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 using namespace std;
  37
  38 class QOpenGLContext;
  39 class QSurface;
  40
  41 #define CHECK_VASTATUS(va_status, func)                                 \
  42     if (va_status != VA_STATUS_SUCCESS) {                               \
  43         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  44         exit(1);                                                        \
  45     }
  46
  47 //#include "loadsurface.h"
  48
  49 #define NAL_REF_IDC_NONE        0
  50 #define NAL_REF_IDC_LOW         1
  51 #define NAL_REF_IDC_MEDIUM      2
  52 #define NAL_REF_IDC_HIGH        3
  53
  54 #define NAL_NON_IDR             1
  55 #define NAL_IDR                 5
  56 #define NAL_SPS                 7
  57 #define NAL_PPS                 8
  58 #define NAL_SEI                 6
  59
  60 #define SLICE_TYPE_P            0
  61 #define SLICE_TYPE_B            1
  62 #define SLICE_TYPE_I            2
  63 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  64 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  65 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  66
  67
  68 #define ENTROPY_MODE_CAVLC      0
  69 #define ENTROPY_MODE_CABAC      1
  70
  71 #define PROFILE_IDC_BASELINE    66
  72 #define PROFILE_IDC_MAIN        77
  73 #define PROFILE_IDC_HIGH        100
  74
  75 #define BITSTREAM_ALLOCATE_STEPPING     4096
  76
  77 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  78 static  VADisplay va_dpy;
  79 static  VAProfile h264_profile = (VAProfile)~0;
  80 static  VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
  81 static  int config_attrib_num = 0, enc_packed_header_idx;
  82
  83 struct GLSurface {
  84         VASurfaceID src_surface, ref_surface;
  85         VABufferID coded_buf;
  86
  87         VAImage surface_image;
  88         GLuint y_tex, cbcr_tex;
  89         EGLImage y_egl_image, cbcr_egl_image;
  90 };
  91 GLSurface gl_surfaces[SURFACE_NUM];
  92
  93 static  VAConfigID config_id;
  94 static  VAContextID context_id;
  95 static  VAEncSequenceParameterBufferH264 seq_param;
  96 static  VAEncPictureParameterBufferH264 pic_param;
  97 static  VAEncSliceParameterBufferH264 slice_param;
  98 static  VAPictureH264 CurrentCurrPic;
  99 static  VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
 100
 101 static  unsigned int MaxFrameNum = (2<<16);
 102 static  unsigned int MaxPicOrderCntLsb = (2<<8);
 103 static  unsigned int Log2MaxFrameNum = 16;
 104 static  unsigned int Log2MaxPicOrderCntLsb = 8;
 105
 106 static  unsigned int num_ref_frames = 2;
 107 static  unsigned int numShortTerm = 0;
 108 static  int constraint_set_flag = 0;
 109 static  int h264_packedheader = 0; /* support pack header? */
 110 static  int h264_maxref = (1<<16|1);
 111 static  int h264_entropy_mode = 1; /* cabac */
 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  double frame_size = 0;
 119 static  int initial_qp = 15;
 120 //static  int initial_qp = 28;
 121 static  int minimal_qp = 0;
 122 static  int intra_period = 30;
 123 static  int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 124 static  int ip_period = 3;
 125 static  int rc_mode = -1;
 126 static  int rc_default_modes[] = {
 127     VA_RC_VBR,
 128     VA_RC_CQP,
 129     VA_RC_VBR_CONSTRAINED,
 130     VA_RC_CBR,
 131     VA_RC_VCM,
 132     VA_RC_NONE,
 133 };
 134 static  unsigned long long current_IDR_display = 0;
 135 static  unsigned int current_frame_num = 0;
 136
 137 static  int misc_priv_type = 0;
 138 static  int misc_priv_value = 0;
 139
 140 /* thread to save coded data */
 141 #define SRC_SURFACE_FREE        0
 142 #define SRC_SURFACE_IN_ENCODING 1
 143
 144 struct __bitstream {
 145     unsigned int *buffer;
 146     int bit_offset;
 147     int max_size_in_dword;
 148 };
 149 typedef struct __bitstream bitstream;
 150
 151 using namespace std;
 152
 153 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 154 // but if we don't delete it here, we get leaks. The GStreamer implementation
 155 // does the same.
 156 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 157 {
 158     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 159     CHECK_VASTATUS(va_status, "vaRenderPicture");
 160
 161     for (int i = 0; i < num_buffers; ++i) {
 162         va_status = vaDestroyBuffer(va_dpy, buffers[i]);
 163         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 164     }
 165 }
 166
 167 static unsigned int
 168 va_swap32(unsigned int val)
 169 {
 170     unsigned char *pval = (unsigned char *)&val;
 171
 172     return ((pval[0] << 24)     |
 173             (pval[1] << 16)     |
 174             (pval[2] << 8)      |
 175             (pval[3] << 0));
 176 }
 177
 178 static void
 179 bitstream_start(bitstream *bs)
 180 {
 181     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 182     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 183     bs->bit_offset = 0;
 184 }
 185
 186 static void
 187 bitstream_end(bitstream *bs)
 188 {
 189     int pos = (bs->bit_offset >> 5);
 190     int bit_offset = (bs->bit_offset & 0x1f);
 191     int bit_left = 32 - bit_offset;
 192
 193     if (bit_offset) {
 194         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 195     }
 196 }
 197
 198 static void
 199 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 200 {
 201     int pos = (bs->bit_offset >> 5);
 202     int bit_offset = (bs->bit_offset & 0x1f);
 203     int bit_left = 32 - bit_offset;
 204
 205     if (!size_in_bits)
 206         return;
 207
 208     bs->bit_offset += size_in_bits;
 209
 210     if (bit_left > size_in_bits) {
 211         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 212     } else {
 213         size_in_bits -= bit_left;
 214         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 215         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 216
 217         if (pos + 1 == bs->max_size_in_dword) {
 218             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 219             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 220         }
 221
 222         bs->buffer[pos + 1] = val;
 223     }
 224 }
 225
 226 static void
 227 bitstream_put_ue(bitstream *bs, unsigned int val)
 228 {
 229     int size_in_bits = 0;
 230     int tmp_val = ++val;
 231
 232     while (tmp_val) {
 233         tmp_val >>= 1;
 234         size_in_bits++;
 235     }
 236
 237     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 238     bitstream_put_ui(bs, val, size_in_bits);
 239 }
 240
 241 static void
 242 bitstream_put_se(bitstream *bs, int val)
 243 {
 244     unsigned int new_val;
 245
 246     if (val <= 0)
 247         new_val = -2 * val;
 248     else
 249         new_val = 2 * val - 1;
 250
 251     bitstream_put_ue(bs, new_val);
 252 }
 253
 254 static void
 255 bitstream_byte_aligning(bitstream *bs, int bit)
 256 {
 257     int bit_offset = (bs->bit_offset & 0x7);
 258     int bit_left = 8 - bit_offset;
 259     int new_val;
 260
 261     if (!bit_offset)
 262         return;
 263
 264     assert(bit == 0 || bit == 1);
 265
 266     if (bit)
 267         new_val = (1 << bit_left) - 1;
 268     else
 269         new_val = 0;
 270
 271     bitstream_put_ui(bs, new_val, bit_left);
 272 }
 273
 274 static void
 275 rbsp_trailing_bits(bitstream *bs)
 276 {
 277     bitstream_put_ui(bs, 1, 1);
 278     bitstream_byte_aligning(bs, 0);
 279 }
 280
 281 static void nal_start_code_prefix(bitstream *bs)
 282 {
 283     bitstream_put_ui(bs, 0x00000001, 32);
 284 }
 285
 286 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 287 {
 288     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 289     bitstream_put_ui(bs, nal_ref_idc, 2);
 290     bitstream_put_ui(bs, nal_unit_type, 5);
 291 }
 292
 293 static void sps_rbsp(bitstream *bs)
 294 {
 295     int profile_idc = PROFILE_IDC_BASELINE;
 296
 297     if (h264_profile  == VAProfileH264High)
 298         profile_idc = PROFILE_IDC_HIGH;
 299     else if (h264_profile  == VAProfileH264Main)
 300         profile_idc = PROFILE_IDC_MAIN;
 301
 302     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 303     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 304     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 305     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 306     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 307     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 308     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 309     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 310
 311     if ( profile_idc == PROFILE_IDC_HIGH) {
 312         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 313         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 314         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 315         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 316         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 317     }
 318
 319     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 320     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 321
 322     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 323         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 324     else {
 325         assert(0);
 326     }
 327
 328     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 329     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 330
 331     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 332     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 333     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 334
 335     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 336         assert(0);
 337     }
 338
 339     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 340     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 341
 342     if (seq_param.frame_cropping_flag) {
 343         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 344         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 345         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 346         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 347     }
 348
 349     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 350     if ( false ) {
 351         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 352     } else {
 353         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 354         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 355         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 356         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 357         {
 358             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 359             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 360             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 361             {
 362                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 363                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 364                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 365             }
 366         }
 367         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 368         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 369         {
 370             bitstream_put_ui(bs, 1, 32);  // FPS
 371             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 372             bitstream_put_ui(bs, 1, 1);
 373         }
 374         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 375         {
 376             // hrd_parameters
 377             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 378             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 379             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 380
 381             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 382             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 383             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 384
 385             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 386             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 387             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 388             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 389         }
 390         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 391         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 392
 393         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 394         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 395     }
 396
 397     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 398 }
 399
 400
 401 static void pps_rbsp(bitstream *bs)
 402 {
 403     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 404     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 405
 406     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 407
 408     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 409
 410     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 411
 412     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 413     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 414
 415     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 416     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 417
 418     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 419     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 420     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 421
 422     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 423     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 424     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 425
 426     /* more_rbsp_data */
 427     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 428     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 429     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 430
 431     rbsp_trailing_bits(bs);
 432 }
 433
 434 static void slice_header(bitstream *bs)
 435 {
 436     int first_mb_in_slice = slice_param.macroblock_address;
 437
 438     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 439     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 440     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 441     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 442
 443     /* frame_mbs_only_flag == 1 */
 444     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 445         /* FIXME: */
 446         assert(0);
 447     }
 448
 449     if (pic_param.pic_fields.bits.idr_pic_flag)
 450         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 451
 452     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 453         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 454         /* pic_order_present_flag == 0 */
 455     } else {
 456         /* FIXME: */
 457         assert(0);
 458     }
 459
 460     /* redundant_pic_cnt_present_flag == 0 */
 461     /* slice type */
 462     if (IS_P_SLICE(slice_param.slice_type)) {
 463         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 464
 465         if (slice_param.num_ref_idx_active_override_flag)
 466             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 467
 468         /* ref_pic_list_reordering */
 469         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 470     } else if (IS_B_SLICE(slice_param.slice_type)) {
 471         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 472
 473         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 474
 475         if (slice_param.num_ref_idx_active_override_flag) {
 476             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 477             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 478         }
 479
 480         /* ref_pic_list_reordering */
 481         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 482         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 483     }
 484
 485     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 486          IS_P_SLICE(slice_param.slice_type)) ||
 487         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 488          IS_B_SLICE(slice_param.slice_type))) {
 489         /* FIXME: fill weight/offset table */
 490         assert(0);
 491     }
 492
 493     /* dec_ref_pic_marking */
 494     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 495         unsigned char no_output_of_prior_pics_flag = 0;
 496         unsigned char long_term_reference_flag = 0;
 497         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 498
 499         if (pic_param.pic_fields.bits.idr_pic_flag) {
 500             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 501             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 502         } else {
 503             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 504         }
 505     }
 506
 507     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 508         !IS_I_SLICE(slice_param.slice_type))
 509         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 510
 511     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 512
 513     /* ignore for SP/SI */
 514
 515     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 516         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 517
 518         if (slice_param.disable_deblocking_filter_idc != 1) {
 519             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 520             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 521         }
 522     }
 523
 524     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 525         bitstream_byte_aligning(bs, 1);
 526     }
 527 }
 528
 529 static int
 530 build_packed_pic_buffer(unsigned char **header_buffer)
 531 {
 532     bitstream bs;
 533
 534     bitstream_start(&bs);
 535     nal_start_code_prefix(&bs);
 536     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 537     pps_rbsp(&bs);
 538     bitstream_end(&bs);
 539
 540     *header_buffer = (unsigned char *)bs.buffer;
 541     return bs.bit_offset;
 542 }
 543
 544 static int
 545 build_packed_seq_buffer(unsigned char **header_buffer)
 546 {
 547     bitstream bs;
 548
 549     bitstream_start(&bs);
 550     nal_start_code_prefix(&bs);
 551     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 552     sps_rbsp(&bs);
 553     bitstream_end(&bs);
 554
 555     *header_buffer = (unsigned char *)bs.buffer;
 556     return bs.bit_offset;
 557 }
 558
 559 static int build_packed_slice_buffer(unsigned char **header_buffer)
 560 {
 561     bitstream bs;
 562     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 563     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 564
 565     bitstream_start(&bs);
 566     nal_start_code_prefix(&bs);
 567
 568     if (IS_I_SLICE(slice_param.slice_type)) {
 569         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 570     } else if (IS_P_SLICE(slice_param.slice_type)) {
 571         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 572     } else {
 573         assert(IS_B_SLICE(slice_param.slice_type));
 574         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 575     }
 576
 577     slice_header(&bs);
 578     bitstream_end(&bs);
 579
 580     *header_buffer = (unsigned char *)bs.buffer;
 581     return bs.bit_offset;
 582 }
 583
 584
 585 /*
 586   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 587   1) period between Frame #X and Frame #N = #X - #N
 588   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 589   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 590   4) intra_period and intra_idr_period take precedence over ip_period
 591   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 592      of I/IDR frames, see bellow examples
 593   -------------------------------------------------------------------
 594   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 595   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 596   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 597   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 598   1            1                ignored    IDR IDR IDR IDR...
 599   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 600   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 601   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 602                                               (PBB)(IBB)(PBB)(IBB)...
 603   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 604                                            IDRPPPPP IPPPPP IPPPPP...
 605   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 606                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 607                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 608                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 609                                            {IDR(PBB)(PBB)}                     (6/6/3)
 610                                            {IDR(PBB)(PBB)}.
 611 */
 612
 613 // General pts/dts strategy:
 614 //
 615 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 616 // bit tricky. We assume first of all that the frame rate never goes _above_
 617 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 618 // in at least this speed, as long at dts <= pts (the frame is not attempted
 619 // presented before it is decoded). Furthermore, we never have longer chains of
 620 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 621 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 622 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 623 // frame, with an order O <= C.)
 624 //
 625 // Many strategies are possible, but we establish these rules:
 626 //
 627 //  - Tip frames have dts = pts - (C-O)*N.
 628 //  - Non-tip frames have dts = dts_last + N.
 629 //
 630 // An example, with C=2 and N=10 and the data flow showed with arrows:
 631 //
 632 //        I  B  P  B  B  P
 633 //   pts: 30 40 50 60 70 80
 634 //        ↓  ↓     ↓
 635 //   dts: 10 30 20 60 50←40
 636 //         |  |  ↑        ↑
 637 //         `--|--'        |
 638 //             `----------'
 639 //
 640 // To show that this works fine also with irregular spacings, let's say that
 641 // the third frame is delayed a bit (something earlier was dropped). Now the
 642 // situation looks like this:
 643 //
 644 //        I  B  P  B  B   P
 645 //   pts: 30 40 80 90 100 110
 646 //        ↓  ↓     ↓
 647 //   dts: 10 30 20 90 50←40
 648 //         |  |  ↑        ↑
 649 //         `--|--'        |
 650 //             `----------'
 651 //
 652 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 653 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 654 //
 655 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 656 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 657 // a tip frame and should be given a dts based on the previous one.
 658 #define FRAME_P 0
 659 #define FRAME_B 1
 660 #define FRAME_I 2
 661 #define FRAME_IDR 7
 662 void encoding2display_order(
 663     int encoding_order, int intra_period,
 664     int intra_idr_period, int ip_period,
 665     int *displaying_order,
 666     int *frame_type, int *pts_lag)
 667 {
 668     int encoding_order_gop = 0;
 669
 670     *pts_lag = 0;
 671
 672     if (intra_period == 1) { /* all are I/IDR frames */
 673         *displaying_order = encoding_order;
 674         if (intra_idr_period == 0)
 675             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 676         else
 677             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 678         return;
 679     }
 680
 681     if (intra_period == 0)
 682         intra_idr_period = 0;
 683
 684     if (ip_period == 1) {
 685         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 686         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 687         *displaying_order = encoding_order;
 688
 689         if (encoding_order_gop == 0) { /* the first frame */
 690             *frame_type = FRAME_IDR;
 691         } else if (intra_period != 0 && /* have I frames */
 692                    encoding_order_gop >= 2 &&
 693                    (encoding_order_gop % intra_period == 0)) {
 694             *frame_type = FRAME_I;
 695         } else {
 696             *frame_type = FRAME_P;
 697         }
 698         return;
 699     }
 700
 701     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 702     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 703     *pts_lag = -1;  // Most frames are not tip frames.
 704
 705     if (encoding_order_gop == 0) { /* the first frame */
 706         *frame_type = FRAME_IDR;
 707         *displaying_order = encoding_order;
 708         // IDR frames are a special case; I honestly can't find the logic behind
 709         // why this is the right thing, but it seems to line up nicely in practice :-)
 710         *pts_lag = TIMEBASE / MAX_FPS;
 711     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 712         *frame_type = FRAME_B;
 713         *displaying_order = encoding_order - 1;
 714         if ((encoding_order_gop % ip_period) == 0) {
 715             *pts_lag = 0;  // Last B-frame.
 716         }
 717     } else if (intra_period != 0 && /* have I frames */
 718                encoding_order_gop >= 2 &&
 719                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 720         *frame_type = FRAME_I;
 721         *displaying_order = encoding_order + ip_period - 1;
 722     } else {
 723         *frame_type = FRAME_P;
 724         *displaying_order = encoding_order + ip_period - 1;
 725     }
 726 }
 727
 728
 729 static const char *rc_to_string(int rcmode)
 730 {
 731     switch (rc_mode) {
 732     case VA_RC_NONE:
 733         return "NONE";
 734     case VA_RC_CBR:
 735         return "CBR";
 736     case VA_RC_VBR:
 737         return "VBR";
 738     case VA_RC_VCM:
 739         return "VCM";
 740     case VA_RC_CQP:
 741         return "CQP";
 742     case VA_RC_VBR_CONSTRAINED:
 743         return "VBR_CONSTRAINED";
 744     default:
 745         return "Unknown";
 746     }
 747 }
 748
 749 #if 0
 750 static int process_cmdline(int argc, char *argv[])
 751 {
 752     char c;
 753     const struct option long_opts[] = {
 754         {"help", no_argument, NULL, 0 },
 755         {"bitrate", required_argument, NULL, 1 },
 756         {"minqp", required_argument, NULL, 2 },
 757         {"initialqp", required_argument, NULL, 3 },
 758         {"intra_period", required_argument, NULL, 4 },
 759         {"idr_period", required_argument, NULL, 5 },
 760         {"ip_period", required_argument, NULL, 6 },
 761         {"rcmode", required_argument, NULL, 7 },
 762         {"srcyuv", required_argument, NULL, 9 },
 763         {"recyuv", required_argument, NULL, 10 },
 764         {"fourcc", required_argument, NULL, 11 },
 765         {"syncmode", no_argument, NULL, 12 },
 766         {"enablePSNR", no_argument, NULL, 13 },
 767         {"prit", required_argument, NULL, 14 },
 768         {"priv", required_argument, NULL, 15 },
 769         {"framecount", required_argument, NULL, 16 },
 770         {"entropy", required_argument, NULL, 17 },
 771         {"profile", required_argument, NULL, 18 },
 772         {NULL, no_argument, NULL, 0 }};
 773     int long_index;
 774
 775     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 776         switch (c) {
 777         case 'w':
 778             frame_width = atoi(optarg);
 779             break;
 780         case 'h':
 781             frame_height = atoi(optarg);
 782             break;
 783         case 'n':
 784         case 'f':
 785             frame_rate = atoi(optarg);
 786             break;
 787         case 'o':
 788             coded_fn = strdup(optarg);
 789             break;
 790         case 0:
 791             print_help();
 792             exit(0);
 793         case 1:
 794             frame_bitrate = atoi(optarg);
 795             break;
 796         case 2:
 797             minimal_qp = atoi(optarg);
 798             break;
 799         case 3:
 800             initial_qp = atoi(optarg);
 801             break;
 802         case 4:
 803             intra_period = atoi(optarg);
 804             break;
 805         case 5:
 806             intra_idr_period = atoi(optarg);
 807             break;
 808         case 6:
 809             ip_period = atoi(optarg);
 810             break;
 811         case 7:
 812             rc_mode = string_to_rc(optarg);
 813             if (rc_mode < 0) {
 814                 print_help();
 815                 exit(1);
 816             }
 817             break;
 818         case 9:
 819             srcyuv_fn = strdup(optarg);
 820             break;
 821         case 11:
 822             srcyuv_fourcc = string_to_fourcc(optarg);
 823             if (srcyuv_fourcc <= 0) {
 824                 print_help();
 825                 exit(1);
 826             }
 827             break;
 828         case 13:
 829             calc_psnr = 1;
 830             break;
 831         case 14:
 832             misc_priv_type = strtol(optarg, NULL, 0);
 833             break;
 834         case 15:
 835             misc_priv_value = strtol(optarg, NULL, 0);
 836             break;
 837         case 17:
 838             h264_entropy_mode = atoi(optarg) ? 1: 0;
 839             break;
 840         case 18:
 841             if (strncmp(optarg, "BP", 2) == 0)
 842                 h264_profile = VAProfileH264Baseline;
 843             else if (strncmp(optarg, "MP", 2) == 0)
 844                 h264_profile = VAProfileH264Main;
 845             else if (strncmp(optarg, "HP", 2) == 0)
 846                 h264_profile = VAProfileH264High;
 847             else
 848                 h264_profile = (VAProfile)0;
 849             break;
 850         case ':':
 851         case '?':
 852             print_help();
 853             exit(0);
 854         }
 855     }
 856
 857     if (ip_period < 1) {
 858         printf(" ip_period must be greater than 0\n");
 859         exit(0);
 860     }
 861     if (intra_period != 1 && intra_period % ip_period != 0) {
 862         printf(" intra_period must be a multiplier of ip_period\n");
 863         exit(0);
 864     }
 865     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 866         printf(" intra_idr_period must be a multiplier of intra_period\n");
 867         exit(0);
 868     }
 869
 870     if (frame_bitrate == 0)
 871         frame_bitrate = frame_width * frame_height * 12 * MAX_FPS / 50;
 872
 873     if (coded_fn == NULL) {
 874         struct stat buf;
 875         if (stat("/tmp", &buf) == 0)
 876             coded_fn = strdup("/tmp/test.264");
 877         else if (stat("/sdcard", &buf) == 0)
 878             coded_fn = strdup("/sdcard/test.264");
 879         else
 880             coded_fn = strdup("./test.264");
 881     }
 882
 883
 884     frame_width_mbaligned = (frame_width + 15) & (~15);
 885     frame_height_mbaligned = (frame_height + 15) & (~15);
 886     if (frame_width != frame_width_mbaligned ||
 887         frame_height != frame_height_mbaligned) {
 888         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 889                frame_width, frame_height,
 890                frame_width_mbaligned, frame_height_mbaligned
 891                );
 892     }
 893
 894     return 0;
 895 }
 896 #endif
 897
 898 static Display *x11_display;
 899 static Window   x11_window;
 900
 901 VADisplay
 902 va_open_display(void)
 903 {
 904     x11_display = XOpenDisplay(NULL);
 905     if (!x11_display) {
 906         fprintf(stderr, "error: can't connect to X server!\n");
 907         return NULL;
 908     }
 909     return vaGetDisplay(x11_display);
 910 }
 911
 912 void
 913 va_close_display(VADisplay va_dpy)
 914 {
 915     if (!x11_display)
 916         return;
 917
 918     if (x11_window) {
 919         XUnmapWindow(x11_display, x11_window);
 920         XDestroyWindow(x11_display, x11_window);
 921         x11_window = None;
 922     }
 923     XCloseDisplay(x11_display);
 924     x11_display = NULL;
 925 }
 926
 927 static int init_va(void)
 928 {
 929     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 930     VAEntrypoint *entrypoints;
 931     int num_entrypoints, slice_entrypoint;
 932     int support_encode = 0;
 933     int major_ver, minor_ver;
 934     VAStatus va_status;
 935     unsigned int i;
 936
 937     va_dpy = va_open_display();
 938     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 939     CHECK_VASTATUS(va_status, "vaInitialize");
 940
 941     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 942     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 943     if (!entrypoints) {
 944         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 945         exit(1);
 946     }
 947
 948     /* use the highest profile */
 949     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 950         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 951             continue;
 952
 953         h264_profile = profile_list[i];
 954         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 955         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 956             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 957                 support_encode = 1;
 958                 break;
 959             }
 960         }
 961         if (support_encode == 1)
 962             break;
 963     }
 964
 965     if (support_encode == 0) {
 966         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 967         exit(1);
 968     } else {
 969         switch (h264_profile) {
 970             case VAProfileH264Baseline:
 971                 ip_period = 1;
 972                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 973                 h264_entropy_mode = 0;
 974                 break;
 975             case VAProfileH264ConstrainedBaseline:
 976                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 977                 ip_period = 1;
 978                 break;
 979
 980             case VAProfileH264Main:
 981                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 982                 break;
 983
 984             case VAProfileH264High:
 985                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 986                 break;
 987             default:
 988                 h264_profile = VAProfileH264Baseline;
 989                 ip_period = 1;
 990                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 991                 break;
 992         }
 993     }
 994
 995     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 996
 997     /* find out the format for the render target, and rate control mode */
 998     for (i = 0; i < VAConfigAttribTypeMax; i++)
 999         attrib[i].type = (VAConfigAttribType)i;
1000
1001     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1002                                       &attrib[0], VAConfigAttribTypeMax);
1003     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1004     /* check the interested configattrib */
1005     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1006         printf("Not find desired YUV420 RT format\n");
1007         exit(1);
1008     } else {
1009         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1010         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1011         config_attrib_num++;
1012     }
1013
1014     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1015         int tmp = attrib[VAConfigAttribRateControl].value;
1016
1017         if (rc_mode == -1 || !(rc_mode & tmp))  {
1018             if (rc_mode != -1) {
1019                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1020             }
1021
1022             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1023                 if (rc_default_modes[i] & tmp) {
1024                     rc_mode = rc_default_modes[i];
1025                     break;
1026                 }
1027             }
1028         }
1029
1030         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1031         config_attrib[config_attrib_num].value = rc_mode;
1032         config_attrib_num++;
1033     }
1034
1035
1036     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1037         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1038
1039         h264_packedheader = 1;
1040         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1041         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1042
1043         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1044             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1045         }
1046
1047         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1048             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1049         }
1050
1051         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1052             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1053         }
1054
1055         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1056             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1057         }
1058
1059         enc_packed_header_idx = config_attrib_num;
1060         config_attrib_num++;
1061     }
1062
1063     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1064         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1065         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1066         config_attrib_num++;
1067     }
1068
1069     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1070         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1071     }
1072
1073     free(entrypoints);
1074     return 0;
1075 }
1076
1077 static int setup_encode()
1078 {
1079     VAStatus va_status;
1080     VASurfaceID *tmp_surfaceid;
1081     int codedbuf_size, i;
1082     static VASurfaceID src_surface[SURFACE_NUM];
1083     static VASurfaceID ref_surface[SURFACE_NUM];
1084
1085     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1086             &config_attrib[0], config_attrib_num, &config_id);
1087     CHECK_VASTATUS(va_status, "vaCreateConfig");
1088
1089     /* create source surfaces */
1090     va_status = vaCreateSurfaces(va_dpy,
1091                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1092                                  &src_surface[0], SURFACE_NUM,
1093                                  NULL, 0);
1094     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1095
1096     /* create reference surfaces */
1097     va_status = vaCreateSurfaces(va_dpy,
1098                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1099                                  &ref_surface[0], SURFACE_NUM,
1100                                  NULL, 0);
1101     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1102
1103     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1104     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1105     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1106
1107     /* Create a context for this encode pipe */
1108     va_status = vaCreateContext(va_dpy, config_id,
1109                                 frame_width_mbaligned, frame_height_mbaligned,
1110                                 VA_PROGRESSIVE,
1111                                 tmp_surfaceid, 2 * SURFACE_NUM,
1112                                 &context_id);
1113     CHECK_VASTATUS(va_status, "vaCreateContext");
1114     free(tmp_surfaceid);
1115
1116     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1117
1118     for (i = 0; i < SURFACE_NUM; i++) {
1119         /* create coded buffer once for all
1120          * other VA buffers which won't be used again after vaRenderPicture.
1121          * so APP can always vaCreateBuffer for every frame
1122          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1123          * so VA won't maintain the coded buffer
1124          */
1125         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1126                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1127         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1128     }
1129
1130     /* create OpenGL objects */
1131     //glGenFramebuffers(SURFACE_NUM, fbos);
1132
1133     for (i = 0; i < SURFACE_NUM; i++) {
1134         glGenTextures(1, &gl_surfaces[i].y_tex);
1135         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1136     }
1137
1138     for (i = 0; i < SURFACE_NUM; i++) {
1139         gl_surfaces[i].src_surface = src_surface[i];
1140         gl_surfaces[i].ref_surface = ref_surface[i];
1141     }
1142
1143     return 0;
1144 }
1145
1146
1147
1148 #define partition(ref, field, key, ascending)   \
1149     while (i <= j) {                            \
1150         if (ascending) {                        \
1151             while (ref[i].field < key)          \
1152                 i++;                            \
1153             while (ref[j].field > key)          \
1154                 j--;                            \
1155         } else {                                \
1156             while (ref[i].field > key)          \
1157                 i++;                            \
1158             while (ref[j].field < key)          \
1159                 j--;                            \
1160         }                                       \
1161         if (i <= j) {                           \
1162             tmp = ref[i];                       \
1163             ref[i] = ref[j];                    \
1164             ref[j] = tmp;                       \
1165             i++;                                \
1166             j--;                                \
1167         }                                       \
1168     }                                           \
1169
1170 static void sort_one(VAPictureH264 ref[], int left, int right,
1171                      int ascending, int frame_idx)
1172 {
1173     int i = left, j = right;
1174     unsigned int key;
1175     VAPictureH264 tmp;
1176
1177     if (frame_idx) {
1178         key = ref[(left + right) / 2].frame_idx;
1179         partition(ref, frame_idx, key, ascending);
1180     } else {
1181         key = ref[(left + right) / 2].TopFieldOrderCnt;
1182         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1183     }
1184
1185     /* recursion */
1186     if (left < j)
1187         sort_one(ref, left, j, ascending, frame_idx);
1188
1189     if (i < right)
1190         sort_one(ref, i, right, ascending, frame_idx);
1191 }
1192
1193 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1194                      int partition_ascending, int list0_ascending, int list1_ascending)
1195 {
1196     int i = left, j = right;
1197     VAPictureH264 tmp;
1198
1199     if (frame_idx) {
1200         partition(ref, frame_idx, key, partition_ascending);
1201     } else {
1202         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1203     }
1204
1205
1206     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1207     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1208 }
1209
1210 static int update_ReferenceFrames(int frame_type)
1211 {
1212     int i;
1213
1214     if (frame_type == FRAME_B)
1215         return 0;
1216
1217     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1218     numShortTerm++;
1219     if (numShortTerm > num_ref_frames)
1220         numShortTerm = num_ref_frames;
1221     for (i=numShortTerm-1; i>0; i--)
1222         ReferenceFrames[i] = ReferenceFrames[i-1];
1223     ReferenceFrames[0] = CurrentCurrPic;
1224
1225     current_frame_num++;
1226     if (current_frame_num > MaxFrameNum)
1227         current_frame_num = 0;
1228
1229     return 0;
1230 }
1231
1232
1233 static int update_RefPicList(int frame_type)
1234 {
1235     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1236
1237     if (frame_type == FRAME_P) {
1238         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1239         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1240     }
1241
1242     if (frame_type == FRAME_B) {
1243         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1244         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1245                  1, 0, 1);
1246
1247         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1248         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1249                  0, 1, 0);
1250     }
1251
1252     return 0;
1253 }
1254
1255
1256 static int render_sequence(void)
1257 {
1258     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1259     VAStatus va_status;
1260     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1261     VAEncMiscParameterRateControl *misc_rate_ctrl;
1262
1263     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1264     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1265     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1266     seq_param.bits_per_second = frame_bitrate;
1267
1268     seq_param.intra_period = intra_period;
1269     seq_param.intra_idr_period = intra_idr_period;
1270     seq_param.ip_period = ip_period;
1271
1272     seq_param.max_num_ref_frames = num_ref_frames;
1273     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1274     seq_param.time_scale = TIMEBASE * 2;
1275     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1276     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1277     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1278     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1279     seq_param.seq_fields.bits.chroma_format_idc = 1;
1280     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1281
1282     if (frame_width != frame_width_mbaligned ||
1283         frame_height != frame_height_mbaligned) {
1284         seq_param.frame_cropping_flag = 1;
1285         seq_param.frame_crop_left_offset = 0;
1286         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1287         seq_param.frame_crop_top_offset = 0;
1288         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1289     }
1290
1291     va_status = vaCreateBuffer(va_dpy, context_id,
1292                                VAEncSequenceParameterBufferType,
1293                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1294     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1295
1296     va_status = vaCreateBuffer(va_dpy, context_id,
1297                                VAEncMiscParameterBufferType,
1298                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1299                                1, NULL, &rc_param_buf);
1300     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1301
1302     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1303     misc_param->type = VAEncMiscParameterTypeRateControl;
1304     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1305     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1306     misc_rate_ctrl->bits_per_second = frame_bitrate;
1307     misc_rate_ctrl->target_percentage = 66;
1308     misc_rate_ctrl->window_size = 1000;
1309     misc_rate_ctrl->initial_qp = initial_qp;
1310     misc_rate_ctrl->min_qp = minimal_qp;
1311     misc_rate_ctrl->basic_unit_size = 0;
1312     vaUnmapBuffer(va_dpy, rc_param_buf);
1313
1314     render_id[0] = seq_param_buf;
1315     render_id[1] = rc_param_buf;
1316
1317     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1318
1319     if (misc_priv_type != 0) {
1320         va_status = vaCreateBuffer(va_dpy, context_id,
1321                                    VAEncMiscParameterBufferType,
1322                                    sizeof(VAEncMiscParameterBuffer),
1323                                    1, NULL, &misc_param_tmpbuf);
1324         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1325         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1326         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1327         misc_param_tmp->data[0] = misc_priv_value;
1328         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1329
1330         render_picture_and_delete(va_dpy, context_id, &misc_param_tmpbuf, 1);
1331     }
1332
1333     return 0;
1334 }
1335
1336 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1337 {
1338     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1339     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1340     int PicOrderCntMsb, TopFieldOrderCnt;
1341
1342     if (frame_type == FRAME_IDR)
1343         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1344     else {
1345         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1346         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1347     }
1348
1349     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1350         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1351         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1352     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1353              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1354         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1355     else
1356         PicOrderCntMsb = prevPicOrderCntMsb;
1357
1358     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1359
1360     if (frame_type != FRAME_B) {
1361         PicOrderCntMsb_ref = PicOrderCntMsb;
1362         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1363     }
1364
1365     return TopFieldOrderCnt;
1366 }
1367
1368 static int render_picture(int frame_type, int display_frame_num)
1369 {
1370     VABufferID pic_param_buf;
1371     VAStatus va_status;
1372     int i = 0;
1373
1374     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1375     pic_param.CurrPic.frame_idx = current_frame_num;
1376     pic_param.CurrPic.flags = 0;
1377     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - current_IDR_display) % MaxPicOrderCntLsb, frame_type);
1378     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1379     CurrentCurrPic = pic_param.CurrPic;
1380
1381     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1382     for (i = numShortTerm; i < SURFACE_NUM; i++) {
1383         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1384         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1385     }
1386
1387     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1388     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1389     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1390     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1391     pic_param.frame_num = current_frame_num;
1392     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1393     pic_param.last_picture = false;  // FIXME
1394     pic_param.pic_init_qp = initial_qp;
1395
1396     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1397                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1398     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1399
1400     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1401
1402     return 0;
1403 }
1404
1405 static int render_packedsequence(void)
1406 {
1407     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1408     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1409     unsigned int length_in_bits;
1410     unsigned char *packedseq_buffer = NULL;
1411     VAStatus va_status;
1412
1413     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1414
1415     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1416
1417     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1418     packedheader_param_buffer.has_emulation_bytes = 0;
1419     va_status = vaCreateBuffer(va_dpy,
1420                                context_id,
1421                                VAEncPackedHeaderParameterBufferType,
1422                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1423                                &packedseq_para_bufid);
1424     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1425
1426     va_status = vaCreateBuffer(va_dpy,
1427                                context_id,
1428                                VAEncPackedHeaderDataBufferType,
1429                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1430                                &packedseq_data_bufid);
1431     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1432
1433     render_id[0] = packedseq_para_bufid;
1434     render_id[1] = packedseq_data_bufid;
1435     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1436
1437     free(packedseq_buffer);
1438
1439     return 0;
1440 }
1441
1442
1443 static int render_packedpicture(void)
1444 {
1445     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1446     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1447     unsigned int length_in_bits;
1448     unsigned char *packedpic_buffer = NULL;
1449     VAStatus va_status;
1450
1451     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1452     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1453     packedheader_param_buffer.bit_length = length_in_bits;
1454     packedheader_param_buffer.has_emulation_bytes = 0;
1455
1456     va_status = vaCreateBuffer(va_dpy,
1457                                context_id,
1458                                VAEncPackedHeaderParameterBufferType,
1459                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1460                                &packedpic_para_bufid);
1461     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1462
1463     va_status = vaCreateBuffer(va_dpy,
1464                                context_id,
1465                                VAEncPackedHeaderDataBufferType,
1466                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1467                                &packedpic_data_bufid);
1468     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1469
1470     render_id[0] = packedpic_para_bufid;
1471     render_id[1] = packedpic_data_bufid;
1472     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1473
1474     free(packedpic_buffer);
1475
1476     return 0;
1477 }
1478
1479 static void render_packedslice()
1480 {
1481     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1482     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1483     unsigned int length_in_bits;
1484     unsigned char *packedslice_buffer = NULL;
1485     VAStatus va_status;
1486
1487     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1488     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1489     packedheader_param_buffer.bit_length = length_in_bits;
1490     packedheader_param_buffer.has_emulation_bytes = 0;
1491
1492     va_status = vaCreateBuffer(va_dpy,
1493                                context_id,
1494                                VAEncPackedHeaderParameterBufferType,
1495                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1496                                &packedslice_para_bufid);
1497     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1498
1499     va_status = vaCreateBuffer(va_dpy,
1500                                context_id,
1501                                VAEncPackedHeaderDataBufferType,
1502                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1503                                &packedslice_data_bufid);
1504     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1505
1506     render_id[0] = packedslice_para_bufid;
1507     render_id[1] = packedslice_data_bufid;
1508     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1509
1510     free(packedslice_buffer);
1511 }
1512
1513 static int render_slice(int encoding_frame_num, int display_frame_num, int frame_type)
1514 {
1515     VABufferID slice_param_buf;
1516     VAStatus va_status;
1517     int i;
1518
1519     update_RefPicList(frame_type);
1520
1521     /* one frame, one slice */
1522     slice_param.macroblock_address = 0;
1523     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1524     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1525     if (frame_type == FRAME_IDR) {
1526         if (encoding_frame_num != 0)
1527             ++slice_param.idr_pic_id;
1528     } else if (frame_type == FRAME_P) {
1529         int refpiclist0_max = h264_maxref & 0xffff;
1530         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1531
1532         for (i = refpiclist0_max; i < 32; i++) {
1533             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1534             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1535         }
1536     } else if (frame_type == FRAME_B) {
1537         int refpiclist0_max = h264_maxref & 0xffff;
1538         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1539
1540         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1541         for (i = refpiclist0_max; i < 32; i++) {
1542             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1543             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1544         }
1545
1546         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1547         for (i = refpiclist1_max; i < 32; i++) {
1548             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1549             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1550         }
1551     }
1552
1553     slice_param.slice_alpha_c0_offset_div2 = 0;
1554     slice_param.slice_beta_offset_div2 = 0;
1555     slice_param.direct_spatial_mv_pred_flag = 1;
1556     slice_param.pic_order_cnt_lsb = (display_frame_num - current_IDR_display) % MaxPicOrderCntLsb;
1557
1558
1559     if (h264_packedheader &&
1560         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1561         render_packedslice();
1562
1563     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1564                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1565     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1566
1567     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1568
1569     return 0;
1570 }
1571
1572
1573
1574 void H264Encoder::save_codeddata(storage_task task)
1575 {
1576     VACodedBufferSegment *buf_list = NULL;
1577     VAStatus va_status;
1578     unsigned int coded_size = 0;
1579
1580     string data;
1581
1582     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1583
1584     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1585     CHECK_VASTATUS(va_status, "vaMapBuffer");
1586     while (buf_list != NULL) {
1587         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1588         buf_list = (VACodedBufferSegment *) buf_list->next;
1589
1590         frame_size += coded_size;
1591     }
1592     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1593
1594     {
1595         // Add video.
1596         AVPacket pkt;
1597         memset(&pkt, 0, sizeof(pkt));
1598         pkt.buf = nullptr;
1599         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1600         pkt.size = data.size();
1601         pkt.stream_index = 0;
1602         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1603             pkt.flags = AV_PKT_FLAG_KEY;
1604         } else {
1605             pkt.flags = 0;
1606         }
1607         //pkt.duration = 1;
1608         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1609     }
1610     // Encode and add all audio frames up to and including the pts of this video frame.
1611     for ( ;; ) {
1612         int64_t audio_pts;
1613         vector<float> audio;
1614         {
1615              unique_lock<mutex> lock(frame_queue_mutex);
1616              frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || !pending_audio_frames.empty(); });
1617              if (copy_thread_should_quit) return;
1618              auto it = pending_audio_frames.begin();
1619              if (it->first > task.pts) break;
1620              audio_pts = it->first;
1621              audio = move(it->second);
1622              pending_audio_frames.erase(it);
1623         }
1624
1625         AVFrame *frame = avcodec_alloc_frame();
1626         frame->nb_samples = audio.size() / 2;
1627         frame->format = AV_SAMPLE_FMT_S32;
1628         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1629
1630         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1631         int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
1632         if (ret < 0) {
1633             fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1634             exit(1);
1635         }
1636         for (int i = 0; i < frame->nb_samples * 2; ++i) {
1637             if (audio[i] >= 1.0f) {
1638                 int_samples[i] = 2147483647;
1639             } else if (audio[i] <= -1.0f) {
1640                 int_samples[i] = -2147483647;
1641             } else {
1642                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1643             }
1644         }
1645
1646         AVPacket pkt;
1647         av_init_packet(&pkt);
1648         pkt.data = nullptr;
1649         pkt.size = 0;
1650         int got_output;
1651         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1652         if (got_output) {
1653             pkt.stream_index = 1;
1654             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1655         }
1656         // TODO: Delayed frames.
1657         avcodec_free_frame(&frame);
1658         av_free_packet(&pkt);
1659         if (audio_pts == task.pts) break;
1660     }
1661
1662 #if 0
1663     printf("\r      "); /* return back to startpoint */
1664     switch (encode_order % 4) {
1665         case 0:
1666             printf("|");
1667             break;
1668         case 1:
1669             printf("/");
1670             break;
1671         case 2:
1672             printf("-");
1673             break;
1674         case 3:
1675             printf("\\");
1676             break;
1677     }
1678     printf("%08lld", encode_order);
1679     printf("(%06d bytes coded)", coded_size);
1680 #endif
1681 }
1682
1683
1684 // this is weird. but it seems to put a new frame onto the queue
1685 void H264Encoder::storage_task_enqueue(storage_task task)
1686 {
1687         unique_lock<mutex> lock(storage_task_queue_mutex);
1688         storage_task_queue.push(move(task));
1689         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1690         storage_task_queue_changed.notify_all();
1691 }
1692
1693 void H264Encoder::storage_task_thread()
1694 {
1695         for ( ;; ) {
1696                 storage_task current;
1697                 {
1698                         // wait until there's an encoded frame
1699                         unique_lock<mutex> lock(storage_task_queue_mutex);
1700                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1701                         if (storage_thread_should_quit) return;
1702                         current = move(storage_task_queue.front());
1703                         storage_task_queue.pop();
1704                 }
1705
1706                 VAStatus va_status;
1707
1708                 // waits for data, then saves it to disk.
1709                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1710                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1711                 save_codeddata(move(current));
1712
1713                 {
1714                         unique_lock<mutex> lock(storage_task_queue_mutex);
1715                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1716                         storage_task_queue_changed.notify_all();
1717                 }
1718         }
1719 }
1720
1721 static int release_encode()
1722 {
1723     int i;
1724
1725     for (i = 0; i < SURFACE_NUM; i++) {
1726         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1727         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1728         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1729     }
1730
1731     vaDestroyContext(va_dpy, context_id);
1732     vaDestroyConfig(va_dpy, config_id);
1733
1734     return 0;
1735 }
1736
1737 static int deinit_va()
1738 {
1739     vaTerminate(va_dpy);
1740
1741     va_close_display(va_dpy);
1742
1743     return 0;
1744 }
1745
1746
1747 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
1748         : current_storage_frame(0), surface(surface), httpd(httpd)
1749 {
1750         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1751         context_audio = avcodec_alloc_context3(codec_audio);
1752         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1753         context_audio->sample_rate = OUTPUT_FREQUENCY;
1754         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1755         context_audio->channels = 2;
1756         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1757         context_audio->time_base = AVRational{1, TIMEBASE};
1758         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1759                 fprintf(stderr, "Could not open codec\n");
1760                 exit(1);
1761         }
1762
1763         frame_width = width;
1764         frame_height = height;
1765         frame_width_mbaligned = (frame_width + 15) & (~15);
1766         frame_height_mbaligned = (frame_height + 15) & (~15);
1767         frame_bitrate = 15000000;  // / 60;
1768
1769         //print_input();
1770
1771         init_va();
1772         setup_encode();
1773
1774         // No frames are ready yet.
1775         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1776
1777         memset(&seq_param, 0, sizeof(seq_param));
1778         memset(&pic_param, 0, sizeof(pic_param));
1779         memset(&slice_param, 0, sizeof(slice_param));
1780
1781         storage_thread = thread(&H264Encoder::storage_task_thread, this);
1782
1783         copy_thread = thread([this]{
1784                 //SDL_GL_MakeCurrent(window, context);
1785                 QOpenGLContext *context = create_context(this->surface);
1786                 eglBindAPI(EGL_OPENGL_API);
1787                 if (!make_current(context, this->surface)) {
1788                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1789                                 eglGetError());
1790                         exit(1);
1791                 }
1792                 copy_thread_func();
1793         });
1794 }
1795
1796 H264Encoder::~H264Encoder()
1797 {
1798         {
1799                 unique_lock<mutex> lock(storage_task_queue_mutex);
1800                 storage_thread_should_quit = true;
1801                 storage_task_queue_changed.notify_all();
1802         }
1803         {
1804                 unique_lock<mutex> lock(frame_queue_mutex);
1805                 copy_thread_should_quit = true;
1806                 frame_queue_nonempty.notify_all();
1807         }
1808         storage_thread.join();
1809         copy_thread.join();
1810
1811         release_encode();
1812         deinit_va();
1813 }
1814
1815 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1816 {
1817         {
1818                 // Wait until this frame slot is done encoding.
1819                 unique_lock<mutex> lock(storage_task_queue_mutex);
1820                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1821                 if (storage_thread_should_quit) return false;
1822         }
1823
1824         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1825         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1826         *y_tex = surf->y_tex;
1827         *cbcr_tex = surf->cbcr_tex;
1828
1829         VASurfaceID surface = surf->src_surface;
1830         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1831         CHECK_VASTATUS(va_status, "vaDeriveImage");
1832
1833         VABufferInfo buf_info;
1834         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1835         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1836         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1837
1838         // Create Y image.
1839         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1840         EGLint y_attribs[] = {
1841                 EGL_WIDTH, frame_width,
1842                 EGL_HEIGHT, frame_height,
1843                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1844                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1845                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1846                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1847                 EGL_NONE
1848         };
1849
1850         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1851         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1852
1853         // Associate Y image to a texture.
1854         glBindTexture(GL_TEXTURE_2D, *y_tex);
1855         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1856
1857         // Create CbCr image.
1858         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1859         EGLint cbcr_attribs[] = {
1860                 EGL_WIDTH, frame_width,
1861                 EGL_HEIGHT, frame_height,
1862                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1863                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1864                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1865                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1866                 EGL_NONE
1867         };
1868
1869         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1870         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1871
1872         // Associate CbCr image to a texture.
1873         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1874         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1875
1876         return true;
1877 }
1878
1879 void H264Encoder::add_audio(int64_t pts, vector<float> audio)
1880 {
1881         {
1882                 unique_lock<mutex> lock(frame_queue_mutex);
1883                 pending_audio_frames[pts] = move(audio);
1884         }
1885         frame_queue_nonempty.notify_all();
1886 }
1887
1888 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
1889 {
1890         {
1891                 unique_lock<mutex> lock(frame_queue_mutex);
1892                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1893                 ++current_storage_frame;
1894         }
1895         frame_queue_nonempty.notify_all();
1896 }
1897
1898 void H264Encoder::copy_thread_func()
1899 {
1900         int64_t last_dts = -1;
1901         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1902                 PendingFrame frame;
1903                 int pts_lag;
1904                 int frame_type, display_frame_num;
1905                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1906                                        &display_frame_num, &frame_type, &pts_lag);
1907                 if (frame_type == FRAME_IDR) {
1908                         numShortTerm = 0;
1909                         current_frame_num = 0;
1910                         current_IDR_display = display_frame_num;
1911                 }
1912
1913                 {
1914                         unique_lock<mutex> lock(frame_queue_mutex);
1915                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1916                                 return copy_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1917                         });
1918                         if (copy_thread_should_quit) {
1919                                 return;
1920                         } else {
1921                                 frame = move(pending_video_frames[display_frame_num]);
1922                                 pending_video_frames.erase(display_frame_num);
1923                         }
1924                 }
1925
1926                 // Determine the dts of this frame.
1927                 int64_t dts;
1928                 if (pts_lag == -1) {
1929                         assert(last_dts != -1);
1930                         dts = last_dts + (TIMEBASE / MAX_FPS);
1931                 } else {
1932                         dts = frame.pts - pts_lag;
1933                 }
1934                 last_dts = dts;
1935
1936                 encode_frame(frame, encoding_frame_num, display_frame_num, frame_type, frame.pts, dts);
1937         }
1938 }
1939
1940 void H264Encoder::encode_frame(H264Encoder::PendingFrame frame, int encoding_frame_num, int display_frame_num,
1941                                int frame_type, int64_t pts, int64_t dts)
1942 {
1943         // Wait for the GPU to be done with the frame.
1944         glClientWaitSync(frame.fence.get(), 0, 0);
1945
1946         // Release back any input frames we needed to render this frame.
1947         frame.input_frames.clear();
1948
1949         // Unmap the image.
1950         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1951         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1952         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1953         VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1954         CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1955         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1956         CHECK_VASTATUS(va_status, "vaDestroyImage");
1957
1958         VASurfaceID surface = surf->src_surface;
1959
1960         // Schedule the frame for encoding.
1961         va_status = vaBeginPicture(va_dpy, context_id, surface);
1962         CHECK_VASTATUS(va_status, "vaBeginPicture");
1963
1964         if (frame_type == FRAME_IDR) {
1965                 render_sequence();
1966                 render_picture(frame_type, display_frame_num);
1967                 if (h264_packedheader) {
1968                         render_packedsequence();
1969                         render_packedpicture();
1970                 }
1971         } else {
1972                 //render_sequence();
1973                 render_picture(frame_type, display_frame_num);
1974         }
1975         render_slice(encoding_frame_num, display_frame_num, frame_type);
1976
1977         va_status = vaEndPicture(va_dpy, context_id);
1978         CHECK_VASTATUS(va_status, "vaEndPicture");
1979
1980         // so now the data is done encoding (well, async job kicked off)...
1981         // we send that to the storage thread
1982         storage_task tmp;
1983         tmp.display_order = display_frame_num;
1984         tmp.frame_type = frame_type;
1985         tmp.pts = pts;
1986         tmp.dts = dts;
1987         storage_task_enqueue(move(tmp));
1988
1989         update_ReferenceFrames(frame_type);
1990 }