisplay va_dpy)
 919 {
 920     if (!x11_display)
 921         return;
 922
 923     if (x11_window) {
 924         XUnmapWindow(x11_display, x11_window);
 925         XDestroyWindow(x11_display, x11_window);
 926         x11_window = None;
 927     }
 928     XCloseDisplay(x11_display);
 929     x11_display = NULL;
 930 }
 931
 932 static int init_va(void)
 933 {
 934     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 935     VAEntrypoint *entrypoints;
 936     int num_entrypoints, slice_entrypoint;
 937     int support_encode = 0;
 938     int major_ver, minor_ver;
 939     VAStatus va_status;
 940     unsigned int i;
 941
 942     va_dpy = va_open_display();
 943     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 944     CHECK_VASTATUS(va_status, "vaInitialize");
 945
 946     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 947     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 948     if (!entrypoints) {
 949         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 950         exit(1);
 951     }
 952
 953     /* use the highest profile */
 954     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 955         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 956             continue;
 957
 958         h264_profile = profile_list[i];
 959         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 960         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 961             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 962                 support_encode = 1;
 963                 break;
 964             }
 965         }
 966         if (support_encode == 1)
 967             break;
 968     }
 969
 970     if (support_encode == 0) {
 971         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 972         exit(1);
 973     } else {
 974         switch (h264_profile) {
 975             case VAProfileH264Baseline:
 976                 printf("Use profile VAProfileH264Baseline\n");
 977                 ip_period = 1;
 978                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 979                 h264_entropy_mode = 0;
 980                 break;
 981             case VAProfileH264ConstrainedBaseline:
 982                 printf("Use profile VAProfileH264ConstrainedBaseline\n");
 983                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 984                 ip_period = 1;
 985                 break;
 986
 987             case VAProfileH264Main:
 988                 printf("Use profile VAProfileH264Main\n");
 989                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 990                 break;
 991
 992             case VAProfileH264High:
 993                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 994                 printf("Use profile VAProfileH264High\n");
 995                 break;
 996             default:
 997                 printf("unknow profile. Set to Baseline");
 998                 h264_profile = VAProfileH264Baseline;
 999                 ip_period = 1;
1000                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1001                 break;
1002         }
1003     }
1004
1005     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1006
1007     /* find out the format for the render target, and rate control mode */
1008     for (i = 0; i < VAConfigAttribTypeMax; i++)
1009         attrib[i].type = (VAConfigAttribType)i;
1010
1011     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1012                                       &attrib[0], VAConfigAttribTypeMax);
1013     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1014     /* check the interested configattrib */
1015     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1016         printf("Not find desired YUV420 RT format\n");
1017         exit(1);
1018     } else {
1019         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1020         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1021         config_attrib_num++;
1022     }
1023
1024     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1025         int tmp = attrib[VAConfigAttribRateControl].value;
1026
1027         printf("Support rate control mode (0x%x):", tmp);
1028
1029         if (tmp & VA_RC_NONE)
1030             printf("NONE ");
1031         if (tmp & VA_RC_CBR)
1032             printf("CBR ");
1033         if (tmp & VA_RC_VBR)
1034             printf("VBR ");
1035         if (tmp & VA_RC_VCM)
1036             printf("VCM ");
1037         if (tmp & VA_RC_CQP)
1038             printf("CQP ");
1039         if (tmp & VA_RC_VBR_CONSTRAINED)
1040             printf("VBR_CONSTRAINED ");
1041
1042         printf("\n");
1043
1044         if (rc_mode == -1 || !(rc_mode & tmp))  {
1045             if (rc_mode != -1) {
1046                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1047             }
1048
1049             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1050                 if (rc_default_modes[i] & tmp) {
1051                     rc_mode = rc_default_modes[i];
1052                     break;
1053                 }
1054             }
1055
1056             printf("RateControl mode: %s\n", rc_to_string(rc_mode));
1057         }
1058
1059         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1060         config_attrib[config_attrib_num].value = rc_mode;
1061         config_attrib_num++;
1062     }
1063
1064
1065     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1066         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1067
1068         printf("Support VAConfigAttribEncPackedHeaders\n");
1069
1070         h264_packedheader = 1;
1071         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1072         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1073
1074         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1075             printf("Support packed sequence headers\n");
1076             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1077         }
1078
1079         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1080             printf("Support packed picture headers\n");
1081             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1082         }
1083
1084         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1085             printf("Support packed slice headers\n");
1086             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1087         }
1088
1089         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1090             printf("Support packed misc headers\n");
1091             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1092         }
1093
1094         enc_packed_header_idx = config_attrib_num;
1095         config_attrib_num++;
1096     }
1097
1098     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1099         int tmp = attrib[VAConfigAttribEncInterlaced].value;
1100
1101         printf("Support VAConfigAttribEncInterlaced\n");
1102
1103         if (tmp & VA_ENC_INTERLACED_FRAME)
1104             printf("support VA_ENC_INTERLACED_FRAME\n");
1105         if (tmp & VA_ENC_INTERLACED_FIELD)
1106             printf("Support VA_ENC_INTERLACED_FIELD\n");
1107         if (tmp & VA_ENC_INTERLACED_MBAFF)
1108             printf("Support VA_ENC_INTERLACED_MBAFF\n");
1109         if (tmp & VA_ENC_INTERLACED_PAFF)
1110             printf("Support VA_ENC_INTERLACED_PAFF\n");
1111
1112         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1113         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1114         config_attrib_num++;
1115     }
1116
1117     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1118         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1119
1120         printf("Support %d RefPicList0 and %d RefPicList1\n",
1121                h264_maxref & 0xffff, (h264_maxref >> 16) & 0xffff );
1122     }
1123
1124     if (attrib[VAConfigAttribEncMaxSlices].value != VA_ATTRIB_NOT_SUPPORTED)
1125         printf("Support %d slices\n", attrib[VAConfigAttribEncMaxSlices].value);
1126
1127     if (attrib[VAConfigAttribEncSliceStructure].value != VA_ATTRIB_NOT_SUPPORTED) {
1128         int tmp = attrib[VAConfigAttribEncSliceStructure].value;
1129
1130         printf("Support VAConfigAttribEncSliceStructure\n");
1131
1132         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS)
1133             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_ROWS\n");
1134         if (tmp & VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS)
1135             printf("Support VA_ENC_SLICE_STRUCTURE_POWER_OF_TWO_ROWS\n");
1136         if (tmp & VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS)
1137             printf("Support VA_ENC_SLICE_STRUCTURE_ARBITRARY_MACROBLOCKS\n");
1138     }
1139     if (attrib[VAConfigAttribEncMacroblockInfo].value != VA_ATTRIB_NOT_SUPPORTED) {
1140         printf("Support VAConfigAttribEncMacroblockInfo\n");
1141     }
1142
1143     free(entrypoints);
1144     return 0;
1145 }
1146
1147 static int setup_encode()
1148 {
1149     VAStatus va_status;
1150     VASurfaceID *tmp_surfaceid;
1151     int codedbuf_size, i;
1152     static VASurfaceID src_surface[SURFACE_NUM];
1153     static VASurfaceID ref_surface[SURFACE_NUM];
1154
1155     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1156             &config_attrib[0], config_attrib_num, &config_id);
1157     CHECK_VASTATUS(va_status, "vaCreateConfig");
1158
1159     /* create source surfaces */
1160     va_status = vaCreateSurfaces(va_dpy,
1161                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1162                                  &src_surface[0], SURFACE_NUM,
1163                                  NULL, 0);
1164     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1165
1166     /* create reference surfaces */
1167     va_status = vaCreateSurfaces(va_dpy,
1168                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1169                                  &ref_surface[0], SURFACE_NUM,
1170                                  NULL, 0);
1171     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1172
1173     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1174     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1175     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1176
1177     /* Create a context for this encode pipe */
1178     va_status = vaCreateContext(va_dpy, config_id,
1179                                 frame_width_mbaligned, frame_height_mbaligned,
1180                                 VA_PROGRESSIVE,
1181                                 tmp_surfaceid, 2 * SURFACE_NUM,
1182                                 &context_id);
1183     CHECK_VASTATUS(va_status, "vaCreateContext");
1184     free(tmp_surfaceid);
1185
1186     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1187
1188     for (i = 0; i < SURFACE_NUM; i++) {
1189         /* create coded buffer once for all
1190          * other VA buffers which won't be used again after vaRenderPicture.
1191          * so APP can always vaCreateBuffer for every frame
1192          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1193          * so VA won't maintain the coded buffer
1194          */
1195         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1196                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1197         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1198     }
1199
1200     /* create OpenGL objects */
1201     //glGenFramebuffers(SURFACE_NUM, fbos);
1202
1203     for (i = 0; i < SURFACE_NUM; i++) {
1204         glGenTextures(1, &gl_surfaces[i].y_tex);
1205         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1206     }
1207
1208     for (i = 0; i < SURFACE_NUM; i++) {
1209         gl_surfaces[i].src_surface = src_surface[i];
1210         gl_surfaces[i].ref_surface = ref_surface[i];
1211     }
1212
1213     return 0;
1214 }
1215
1216
1217
1218 #define partition(ref, field, key, ascending)   \
1219     while (i <= j) {                            \
1220         if (ascending) {                        \
1221             while (ref[i].field < key)          \
1222                 i++;                            \
1223             while (ref[j].field > key)          \
1224                 j--;                            \
1225         } else {                                \
1226             while (ref[i].field > key)          \
1227                 i++;                            \
1228             while (ref[j].field < key)          \
1229                 j--;                            \
1230         }                                       \
1231         if (i <= j) {                           \
1232             tmp = ref[i];                       \
1233             ref[i] = ref[j];                    \
1234             ref[j] = tmp;                       \
1235             i++;                                \
1236             j--;                                \
1237         }                                       \
1238     }                                           \
1239
1240 static void sort_one(VAPictureH264 ref[], int left, int right,
1241                      int ascending, int frame_idx)
1242 {
1243     int i = left, j = right;
1244     unsigned int key;
1245     VAPictureH264 tmp;
1246
1247     if (frame_idx) {
1248         key = ref[(left + right) / 2].frame_idx;
1249         partition(ref, frame_idx, key, ascending);
1250     } else {
1251         key = ref[(left + right) / 2].TopFieldOrderCnt;
1252         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1253     }
1254
1255     /* recursion */
1256     if (left < j)
1257         sort_one(ref, left, j, ascending, frame_idx);
1258
1259     if (i < right)
1260         sort_one(ref, i, right, ascending, frame_idx);
1261 }
1262
1263 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1264                      int partition_ascending, int list0_ascending, int list1_ascending)
1265 {
1266     int i = left, j = right;
1267     VAPictureH264 tmp;
1268
1269     if (frame_idx) {
1270         partition(ref, frame_idx, key, partition_ascending);
1271     } else {
1272         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1273     }
1274
1275
1276     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1277     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1278 }
1279
1280 static int update_ReferenceFrames(void)
1281 {
1282     int i;
1283
1284     if (current_frame_type == FRAME_B)
1285         return 0;
1286
1287     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1288     numShortTerm++;
1289     if (numShortTerm > num_ref_frames)
1290         numShortTerm = num_ref_frames;
1291     for (i=numShortTerm-1; i>0; i--)
1292         ReferenceFrames[i] = ReferenceFrames[i-1];
1293     ReferenceFrames[0] = CurrentCurrPic;
1294
1295     if (current_frame_type != FRAME_B)
1296         current_frame_num++;
1297     if (current_frame_num > MaxFrameNum)
1298         current_frame_num = 0;
1299
1300     return 0;
1301 }
1302
1303
1304 static int update_RefPicList(void)
1305 {
1306     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1307
1308     if (current_frame_type == FRAME_P) {
1309         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1310         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1311     }
1312
1313     if (current_frame_type == FRAME_B) {
1314         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1315         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1316                  1, 0, 1);
1317
1318         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1319         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1320                  0, 1, 0);
1321     }
1322
1323     return 0;
1324 }
1325
1326
1327 static int render_sequence(void)
1328 {
1329     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1330     VAStatus va_status;
1331     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1332     VAEncMiscParameterRateControl *misc_rate_ctrl;
1333
1334     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1335     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1336     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1337     seq_param.bits_per_second = frame_bitrate;
1338
1339     seq_param.intra_period = intra_period;
1340     seq_param.intra_idr_period = intra_idr_period;
1341     seq_param.ip_period = ip_period;
1342
1343     seq_param.max_num_ref_frames = num_ref_frames;
1344     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1345     seq_param.time_scale = TIMEBASE * 2;
1346     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1347     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1348     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1349     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1350     seq_param.seq_fields.bits.chroma_format_idc = 1;
1351     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1352
1353     if (frame_width != frame_width_mbaligned ||
1354         frame_height != frame_height_mbaligned) {
1355         seq_param.frame_cropping_flag = 1;
1356         seq_param.frame_crop_left_offset = 0;
1357         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1358         seq_param.frame_crop_top_offset = 0;
1359         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1360     }
1361
1362     va_status = vaCreateBuffer(va_dpy, context_id,
1363                                VAEncSequenceParameterBufferType,
1364                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1365     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1366
1367     va_status = vaCreateBuffer(va_dpy, context_id,
1368                                VAEncMiscParameterBufferType,
1369                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1370                                1, NULL, &rc_param_buf);
1371     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1372
1373     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1374     misc_param->type = VAEncMiscParameterTypeRateControl;
1375     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1376     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1377     misc_rate_ctrl->bits_per_second = frame_bitrate;
1378     misc_rate_ctrl->target_percentage = 66;
1379     misc_rate_ctrl->window_size = 1000;
1380     misc_rate_ctrl->initial_qp = initial_qp;
1381     misc_rate_ctrl->min_qp = minimal_qp;
1382     misc_rate_ctrl->basic_unit_size = 0;
1383     vaUnmapBuffer(va_dpy, rc_param_buf);
1384
1385     render_id[0] = seq_param_buf;
1386     render_id[1] = rc_param_buf;
1387
1388     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1389
1390     if (misc_priv_type != 0) {
1391         va_status = vaCreateBuffer(va_dpy, context_id,
1392                                    VAEncMiscParameterBufferType,
1393                                    sizeof(VAEncMiscParameterBuffer),
1394                                    1, NULL, &misc_param_tmpbuf);
1395         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1396         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1397         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1398         misc_param_tmp->data[0] = misc_priv_value;
1399         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1400
1401         render_picture_and_delete(va_dpy, context_id, &misc_param_tmpbuf, 1);
1402     }
1403
1404     return 0;
1405 }
1406
1407 static int calc_poc(int pic_order_cnt_lsb)
1408 {
1409     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1410     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1411     int PicOrderCntMsb, TopFieldOrderCnt;
1412
1413     if (current_frame_type == FRAME_IDR)
1414         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1415     else {
1416         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1417         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1418     }
1419
1420     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1421         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1422         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1423     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1424              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1425         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1426     else
1427         PicOrderCntMsb = prevPicOrderCntMsb;
1428
1429     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1430
1431     if (current_frame_type != FRAME_B) {
1432         PicOrderCntMsb_ref = PicOrderCntMsb;
1433         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1434     }
1435
1436     return TopFieldOrderCnt;
1437 }
1438
1439 static int render_picture(void)
1440 {
1441     VABufferID pic_param_buf;
1442     VAStatus va_status;
1443     int i = 0;
1444
1445     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1446     pic_param.CurrPic.frame_idx = current_frame_num;
1447     pic_param.CurrPic.flags = 0;
1448     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1449     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1450     CurrentCurrPic = pic_param.CurrPic;
1451
1452     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1453         update_RefPicList(); /* calc RefPicList */
1454         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1455         if (current_frame_type == FRAME_P) {
1456             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1457         } else if (current_frame_type == FRAME_B) {
1458             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1459             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1460         }
1461     } else {
1462         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1463         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1464             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1465             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1466         }
1467     }
1468
1469     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1470     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1471     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1472     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1473     pic_param.frame_num = current_frame_num;
1474     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1475     pic_param.last_picture = false;  // FIXME
1476     pic_param.pic_init_qp = initial_qp;
1477
1478     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1479                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1480     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1481
1482     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1483
1484     return 0;
1485 }
1486
1487 static int render_packedsequence(void)
1488 {
1489     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1490     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1491     unsigned int length_in_bits;
1492     unsigned char *packedseq_buffer = NULL;
1493     VAStatus va_status;
1494
1495     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1496
1497     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1498
1499     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1500     packedheader_param_buffer.has_emulation_bytes = 0;
1501     va_status = vaCreateBuffer(va_dpy,
1502                                context_id,
1503                                VAEncPackedHeaderParameterBufferType,
1504                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1505                                &packedseq_para_bufid);
1506     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1507
1508     va_status = vaCreateBuffer(va_dpy,
1509                                context_id,
1510                                VAEncPackedHeaderDataBufferType,
1511                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1512                                &packedseq_data_bufid);
1513     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1514
1515     render_id[0] = packedseq_para_bufid;
1516     render_id[1] = packedseq_data_bufid;
1517     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1518
1519     free(packedseq_buffer);
1520
1521     return 0;
1522 }
1523
1524
1525 static int render_packedpicture(void)
1526 {
1527     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1528     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1529     unsigned int length_in_bits;
1530     unsigned char *packedpic_buffer = NULL;
1531     VAStatus va_status;
1532
1533     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1534     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1535     packedheader_param_buffer.bit_length = length_in_bits;
1536     packedheader_param_buffer.has_emulation_bytes = 0;
1537
1538     va_status = vaCreateBuffer(va_dpy,
1539                                context_id,
1540                                VAEncPackedHeaderParameterBufferType,
1541                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1542                                &packedpic_para_bufid);
1543     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1544
1545     va_status = vaCreateBuffer(va_dpy,
1546                                context_id,
1547                                VAEncPackedHeaderDataBufferType,
1548                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1549                                &packedpic_data_bufid);
1550     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1551
1552     render_id[0] = packedpic_para_bufid;
1553     render_id[1] = packedpic_data_bufid;
1554     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1555
1556     free(packedpic_buffer);
1557
1558     return 0;
1559 }
1560
1561 static void render_packedslice()
1562 {
1563     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1564     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1565     unsigned int length_in_bits;
1566     unsigned char *packedslice_buffer = NULL;
1567     VAStatus va_status;
1568
1569     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1570     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1571     packedheader_param_buffer.bit_length = length_in_bits;
1572     packedheader_param_buffer.has_emulation_bytes = 0;
1573
1574     va_status = vaCreateBuffer(va_dpy,
1575                                context_id,
1576                                VAEncPackedHeaderParameterBufferType,
1577                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1578                                &packedslice_para_bufid);
1579     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1580
1581     va_status = vaCreateBuffer(va_dpy,
1582                                context_id,
1583                                VAEncPackedHeaderDataBufferType,
1584                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1585                                &packedslice_data_bufid);
1586     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1587
1588     render_id[0] = packedslice_para_bufid;
1589     render_id[1] = packedslice_data_bufid;
1590     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1591
1592     free(packedslice_buffer);
1593 }
1594
1595 static int render_slice(void)
1596 {
1597     VABufferID slice_param_buf;
1598     VAStatus va_status;
1599     int i;
1600
1601     update_RefPicList();
1602
1603     /* one frame, one slice */
1604     slice_param.macroblock_address = 0;
1605     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1606     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1607     if (current_frame_type == FRAME_IDR) {
1608         if (current_frame_encoding != 0)
1609             ++slice_param.idr_pic_id;
1610     } else if (current_frame_type == FRAME_P) {
1611         int refpiclist0_max = h264_maxref & 0xffff;
1612         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1613
1614         for (i = refpiclist0_max; i < 32; i++) {
1615             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1616             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1617         }
1618     } else if (current_frame_type == FRAME_B) {
1619         int refpiclist0_max = h264_maxref & 0xffff;
1620         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1621
1622         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1623         for (i = refpiclist0_max; i < 32; i++) {
1624             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1625             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1626         }
1627
1628         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1629         for (i = refpiclist1_max; i < 32; i++) {
1630             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1631             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1632         }
1633     }
1634
1635     slice_param.slice_alpha_c0_offset_div2 = 0;
1636     slice_param.slice_beta_offset_div2 = 0;
1637     slice_param.direct_spatial_mv_pred_flag = 1;
1638     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1639
1640
1641     if (h264_packedheader &&
1642         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1643         render_packedslice();
1644
1645     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1646                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1647     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1648
1649     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1650
1651     return 0;
1652 }
1653
1654
1655
1656 int H264Encoder::save_codeddata(storage_task task)
1657 {
1658     VACodedBufferSegment *buf_list = NULL;
1659     VAStatus va_status;
1660     unsigned int coded_size = 0;
1661
1662     string data;
1663
1664     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / frame_rate);  // So we never get negative dts.
1665
1666     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1667     CHECK_VASTATUS(va_status, "vaMapBuffer");
1668     while (buf_list != NULL) {
1669         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1670         buf_list = (VACodedBufferSegment *) buf_list->next;
1671
1672         frame_size += coded_size;
1673     }
1674     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1675
1676     {
1677         // Add video.
1678         AVPacket pkt;
1679         memset(&pkt, 0, sizeof(pkt));
1680         pkt.buf = nullptr;
1681         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1682         pkt.size = data.size();
1683         pkt.stream_index = 0;
1684         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1685             pkt.flags = AV_PKT_FLAG_KEY;
1686         } else {
1687             pkt.flags = 0;
1688         }
1689         //pkt.duration = 1;
1690         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1691     }
1692     // Encode and add all audio frames up to and including the pts of this video frame.
1693     // (They can never be queued to us after the video frame they belong to, only before.)
1694     for ( ;; ) {
1695         int64_t audio_pts;
1696         std::vector<float> audio;
1697         {
1698              unique_lock<mutex> lock(frame_queue_mutex);
1699              if (pending_audio_frames.empty()) break;
1700              auto it = pending_audio_frames.begin();
1701              if (it->first > task.pts) break;
1702              audio_pts = it->first;
1703              audio = move(it->second);
1704              pending_audio_frames.erase(it);
1705         }
1706
1707         AVFrame *frame = avcodec_alloc_frame();
1708         frame->nb_samples = audio.size() / 2;
1709         frame->format = AV_SAMPLE_FMT_S32;
1710         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1711
1712         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1713         avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 0);
1714         for (int i = 0; i < frame->nb_samples * 2; ++i) {
1715             if (audio[i] >= 1.0f) {
1716                 int_samples[i] = 2147483647;
1717             } else if (audio[i] <= -1.0f) {
1718                 int_samples[i] = -2147483647;
1719             } else {
1720                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1721             }
1722         }
1723
1724         AVPacket pkt;
1725         av_init_packet(&pkt);
1726         pkt.data = nullptr;
1727         pkt.size = 0;
1728         int got_output;
1729         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1730         if (got_output) {
1731             pkt.stream_index = 1;
1732             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1733         }
1734         // TODO: Delayed frames.
1735         avcodec_free_frame(&frame);
1736         av_free_packet(&pkt);
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     return 0;
1760 }
1761
1762
1763 // this is weird. but it seems to put a new frame onto the queue
1764 void H264Encoder::storage_task_enqueue(storage_task task)
1765 {
1766         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1767         storage_task_queue.push(move(task));
1768         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1769         storage_task_queue_changed.notify_all();
1770 }
1771
1772 void H264Encoder::storage_task_thread()
1773 {
1774         for ( ;; ) {
1775                 storage_task current;
1776                 {
1777                         // wait until there's an encoded frame
1778                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1779                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1780                         if (storage_thread_should_quit) return;
1781                         current = move(storage_task_queue.front());
1782                         storage_task_queue.pop();
1783                 }
1784
1785                 VAStatus va_status;
1786
1787                 // waits for data, then saves it to disk.
1788                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1789                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1790                 save_codeddata(move(current));
1791
1792                 {
1793                         std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1794                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1795                         storage_task_queue_changed.notify_all();
1796                 }
1797         }
1798 }
1799
1800 static int release_encode()
1801 {
1802     int i;
1803
1804     for (i = 0; i < SURFACE_NUM; i++) {
1805         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1806         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1807         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1808     }
1809
1810     vaDestroyContext(va_dpy, context_id);
1811     vaDestroyConfig(va_dpy, config_id);
1812
1813     return 0;
1814 }
1815
1816 static int deinit_va()
1817 {
1818     vaTerminate(va_dpy);
1819
1820     va_close_display(va_dpy);
1821
1822     return 0;
1823 }
1824
1825
1826 static int print_input()
1827 {
1828     printf("\n\nINPUT:Try to encode H264...\n");
1829     if (rc_mode != -1)
1830         printf("INPUT: RateControl  : %s\n", rc_to_string(rc_mode));
1831     printf("INPUT: Resolution   : %dx%dframes\n", frame_width, frame_height);
1832     printf("INPUT: FrameRate    : %d\n", frame_rate);
1833     printf("INPUT: Bitrate      : %d\n", frame_bitrate);
1834     printf("INPUT: Slieces      : %d\n", frame_slices);
1835     printf("INPUT: IntraPeriod  : %d\n", intra_period);
1836     printf("INPUT: IDRPeriod    : %d\n", intra_idr_period);
1837     printf("INPUT: IpPeriod     : %d\n", ip_period);
1838     printf("INPUT: Initial QP   : %d\n", initial_qp);
1839     printf("INPUT: Min QP       : %d\n", minimal_qp);
1840     printf("INPUT: Coded Clip   : %s\n", coded_fn);
1841
1842     printf("\n\n"); /* return back to startpoint */
1843
1844     return 0;
1845 }
1846
1847 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
1848         : current_storage_frame(0), surface(surface), httpd(httpd)
1849 {
1850         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1851         context_audio = avcodec_alloc_context3(codec_audio);
1852         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1853         context_audio->sample_rate = OUTPUT_FREQUENCY;
1854         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1855         context_audio->channels = 2;
1856         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1857         context_audio->time_base = AVRational{1, TIMEBASE};
1858         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1859                 fprintf(stderr, "Could not open codec\n");
1860                 exit(1);
1861         }
1862
1863         frame_width = width;
1864         frame_height = height;
1865         frame_width_mbaligned = (frame_width + 15) & (~15);
1866         frame_height_mbaligned = (frame_height + 15) & (~15);
1867         frame_bitrate = 15000000;  // / 60;
1868         current_frame_encoding = 0;
1869
1870         print_input();
1871
1872         init_va();
1873         setup_encode();
1874
1875         // No frames are ready yet.
1876         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1877
1878         memset(&seq_param, 0, sizeof(seq_param));
1879         memset(&pic_param, 0, sizeof(pic_param));
1880         memset(&slice_param, 0, sizeof(slice_param));
1881
1882         storage_thread = std::thread(&H264Encoder::storage_task_thread, this);
1883
1884         copy_thread = std::thread([this]{
1885                 //SDL_GL_MakeCurrent(window, context);
1886                 QOpenGLContext *context = create_context(this->surface);
1887                 eglBindAPI(EGL_OPENGL_API);
1888                 if (!make_current(context, this->surface)) {
1889                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1890                                 eglGetError());
1891                         exit(1);
1892                 }
1893                 copy_thread_func();
1894         });
1895 }
1896
1897 H264Encoder::~H264Encoder()
1898 {
1899         {
1900                 unique_lock<mutex> lock(storage_task_queue_mutex);
1901                 storage_thread_should_quit = true;
1902                 storage_task_queue_changed.notify_all();
1903         }
1904         {
1905                 unique_lock<mutex> lock(frame_queue_mutex);
1906                 copy_thread_should_quit = true;
1907                 frame_queue_nonempty.notify_one();
1908         }
1909         storage_thread.join();
1910         copy_thread.join();
1911
1912         release_encode();
1913         deinit_va();
1914 }
1915
1916 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1917 {
1918         {
1919                 // Wait until this frame slot is done encoding.
1920                 std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
1921                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1922                 if (storage_thread_should_quit) return false;
1923         }
1924
1925         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1926         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1927         *y_tex = surf->y_tex;
1928         *cbcr_tex = surf->cbcr_tex;
1929
1930         VASurfaceID surface = surf->src_surface;
1931         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1932         CHECK_VASTATUS(va_status, "vaDeriveImage");
1933
1934         VABufferInfo buf_info;
1935         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1936         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1937         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1938
1939         // Create Y image.
1940         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1941         EGLint y_attribs[] = {
1942                 EGL_WIDTH, frame_width,
1943                 EGL_HEIGHT, frame_height,
1944                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1945                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1946                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1947                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1948                 EGL_NONE
1949         };
1950
1951         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1952         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1953
1954         // Associate Y image to a texture.
1955         glBindTexture(GL_TEXTURE_2D, *y_tex);
1956         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1957
1958         // Create CbCr image.
1959         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1960         EGLint cbcr_attribs[] = {
1961                 EGL_WIDTH, frame_width,
1962                 EGL_HEIGHT, frame_height,
1963                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1964                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1965                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1966                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1967                 EGL_NONE
1968         };
1969
1970         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1971         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1972
1973         // Associate CbCr image to a texture.
1974         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1975         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1976
1977         return true;
1978 }
1979
1980 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
1981 {
1982         {
1983                 unique_lock<mutex> lock(frame_queue_mutex);
1984                 pending_audio_frames[pts] = move(audio);
1985         }
1986         frame_queue_nonempty.notify_one();
1987 }
1988
1989
1990 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
1991 {
1992         {
1993                 unique_lock<mutex> lock(frame_queue_mutex);
1994                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1995                 ++current_storage_frame;
1996         }
1997         frame_queue_nonempty.notify_one();
1998 }
1999
2000 void H264Encoder::copy_thread_func()
2001 {
2002         int64_t last_dts = -1;
2003         for ( ;; ) {
2004                 PendingFrame frame;
2005                 int pts_lag;
2006                 encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
2007                                        &current_frame_display, &current_frame_type, &pts_lag);
2008                 if (current_frame_type == FRAME_IDR) {
2009                         numShortTerm = 0;
2010                         current_frame_num = 0;
2011                         current_IDR_display = current_frame_display;
2012                 }
2013
2014                 {
2015                         unique_lock<mutex> lock(frame_queue_mutex);
2016                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; });
2017                         if (copy_thread_should_quit) return;
2018                         frame = move(pending_video_frames[current_frame_display]);
2019                         pending_video_frames.erase(current_frame_display);
2020                 }
2021
2022                 // Wait for the GPU to be done with the frame.
2023                 glClientWaitSync(frame.fence.get(), 0, 0);
2024
2025                 // Release back any input frames we needed to render this frame.
2026                 frame.input_frames.clear();
2027
2028                 // Unmap the image.
2029                 GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
2030                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2031                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2032                 VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2033                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2034                 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2035                 CHECK_VASTATUS(va_status, "vaDestroyImage");
2036
2037                 VASurfaceID surface = surf->src_surface;
2038
2039                 // Schedule the frame for encoding.
2040                 va_status = vaBeginPicture(va_dpy, context_id, surface);
2041                 CHECK_VASTATUS(va_status, "vaBeginPicture");
2042
2043                 if (current_frame_type == FRAME_IDR) {
2044                         render_sequence();
2045                         render_picture();
2046                         if (h264_packedheader) {
2047                                 render_packedsequence();
2048                                 render_packedpicture();
2049                         }
2050                 } else {
2051                         //render_sequence();
2052                         render_picture();
2053                 }
2054                 render_slice();
2055
2056                 va_status = vaEndPicture(va_dpy, context_id);
2057                 CHECK_VASTATUS(va_status, "vaEndPicture");
2058
2059                 // Determine the pts and dts of this frame.
2060                 int64_t pts = frame.pts;
2061                 int64_t dts;
2062                 if (pts_lag == -1) {
2063                         assert(last_dts != -1);
2064                         dts = last_dts + (TIMEBASE / frame_rate);
2065                 } else {
2066                         dts = pts - pts_lag;
2067                 }
2068                 last_dts = dts;
2069
2070                 // so now the data is done encoding (well, async job kicked off)...
2071                 // we send that to the storage thread
2072                 storage_task tmp;
2073                 tmp.display_order = current_frame_display;
2074                 tmp.frame_type = current_frame_type;
2075                 tmp.pts = pts;
2076                 tmp.dts = dts;
2077                 storage_task_enqueue(move(tmp));
2078
2079                 update_ReferenceFrames();
2080                 ++current_frame_encoding;
2081         }
2082 }