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_frame_display = 0;
 135 static  unsigned long long current_IDR_display = 0;
 136 static  unsigned int current_frame_num = 0;
 137 static  int current_frame_type;
 138
 139 static  int misc_priv_type = 0;
 140 static  int misc_priv_value = 0;
 141
 142 /* thread to save coded data */
 143 #define SRC_SURFACE_FREE        0
 144 #define SRC_SURFACE_IN_ENCODING 1
 145
 146 struct __bitstream {
 147     unsigned int *buffer;
 148     int bit_offset;
 149     int max_size_in_dword;
 150 };
 151 typedef struct __bitstream bitstream;
 152
 153 using namespace std;
 154
 155 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 156 // but if we don't delete it here, we get leaks. The GStreamer implementation
 157 // does the same.
 158 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 159 {
 160     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 161     CHECK_VASTATUS(va_status, "vaRenderPicture");
 162
 163     for (int i = 0; i < num_buffers; ++i) {
 164         va_status = vaDestroyBuffer(va_dpy, buffers[i]);
 165         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 166     }
 167 }
 168
 169 static unsigned int
 170 va_swap32(unsigned int val)
 171 {
 172     unsigned char *pval = (unsigned char *)&val;
 173
 174     return ((pval[0] << 24)     |
 175             (pval[1] << 16)     |
 176             (pval[2] << 8)      |
 177             (pval[3] << 0));
 178 }
 179
 180 static void
 181 bitstream_start(bitstream *bs)
 182 {
 183     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 184     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 185     bs->bit_offset = 0;
 186 }
 187
 188 static void
 189 bitstream_end(bitstream *bs)
 190 {
 191     int pos = (bs->bit_offset >> 5);
 192     int bit_offset = (bs->bit_offset & 0x1f);
 193     int bit_left = 32 - bit_offset;
 194
 195     if (bit_offset) {
 196         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 197     }
 198 }
 199
 200 static void
 201 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 202 {
 203     int pos = (bs->bit_offset >> 5);
 204     int bit_offset = (bs->bit_offset & 0x1f);
 205     int bit_left = 32 - bit_offset;
 206
 207     if (!size_in_bits)
 208         return;
 209
 210     bs->bit_offset += size_in_bits;
 211
 212     if (bit_left > size_in_bits) {
 213         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 214     } else {
 215         size_in_bits -= bit_left;
 216         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 217         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 218
 219         if (pos + 1 == bs->max_size_in_dword) {
 220             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 221             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 222         }
 223
 224         bs->buffer[pos + 1] = val;
 225     }
 226 }
 227
 228 static void
 229 bitstream_put_ue(bitstream *bs, unsigned int val)
 230 {
 231     int size_in_bits = 0;
 232     int tmp_val = ++val;
 233
 234     while (tmp_val) {
 235         tmp_val >>= 1;
 236         size_in_bits++;
 237     }
 238
 239     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 240     bitstream_put_ui(bs, val, size_in_bits);
 241 }
 242
 243 static void
 244 bitstream_put_se(bitstream *bs, int val)
 245 {
 246     unsigned int new_val;
 247
 248     if (val <= 0)
 249         new_val = -2 * val;
 250     else
 251         new_val = 2 * val - 1;
 252
 253     bitstream_put_ue(bs, new_val);
 254 }
 255
 256 static void
 257 bitstream_byte_aligning(bitstream *bs, int bit)
 258 {
 259     int bit_offset = (bs->bit_offset & 0x7);
 260     int bit_left = 8 - bit_offset;
 261     int new_val;
 262
 263     if (!bit_offset)
 264         return;
 265
 266     assert(bit == 0 || bit == 1);
 267
 268     if (bit)
 269         new_val = (1 << bit_left) - 1;
 270     else
 271         new_val = 0;
 272
 273     bitstream_put_ui(bs, new_val, bit_left);
 274 }
 275
 276 static void
 277 rbsp_trailing_bits(bitstream *bs)
 278 {
 279     bitstream_put_ui(bs, 1, 1);
 280     bitstream_byte_aligning(bs, 0);
 281 }
 282
 283 static void nal_start_code_prefix(bitstream *bs)
 284 {
 285     bitstream_put_ui(bs, 0x00000001, 32);
 286 }
 287
 288 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 289 {
 290     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 291     bitstream_put_ui(bs, nal_ref_idc, 2);
 292     bitstream_put_ui(bs, nal_unit_type, 5);
 293 }
 294
 295 static void sps_rbsp(bitstream *bs)
 296 {
 297     int profile_idc = PROFILE_IDC_BASELINE;
 298
 299     if (h264_profile  == VAProfileH264High)
 300         profile_idc = PROFILE_IDC_HIGH;
 301     else if (h264_profile  == VAProfileH264Main)
 302         profile_idc = PROFILE_IDC_MAIN;
 303
 304     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 305     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 306     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 307     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 308     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 309     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 310     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 311     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 312
 313     if ( profile_idc == PROFILE_IDC_HIGH) {
 314         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 315         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 316         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 317         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 318         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 319     }
 320
 321     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 322     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 323
 324     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 325         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 326     else {
 327         assert(0);
 328     }
 329
 330     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 331     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 332
 333     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 334     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 335     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 336
 337     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 338         assert(0);
 339     }
 340
 341     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 342     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 343
 344     if (seq_param.frame_cropping_flag) {
 345         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 346         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 347         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 348         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 349     }
 350
 351     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 352     if ( false ) {
 353         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 354     } else {
 355         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 356         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 357         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 358         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 359         {
 360             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 361             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 362             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 363             {
 364                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 365                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 366                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 367             }
 368         }
 369         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 370         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 371         {
 372             bitstream_put_ui(bs, 1, 32);  // FPS
 373             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 374             bitstream_put_ui(bs, 1, 1);
 375         }
 376         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 377         {
 378             // hrd_parameters
 379             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 380             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 381             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 382
 383             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 384             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 385             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 386
 387             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 388             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 389             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 390             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 391         }
 392         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 393         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 394
 395         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 396         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 397     }
 398
 399     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 400 }
 401
 402
 403 static void pps_rbsp(bitstream *bs)
 404 {
 405     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 406     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 407
 408     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 409
 410     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 411
 412     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 413
 414     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 415     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 416
 417     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 418     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 419
 420     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 421     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 422     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 423
 424     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 425     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 426     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 427
 428     /* more_rbsp_data */
 429     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 430     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 431     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 432
 433     rbsp_trailing_bits(bs);
 434 }
 435
 436 static void slice_header(bitstream *bs)
 437 {
 438     int first_mb_in_slice = slice_param.macroblock_address;
 439
 440     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 441     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 442     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 443     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 444
 445     /* frame_mbs_only_flag == 1 */
 446     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 447         /* FIXME: */
 448         assert(0);
 449     }
 450
 451     if (pic_param.pic_fields.bits.idr_pic_flag)
 452         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 453
 454     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 455         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 456         /* pic_order_present_flag == 0 */
 457     } else {
 458         /* FIXME: */
 459         assert(0);
 460     }
 461
 462     /* redundant_pic_cnt_present_flag == 0 */
 463     /* slice type */
 464     if (IS_P_SLICE(slice_param.slice_type)) {
 465         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 466
 467         if (slice_param.num_ref_idx_active_override_flag)
 468             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 469
 470         /* ref_pic_list_reordering */
 471         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 472     } else if (IS_B_SLICE(slice_param.slice_type)) {
 473         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 474
 475         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 476
 477         if (slice_param.num_ref_idx_active_override_flag) {
 478             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 479             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 480         }
 481
 482         /* ref_pic_list_reordering */
 483         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 484         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 485     }
 486
 487     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 488          IS_P_SLICE(slice_param.slice_type)) ||
 489         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 490          IS_B_SLICE(slice_param.slice_type))) {
 491         /* FIXME: fill weight/offset table */
 492         assert(0);
 493     }
 494
 495     /* dec_ref_pic_marking */
 496     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 497         unsigned char no_output_of_prior_pics_flag = 0;
 498         unsigned char long_term_reference_flag = 0;
 499         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 500
 501         if (pic_param.pic_fields.bits.idr_pic_flag) {
 502             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 503             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 504         } else {
 505             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 506         }
 507     }
 508
 509     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 510         !IS_I_SLICE(slice_param.slice_type))
 511         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 512
 513     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 514
 515     /* ignore for SP/SI */
 516
 517     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 518         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 519
 520         if (slice_param.disable_deblocking_filter_idc != 1) {
 521             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 522             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 523         }
 524     }
 525
 526     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 527         bitstream_byte_aligning(bs, 1);
 528     }
 529 }
 530
 531 static int
 532 build_packed_pic_buffer(unsigned char **header_buffer)
 533 {
 534     bitstream bs;
 535
 536     bitstream_start(&bs);
 537     nal_start_code_prefix(&bs);
 538     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 539     pps_rbsp(&bs);
 540     bitstream_end(&bs);
 541
 542     *header_buffer = (unsigned char *)bs.buffer;
 543     return bs.bit_offset;
 544 }
 545
 546 static int
 547 build_packed_seq_buffer(unsigned char **header_buffer)
 548 {
 549     bitstream bs;
 550
 551     bitstream_start(&bs);
 552     nal_start_code_prefix(&bs);
 553     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 554     sps_rbsp(&bs);
 555     bitstream_end(&bs);
 556
 557     *header_buffer = (unsigned char *)bs.buffer;
 558     return bs.bit_offset;
 559 }
 560
 561 static int build_packed_slice_buffer(unsigned char **header_buffer)
 562 {
 563     bitstream bs;
 564     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 565     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 566
 567     bitstream_start(&bs);
 568     nal_start_code_prefix(&bs);
 569
 570     if (IS_I_SLICE(slice_param.slice_type)) {
 571         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 572     } else if (IS_P_SLICE(slice_param.slice_type)) {
 573         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 574     } else {
 575         assert(IS_B_SLICE(slice_param.slice_type));
 576         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 577     }
 578
 579     slice_header(&bs);
 580     bitstream_end(&bs);
 581
 582     *header_buffer = (unsigned char *)bs.buffer;
 583     return bs.bit_offset;
 584 }
 585
 586
 587 /*
 588   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 589   1) period between Frame #X and Frame #N = #X - #N
 590   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 591   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 592   4) intra_period and intra_idr_period take precedence over ip_period
 593   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 594      of I/IDR frames, see bellow examples
 595   -------------------------------------------------------------------
 596   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 597   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 598   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 599   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 600   1            1                ignored    IDR IDR IDR IDR...
 601   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 602   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 603   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 604                                               (PBB)(IBB)(PBB)(IBB)...
 605   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 606                                            IDRPPPPP IPPPPP IPPPPP...
 607   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 608                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 609                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 610                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 611                                            {IDR(PBB)(PBB)}                     (6/6/3)
 612                                            {IDR(PBB)(PBB)}.
 613 */
 614
 615 // General pts/dts strategy:
 616 //
 617 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 618 // bit tricky. We assume first of all that the frame rate never goes _above_
 619 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 620 // in at least this speed, as long at dts <= pts (the frame is not attempted
 621 // presented before it is decoded). Furthermore, we never have longer chains of
 622 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 623 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 624 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 625 // frame, with an order O <= C.)
 626 //
 627 // Many strategies are possible, but we establish these rules:
 628 //
 629 //  - Tip frames have dts = pts - (C-O)*N.
 630 //  - Non-tip frames have dts = dts_last + N.
 631 //
 632 // An example, with C=2 and N=10 and the data flow showed with arrows:
 633 //
 634 //        I  B  P  B  B  P
 635 //   pts: 30 40 50 60 70 80
 636 //        ↓  ↓     ↓
 637 //   dts: 10 30 20 60 50←40
 638 //         |  |  ↑        ↑
 639 //         `--|--'        |
 640 //             `----------'
 641 //
 642 // To show that this works fine also with irregular spacings, let's say that
 643 // the third frame is delayed a bit (something earlier was dropped). Now the
 644 // situation looks like this:
 645 //
 646 //        I  B  P  B  B   P
 647 //   pts: 30 40 80 90 100 110
 648 //        ↓  ↓     ↓
 649 //   dts: 10 30 20 90 50←40
 650 //         |  |  ↑        ↑
 651 //         `--|--'        |
 652 //             `----------'
 653 //
 654 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 655 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 656 //
 657 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 658 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 659 // a tip frame and should be given a dts based on the previous one.
 660 #define FRAME_P 0
 661 #define FRAME_B 1
 662 #define FRAME_I 2
 663 #define FRAME_IDR 7
 664 void encoding2display_order(
 665     unsigned long long encoding_order, int intra_period,
 666     int intra_idr_period, int ip_period,
 667     unsigned long long *displaying_order,
 668     int *frame_type, int *pts_lag)
 669 {
 670     int encoding_order_gop = 0;
 671
 672     *pts_lag = 0;
 673
 674     if (intra_period == 1) { /* all are I/IDR frames */
 675         *displaying_order = encoding_order;
 676         if (intra_idr_period == 0)
 677             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 678         else
 679             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 680         return;
 681     }
 682
 683     if (intra_period == 0)
 684         intra_idr_period = 0;
 685
 686     if (ip_period == 1) {
 687         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 688         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 689         *displaying_order = encoding_order;
 690
 691         if (encoding_order_gop == 0) { /* the first frame */
 692             *frame_type = FRAME_IDR;
 693         } else if (intra_period != 0 && /* have I frames */
 694                    encoding_order_gop >= 2 &&
 695                    (encoding_order_gop % intra_period == 0)) {
 696             *frame_type = FRAME_I;
 697         } else {
 698             *frame_type = FRAME_P;
 699         }
 700         return;
 701     }
 702
 703     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 704     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 705     *pts_lag = -1;  // Most frames are not tip frames.
 706
 707     if (encoding_order_gop == 0) { /* the first frame */
 708         *frame_type = FRAME_IDR;
 709         *displaying_order = encoding_order;
 710         // IDR frames are a special case; I honestly can't find the logic behind
 711         // why this is the right thing, but it seems to line up nicely in practice :-)
 712         *pts_lag = TIMEBASE / MAX_FPS;
 713     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 714         *frame_type = FRAME_B;
 715         *displaying_order = encoding_order - 1;
 716         if ((encoding_order_gop % ip_period) == 0) {
 717             *pts_lag = 0;  // Last B-frame.
 718         }
 719     } else if (intra_period != 0 && /* have I frames */
 720                encoding_order_gop >= 2 &&
 721                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 722         *frame_type = FRAME_I;
 723         *displaying_order = encoding_order + ip_period - 1;
 724     } else {
 725         *frame_type = FRAME_P;
 726         *displaying_order = encoding_order + ip_period - 1;
 727     }
 728 }
 729
 730
 731 static const char *rc_to_string(int rcmode)
 732 {
 733     switch (rc_mode) {
 734     case VA_RC_NONE:
 735         return "NONE";
 736     case VA_RC_CBR:
 737         return "CBR";
 738     case VA_RC_VBR:
 739         return "VBR";
 740     case VA_RC_VCM:
 741         return "VCM";
 742     case VA_RC_CQP:
 743         return "CQP";
 744     case VA_RC_VBR_CONSTRAINED:
 745         return "VBR_CONSTRAINED";
 746     default:
 747         return "Unknown";
 748     }
 749 }
 750
 751 #if 0
 752 static int process_cmdline(int argc, char *argv[])
 753 {
 754     char c;
 755     const struct option long_opts[] = {
 756         {"help", no_argument, NULL, 0 },
 757         {"bitrate", required_argument, NULL, 1 },
 758         {"minqp", required_argument, NULL, 2 },
 759         {"initialqp", required_argument, NULL, 3 },
 760         {"intra_period", required_argument, NULL, 4 },
 761         {"idr_period", required_argument, NULL, 5 },
 762         {"ip_period", required_argument, NULL, 6 },
 763         {"rcmode", required_argument, NULL, 7 },
 764         {"srcyuv", required_argument, NULL, 9 },
 765         {"recyuv", required_argument, NULL, 10 },
 766         {"fourcc", required_argument, NULL, 11 },
 767         {"syncmode", no_argument, NULL, 12 },
 768         {"enablePSNR", no_argument, NULL, 13 },
 769         {"prit", required_argument, NULL, 14 },
 770         {"priv", required_argument, NULL, 15 },
 771         {"framecount", required_argument, NULL, 16 },
 772         {"entropy", required_argument, NULL, 17 },
 773         {"profile", required_argument, NULL, 18 },
 774         {NULL, no_argument, NULL, 0 }};
 775     int long_index;
 776
 777     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 778         switch (c) {
 779         case 'w':
 780             frame_width = atoi(optarg);
 781             break;
 782         case 'h':
 783             frame_height = atoi(optarg);
 784             break;
 785         case 'n':
 786         case 'f':
 787             frame_rate = atoi(optarg);
 788             break;
 789         case 'o':
 790             coded_fn = strdup(optarg);
 791             break;
 792         case 0:
 793             print_help();
 794             exit(0);
 795         case 1:
 796             frame_bitrate = atoi(optarg);
 797             break;
 798         case 2:
 799             minimal_qp = atoi(optarg);
 800             break;
 801         case 3:
 802             initial_qp = atoi(optarg);
 803             break;
 804         case 4:
 805             intra_period = atoi(optarg);
 806             break;
 807         case 5:
 808             intra_idr_period = atoi(optarg);
 809             break;
 810         case 6:
 811             ip_period = atoi(optarg);
 812             break;
 813         case 7:
 814             rc_mode = string_to_rc(optarg);
 815             if (rc_mode < 0) {
 816                 print_help();
 817                 exit(1);
 818             }
 819             break;
 820         case 9:
 821             srcyuv_fn = strdup(optarg);
 822             break;
 823         case 11:
 824             srcyuv_fourcc = string_to_fourcc(optarg);
 825             if (srcyuv_fourcc <= 0) {
 826                 print_help();
 827                 exit(1);
 828             }
 829             break;
 830         case 13:
 831             calc_psnr = 1;
 832             break;
 833         case 14:
 834             misc_priv_type = strtol(optarg, NULL, 0);
 835             break;
 836         case 15:
 837             misc_priv_value = strtol(optarg, NULL, 0);
 838             break;
 839         case 17:
 840             h264_entropy_mode = atoi(optarg) ? 1: 0;
 841             break;
 842         case 18:
 843             if (strncmp(optarg, "BP", 2) == 0)
 844                 h264_profile = VAProfileH264Baseline;
 845             else if (strncmp(optarg, "MP", 2) == 0)
 846                 h264_profile = VAProfileH264Main;
 847             else if (strncmp(optarg, "HP", 2) == 0)
 848                 h264_profile = VAProfileH264High;
 849             else
 850                 h264_profile = (VAProfile)0;
 851             break;
 852         case ':':
 853         case '?':
 854             print_help();
 855             exit(0);
 856         }
 857     }
 858
 859     if (ip_period < 1) {
 860         printf(" ip_period must be greater than 0\n");
 861         exit(0);
 862     }
 863     if (intra_period != 1 && intra_period % ip_period != 0) {
 864         printf(" intra_period must be a multiplier of ip_period\n");
 865         exit(0);
 866     }
 867     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 868         printf(" intra_idr_period must be a multiplier of intra_period\n");
 869         exit(0);
 870     }
 871
 872     if (frame_bitrate == 0)
 873         frame_bitrate = frame_width * frame_height * 12 * MAX_FPS / 50;
 874
 875     if (coded_fn == NULL) {
 876         struct stat buf;
 877         if (stat("/tmp", &buf) == 0)
 878             coded_fn = strdup("/tmp/test.264");
 879         else if (stat("/sdcard", &buf) == 0)
 880             coded_fn = strdup("/sdcard/test.264");
 881         else
 882             coded_fn = strdup("./test.264");
 883     }
 884
 885
 886     frame_width_mbaligned = (frame_width + 15) & (~15);
 887     frame_height_mbaligned = (frame_height + 15) & (~15);
 888     if (frame_width != frame_width_mbaligned ||
 889         frame_height != frame_height_mbaligned) {
 890         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 891                frame_width, frame_height,
 892                frame_width_mbaligned, frame_height_mbaligned
 893                );
 894     }
 895
 896     return 0;
 897 }
 898 #endif
 899
 900 static Display *x11_display;
 901 static Window   x11_window;
 902
 903 VADisplay
 904 va_open_display(void)
 905 {
 906     x11_display = XOpenDisplay(NULL);
 907     if (!x11_display) {
 908         fprintf(stderr, "error: can't connect to X server!\n");
 909         return NULL;
 910     }
 911     return vaGetDisplay(x11_display);
 912 }
 913
 914 void
 915 va_close_display(VADisplay va_dpy)
 916 {
 917     if (!x11_display)
 918         return;
 919
 920     if (x11_window) {
 921         XUnmapWindow(x11_display, x11_window);
 922         XDestroyWindow(x11_display, x11_window);
 923         x11_window = None;
 924     }
 925     XCloseDisplay(x11_display);
 926     x11_display = NULL;
 927 }
 928
 929 static int init_va(void)
 930 {
 931     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 932     VAEntrypoint *entrypoints;
 933     int num_entrypoints, slice_entrypoint;
 934     int support_encode = 0;
 935     int major_ver, minor_ver;
 936     VAStatus va_status;
 937     unsigned int i;
 938
 939     va_dpy = va_open_display();
 940     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 941     CHECK_VASTATUS(va_status, "vaInitialize");
 942
 943     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 944     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 945     if (!entrypoints) {
 946         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 947         exit(1);
 948     }
 949
 950     /* use the highest profile */
 951     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 952         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 953             continue;
 954
 955         h264_profile = profile_list[i];
 956         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 957         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 958             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 959                 support_encode = 1;
 960                 break;
 961             }
 962         }
 963         if (support_encode == 1)
 964             break;
 965     }
 966
 967     if (support_encode == 0) {
 968         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
 969         exit(1);
 970     } else {
 971         switch (h264_profile) {
 972             case VAProfileH264Baseline:
 973                 ip_period = 1;
 974                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 975                 h264_entropy_mode = 0;
 976                 break;
 977             case VAProfileH264ConstrainedBaseline:
 978                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 979                 ip_period = 1;
 980                 break;
 981
 982             case VAProfileH264Main:
 983                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 984                 break;
 985
 986             case VAProfileH264High:
 987                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 988                 break;
 989             default:
 990                 h264_profile = VAProfileH264Baseline;
 991                 ip_period = 1;
 992                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 993                 break;
 994         }
 995     }
 996
 997     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 998
 999     /* find out the format for the render target, and rate control mode */
1000     for (i = 0; i < VAConfigAttribTypeMax; i++)
1001         attrib[i].type = (VAConfigAttribType)i;
1002
1003     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1004                                       &attrib[0], VAConfigAttribTypeMax);
1005     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1006     /* check the interested configattrib */
1007     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1008         printf("Not find desired YUV420 RT format\n");
1009         exit(1);
1010     } else {
1011         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1012         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1013         config_attrib_num++;
1014     }
1015
1016     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1017         int tmp = attrib[VAConfigAttribRateControl].value;
1018
1019         if (rc_mode == -1 || !(rc_mode & tmp))  {
1020             if (rc_mode != -1) {
1021                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1022             }
1023
1024             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1025                 if (rc_default_modes[i] & tmp) {
1026                     rc_mode = rc_default_modes[i];
1027                     break;
1028                 }
1029             }
1030         }
1031
1032         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1033         config_attrib[config_attrib_num].value = rc_mode;
1034         config_attrib_num++;
1035     }
1036
1037
1038     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1039         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1040
1041         h264_packedheader = 1;
1042         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1043         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1044
1045         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1046             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1047         }
1048
1049         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1050             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1051         }
1052
1053         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1054             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1055         }
1056
1057         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1058             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1059         }
1060
1061         enc_packed_header_idx = config_attrib_num;
1062         config_attrib_num++;
1063     }
1064
1065     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1066         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1067         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1068         config_attrib_num++;
1069     }
1070
1071     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1072         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1073     }
1074
1075     free(entrypoints);
1076     return 0;
1077 }
1078
1079 static int setup_encode()
1080 {
1081     VAStatus va_status;
1082     VASurfaceID *tmp_surfaceid;
1083     int codedbuf_size, i;
1084     static VASurfaceID src_surface[SURFACE_NUM];
1085     static VASurfaceID ref_surface[SURFACE_NUM];
1086
1087     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1088             &config_attrib[0], config_attrib_num, &config_id);
1089     CHECK_VASTATUS(va_status, "vaCreateConfig");
1090
1091     /* create source surfaces */
1092     va_status = vaCreateSurfaces(va_dpy,
1093                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1094                                  &src_surface[0], SURFACE_NUM,
1095                                  NULL, 0);
1096     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1097
1098     /* create reference surfaces */
1099     va_status = vaCreateSurfaces(va_dpy,
1100                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1101                                  &ref_surface[0], SURFACE_NUM,
1102                                  NULL, 0);
1103     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1104
1105     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1106     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1107     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1108
1109     /* Create a context for this encode pipe */
1110     va_status = vaCreateContext(va_dpy, config_id,
1111                                 frame_width_mbaligned, frame_height_mbaligned,
1112                                 VA_PROGRESSIVE,
1113                                 tmp_surfaceid, 2 * SURFACE_NUM,
1114                                 &context_id);
1115     CHECK_VASTATUS(va_status, "vaCreateContext");
1116     free(tmp_surfaceid);
1117
1118     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1119
1120     for (i = 0; i < SURFACE_NUM; i++) {
1121         /* create coded buffer once for all
1122          * other VA buffers which won't be used again after vaRenderPicture.
1123          * so APP can always vaCreateBuffer for every frame
1124          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1125          * so VA won't maintain the coded buffer
1126          */
1127         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1128                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1129         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1130     }
1131
1132     /* create OpenGL objects */
1133     //glGenFramebuffers(SURFACE_NUM, fbos);
1134
1135     for (i = 0; i < SURFACE_NUM; i++) {
1136         glGenTextures(1, &gl_surfaces[i].y_tex);
1137         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1138     }
1139
1140     for (i = 0; i < SURFACE_NUM; i++) {
1141         gl_surfaces[i].src_surface = src_surface[i];
1142         gl_surfaces[i].ref_surface = ref_surface[i];
1143     }
1144
1145     return 0;
1146 }
1147
1148
1149
1150 #define partition(ref, field, key, ascending)   \
1151     while (i <= j) {                            \
1152         if (ascending) {                        \
1153             while (ref[i].field < key)          \
1154                 i++;                            \
1155             while (ref[j].field > key)          \
1156                 j--;                            \
1157         } else {                                \
1158             while (ref[i].field > key)          \
1159                 i++;                            \
1160             while (ref[j].field < key)          \
1161                 j--;                            \
1162         }                                       \
1163         if (i <= j) {                           \
1164             tmp = ref[i];                       \
1165             ref[i] = ref[j];                    \
1166             ref[j] = tmp;                       \
1167             i++;                                \
1168             j--;                                \
1169         }                                       \
1170     }                                           \
1171
1172 static void sort_one(VAPictureH264 ref[], int left, int right,
1173                      int ascending, int frame_idx)
1174 {
1175     int i = left, j = right;
1176     unsigned int key;
1177     VAPictureH264 tmp;
1178
1179     if (frame_idx) {
1180         key = ref[(left + right) / 2].frame_idx;
1181         partition(ref, frame_idx, key, ascending);
1182     } else {
1183         key = ref[(left + right) / 2].TopFieldOrderCnt;
1184         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1185     }
1186
1187     /* recursion */
1188     if (left < j)
1189         sort_one(ref, left, j, ascending, frame_idx);
1190
1191     if (i < right)
1192         sort_one(ref, i, right, ascending, frame_idx);
1193 }
1194
1195 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1196                      int partition_ascending, int list0_ascending, int list1_ascending)
1197 {
1198     int i = left, j = right;
1199     VAPictureH264 tmp;
1200
1201     if (frame_idx) {
1202         partition(ref, frame_idx, key, partition_ascending);
1203     } else {
1204         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1205     }
1206
1207
1208     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1209     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1210 }
1211
1212 static int update_ReferenceFrames(void)
1213 {
1214     int i;
1215
1216     if (current_frame_type == FRAME_B)
1217         return 0;
1218
1219     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1220     numShortTerm++;
1221     if (numShortTerm > num_ref_frames)
1222         numShortTerm = num_ref_frames;
1223     for (i=numShortTerm-1; i>0; i--)
1224         ReferenceFrames[i] = ReferenceFrames[i-1];
1225     ReferenceFrames[0] = CurrentCurrPic;
1226
1227     if (current_frame_type != FRAME_B)
1228         current_frame_num++;
1229     if (current_frame_num > MaxFrameNum)
1230         current_frame_num = 0;
1231
1232     return 0;
1233 }
1234
1235
1236 static int update_RefPicList(void)
1237 {
1238     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1239
1240     if (current_frame_type == FRAME_P) {
1241         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1242         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1243     }
1244
1245     if (current_frame_type == FRAME_B) {
1246         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1247         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1248                  1, 0, 1);
1249
1250         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1251         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1252                  0, 1, 0);
1253     }
1254
1255     return 0;
1256 }
1257
1258
1259 static int render_sequence(void)
1260 {
1261     VABufferID seq_param_buf, rc_param_buf, misc_param_tmpbuf, render_id[2];
1262     VAStatus va_status;
1263     VAEncMiscParameterBuffer *misc_param, *misc_param_tmp;
1264     VAEncMiscParameterRateControl *misc_rate_ctrl;
1265
1266     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1267     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1268     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1269     seq_param.bits_per_second = frame_bitrate;
1270
1271     seq_param.intra_period = intra_period;
1272     seq_param.intra_idr_period = intra_idr_period;
1273     seq_param.ip_period = ip_period;
1274
1275     seq_param.max_num_ref_frames = num_ref_frames;
1276     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1277     seq_param.time_scale = TIMEBASE * 2;
1278     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1279     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1280     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1281     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1282     seq_param.seq_fields.bits.chroma_format_idc = 1;
1283     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1284
1285     if (frame_width != frame_width_mbaligned ||
1286         frame_height != frame_height_mbaligned) {
1287         seq_param.frame_cropping_flag = 1;
1288         seq_param.frame_crop_left_offset = 0;
1289         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1290         seq_param.frame_crop_top_offset = 0;
1291         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1292     }
1293
1294     va_status = vaCreateBuffer(va_dpy, context_id,
1295                                VAEncSequenceParameterBufferType,
1296                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1297     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1298
1299     va_status = vaCreateBuffer(va_dpy, context_id,
1300                                VAEncMiscParameterBufferType,
1301                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1302                                1, NULL, &rc_param_buf);
1303     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1304
1305     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1306     misc_param->type = VAEncMiscParameterTypeRateControl;
1307     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1308     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1309     misc_rate_ctrl->bits_per_second = frame_bitrate;
1310     misc_rate_ctrl->target_percentage = 66;
1311     misc_rate_ctrl->window_size = 1000;
1312     misc_rate_ctrl->initial_qp = initial_qp;
1313     misc_rate_ctrl->min_qp = minimal_qp;
1314     misc_rate_ctrl->basic_unit_size = 0;
1315     vaUnmapBuffer(va_dpy, rc_param_buf);
1316
1317     render_id[0] = seq_param_buf;
1318     render_id[1] = rc_param_buf;
1319
1320     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1321
1322     if (misc_priv_type != 0) {
1323         va_status = vaCreateBuffer(va_dpy, context_id,
1324                                    VAEncMiscParameterBufferType,
1325                                    sizeof(VAEncMiscParameterBuffer),
1326                                    1, NULL, &misc_param_tmpbuf);
1327         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1328         vaMapBuffer(va_dpy, misc_param_tmpbuf, (void **)&misc_param_tmp);
1329         misc_param_tmp->type = (VAEncMiscParameterType)misc_priv_type;
1330         misc_param_tmp->data[0] = misc_priv_value;
1331         vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
1332
1333         render_picture_and_delete(va_dpy, context_id, &misc_param_tmpbuf, 1);
1334     }
1335
1336     return 0;
1337 }
1338
1339 static int calc_poc(int pic_order_cnt_lsb)
1340 {
1341     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1342     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1343     int PicOrderCntMsb, TopFieldOrderCnt;
1344
1345     if (current_frame_type == FRAME_IDR)
1346         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1347     else {
1348         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1349         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1350     }
1351
1352     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1353         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1354         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1355     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1356              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1357         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1358     else
1359         PicOrderCntMsb = prevPicOrderCntMsb;
1360
1361     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1362
1363     if (current_frame_type != FRAME_B) {
1364         PicOrderCntMsb_ref = PicOrderCntMsb;
1365         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1366     }
1367
1368     return TopFieldOrderCnt;
1369 }
1370
1371 static int render_picture(void)
1372 {
1373     VABufferID pic_param_buf;
1374     VAStatus va_status;
1375     int i = 0;
1376
1377     pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface;
1378     pic_param.CurrPic.frame_idx = current_frame_num;
1379     pic_param.CurrPic.flags = 0;
1380     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb);
1381     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1382     CurrentCurrPic = pic_param.CurrPic;
1383
1384     if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */
1385         update_RefPicList(); /* calc RefPicList */
1386         memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */
1387         if (current_frame_type == FRAME_P) {
1388             pic_param.ReferenceFrames[0] = RefPicList0_P[0];
1389         } else if (current_frame_type == FRAME_B) {
1390             pic_param.ReferenceFrames[0] = RefPicList0_B[0];
1391             pic_param.ReferenceFrames[1] = RefPicList1_B[0];
1392         }
1393     } else {
1394         memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1395         for (i = numShortTerm; i < SURFACE_NUM; i++) {
1396             pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1397             pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1398         }
1399     }
1400
1401     pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR);
1402     pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B);
1403     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1404     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1405     pic_param.frame_num = current_frame_num;
1406     pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf;
1407     pic_param.last_picture = false;  // FIXME
1408     pic_param.pic_init_qp = initial_qp;
1409
1410     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1411                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1412     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1413
1414     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1415
1416     return 0;
1417 }
1418
1419 static int render_packedsequence(void)
1420 {
1421     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1422     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1423     unsigned int length_in_bits;
1424     unsigned char *packedseq_buffer = NULL;
1425     VAStatus va_status;
1426
1427     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1428
1429     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1430
1431     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1432     packedheader_param_buffer.has_emulation_bytes = 0;
1433     va_status = vaCreateBuffer(va_dpy,
1434                                context_id,
1435                                VAEncPackedHeaderParameterBufferType,
1436                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1437                                &packedseq_para_bufid);
1438     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1439
1440     va_status = vaCreateBuffer(va_dpy,
1441                                context_id,
1442                                VAEncPackedHeaderDataBufferType,
1443                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1444                                &packedseq_data_bufid);
1445     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1446
1447     render_id[0] = packedseq_para_bufid;
1448     render_id[1] = packedseq_data_bufid;
1449     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1450
1451     free(packedseq_buffer);
1452
1453     return 0;
1454 }
1455
1456
1457 static int render_packedpicture(void)
1458 {
1459     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1460     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1461     unsigned int length_in_bits;
1462     unsigned char *packedpic_buffer = NULL;
1463     VAStatus va_status;
1464
1465     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1466     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1467     packedheader_param_buffer.bit_length = length_in_bits;
1468     packedheader_param_buffer.has_emulation_bytes = 0;
1469
1470     va_status = vaCreateBuffer(va_dpy,
1471                                context_id,
1472                                VAEncPackedHeaderParameterBufferType,
1473                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1474                                &packedpic_para_bufid);
1475     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1476
1477     va_status = vaCreateBuffer(va_dpy,
1478                                context_id,
1479                                VAEncPackedHeaderDataBufferType,
1480                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1481                                &packedpic_data_bufid);
1482     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1483
1484     render_id[0] = packedpic_para_bufid;
1485     render_id[1] = packedpic_data_bufid;
1486     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1487
1488     free(packedpic_buffer);
1489
1490     return 0;
1491 }
1492
1493 static void render_packedslice()
1494 {
1495     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1496     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1497     unsigned int length_in_bits;
1498     unsigned char *packedslice_buffer = NULL;
1499     VAStatus va_status;
1500
1501     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1502     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1503     packedheader_param_buffer.bit_length = length_in_bits;
1504     packedheader_param_buffer.has_emulation_bytes = 0;
1505
1506     va_status = vaCreateBuffer(va_dpy,
1507                                context_id,
1508                                VAEncPackedHeaderParameterBufferType,
1509                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1510                                &packedslice_para_bufid);
1511     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1512
1513     va_status = vaCreateBuffer(va_dpy,
1514                                context_id,
1515                                VAEncPackedHeaderDataBufferType,
1516                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1517                                &packedslice_data_bufid);
1518     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1519
1520     render_id[0] = packedslice_para_bufid;
1521     render_id[1] = packedslice_data_bufid;
1522     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1523
1524     free(packedslice_buffer);
1525 }
1526
1527 static int render_slice(int encoding_frame_num)
1528 {
1529     VABufferID slice_param_buf;
1530     VAStatus va_status;
1531     int i;
1532
1533     update_RefPicList();
1534
1535     /* one frame, one slice */
1536     slice_param.macroblock_address = 0;
1537     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1538     slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type;
1539     if (current_frame_type == FRAME_IDR) {
1540         if (encoding_frame_num != 0)
1541             ++slice_param.idr_pic_id;
1542     } else if (current_frame_type == FRAME_P) {
1543         int refpiclist0_max = h264_maxref & 0xffff;
1544         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1545
1546         for (i = refpiclist0_max; i < 32; i++) {
1547             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1548             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1549         }
1550     } else if (current_frame_type == FRAME_B) {
1551         int refpiclist0_max = h264_maxref & 0xffff;
1552         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1553
1554         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1555         for (i = refpiclist0_max; i < 32; i++) {
1556             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1557             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1558         }
1559
1560         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1561         for (i = refpiclist1_max; i < 32; i++) {
1562             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1563             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1564         }
1565     }
1566
1567     slice_param.slice_alpha_c0_offset_div2 = 0;
1568     slice_param.slice_beta_offset_div2 = 0;
1569     slice_param.direct_spatial_mv_pred_flag = 1;
1570     slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb;
1571
1572
1573     if (h264_packedheader &&
1574         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1575         render_packedslice();
1576
1577     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1578                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1579     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1580
1581     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1582
1583     return 0;
1584 }
1585
1586
1587
1588 void H264Encoder::save_codeddata(storage_task task)
1589 {
1590     VACodedBufferSegment *buf_list = NULL;
1591     VAStatus va_status;
1592     unsigned int coded_size = 0;
1593
1594     string data;
1595
1596     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1597
1598     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1599     CHECK_VASTATUS(va_status, "vaMapBuffer");
1600     while (buf_list != NULL) {
1601         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1602         buf_list = (VACodedBufferSegment *) buf_list->next;
1603
1604         frame_size += coded_size;
1605     }
1606     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1607
1608     {
1609         // Add video.
1610         AVPacket pkt;
1611         memset(&pkt, 0, sizeof(pkt));
1612         pkt.buf = nullptr;
1613         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1614         pkt.size = data.size();
1615         pkt.stream_index = 0;
1616         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1617             pkt.flags = AV_PKT_FLAG_KEY;
1618         } else {
1619             pkt.flags = 0;
1620         }
1621         //pkt.duration = 1;
1622         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1623     }
1624     // Encode and add all audio frames up to and including the pts of this video frame.
1625     for ( ;; ) {
1626         int64_t audio_pts;
1627         vector<float> audio;
1628         {
1629              unique_lock<mutex> lock(frame_queue_mutex);
1630              frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || !pending_audio_frames.empty(); });
1631              if (copy_thread_should_quit) return;
1632              auto it = pending_audio_frames.begin();
1633              if (it->first > task.pts) break;
1634              audio_pts = it->first;
1635              audio = move(it->second);
1636              pending_audio_frames.erase(it);
1637         }
1638
1639         AVFrame *frame = avcodec_alloc_frame();
1640         frame->nb_samples = audio.size() / 2;
1641         frame->format = AV_SAMPLE_FMT_S32;
1642         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1643
1644         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1645         int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
1646         if (ret < 0) {
1647             fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1648             exit(1);
1649         }
1650         for (int i = 0; i < frame->nb_samples * 2; ++i) {
1651             if (audio[i] >= 1.0f) {
1652                 int_samples[i] = 2147483647;
1653             } else if (audio[i] <= -1.0f) {
1654                 int_samples[i] = -2147483647;
1655             } else {
1656                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1657             }
1658         }
1659
1660         AVPacket pkt;
1661         av_init_packet(&pkt);
1662         pkt.data = nullptr;
1663         pkt.size = 0;
1664         int got_output;
1665         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1666         if (got_output) {
1667             pkt.stream_index = 1;
1668             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1669         }
1670         // TODO: Delayed frames.
1671         avcodec_free_frame(&frame);
1672         av_free_packet(&pkt);
1673         if (audio_pts == task.pts) break;
1674     }
1675
1676 #if 0
1677     printf("\r      "); /* return back to startpoint */
1678     switch (encode_order % 4) {
1679         case 0:
1680             printf("|");
1681             break;
1682         case 1:
1683             printf("/");
1684             break;
1685         case 2:
1686             printf("-");
1687             break;
1688         case 3:
1689             printf("\\");
1690             break;
1691     }
1692     printf("%08lld", encode_order);
1693     printf("(%06d bytes coded)", coded_size);
1694 #endif
1695 }
1696
1697
1698 // this is weird. but it seems to put a new frame onto the queue
1699 void H264Encoder::storage_task_enqueue(storage_task task)
1700 {
1701         unique_lock<mutex> lock(storage_task_queue_mutex);
1702         storage_task_queue.push(move(task));
1703         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1704         storage_task_queue_changed.notify_all();
1705 }
1706
1707 void H264Encoder::storage_task_thread()
1708 {
1709         for ( ;; ) {
1710                 storage_task current;
1711                 {
1712                         // wait until there's an encoded frame
1713                         unique_lock<mutex> lock(storage_task_queue_mutex);
1714                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1715                         if (storage_thread_should_quit) return;
1716                         current = move(storage_task_queue.front());
1717                         storage_task_queue.pop();
1718                 }
1719
1720                 VAStatus va_status;
1721
1722                 // waits for data, then saves it to disk.
1723                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1724                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1725                 save_codeddata(move(current));
1726
1727                 {
1728                         unique_lock<mutex> lock(storage_task_queue_mutex);
1729                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1730                         storage_task_queue_changed.notify_all();
1731                 }
1732         }
1733 }
1734
1735 static int release_encode()
1736 {
1737     int i;
1738
1739     for (i = 0; i < SURFACE_NUM; i++) {
1740         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1741         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1742         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1743     }
1744
1745     vaDestroyContext(va_dpy, context_id);
1746     vaDestroyConfig(va_dpy, config_id);
1747
1748     return 0;
1749 }
1750
1751 static int deinit_va()
1752 {
1753     vaTerminate(va_dpy);
1754
1755     va_close_display(va_dpy);
1756
1757     return 0;
1758 }
1759
1760
1761 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
1762         : current_storage_frame(0), surface(surface), httpd(httpd)
1763 {
1764         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1765         context_audio = avcodec_alloc_context3(codec_audio);
1766         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1767         context_audio->sample_rate = OUTPUT_FREQUENCY;
1768         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1769         context_audio->channels = 2;
1770         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1771         context_audio->time_base = AVRational{1, TIMEBASE};
1772         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1773                 fprintf(stderr, "Could not open codec\n");
1774                 exit(1);
1775         }
1776
1777         frame_width = width;
1778         frame_height = height;
1779         frame_width_mbaligned = (frame_width + 15) & (~15);
1780         frame_height_mbaligned = (frame_height + 15) & (~15);
1781         frame_bitrate = 15000000;  // / 60;
1782
1783         //print_input();
1784
1785         init_va();
1786         setup_encode();
1787
1788         // No frames are ready yet.
1789         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1790
1791         memset(&seq_param, 0, sizeof(seq_param));
1792         memset(&pic_param, 0, sizeof(pic_param));
1793         memset(&slice_param, 0, sizeof(slice_param));
1794
1795         storage_thread = thread(&H264Encoder::storage_task_thread, this);
1796
1797         copy_thread = thread([this]{
1798                 //SDL_GL_MakeCurrent(window, context);
1799                 QOpenGLContext *context = create_context(this->surface);
1800                 eglBindAPI(EGL_OPENGL_API);
1801                 if (!make_current(context, this->surface)) {
1802                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1803                                 eglGetError());
1804                         exit(1);
1805                 }
1806                 copy_thread_func();
1807         });
1808 }
1809
1810 H264Encoder::~H264Encoder()
1811 {
1812         {
1813                 unique_lock<mutex> lock(storage_task_queue_mutex);
1814                 storage_thread_should_quit = true;
1815                 storage_task_queue_changed.notify_all();
1816         }
1817         {
1818                 unique_lock<mutex> lock(frame_queue_mutex);
1819                 copy_thread_should_quit = true;
1820                 frame_queue_nonempty.notify_all();
1821         }
1822         storage_thread.join();
1823         copy_thread.join();
1824
1825         release_encode();
1826         deinit_va();
1827 }
1828
1829 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1830 {
1831         {
1832                 // Wait until this frame slot is done encoding.
1833                 unique_lock<mutex> lock(storage_task_queue_mutex);
1834                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1835                 if (storage_thread_should_quit) return false;
1836         }
1837
1838         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1839         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1840         *y_tex = surf->y_tex;
1841         *cbcr_tex = surf->cbcr_tex;
1842
1843         VASurfaceID surface = surf->src_surface;
1844         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1845         CHECK_VASTATUS(va_status, "vaDeriveImage");
1846
1847         VABufferInfo buf_info;
1848         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1849         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1850         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1851
1852         // Create Y image.
1853         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1854         EGLint y_attribs[] = {
1855                 EGL_WIDTH, frame_width,
1856                 EGL_HEIGHT, frame_height,
1857                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1858                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1859                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1860                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1861                 EGL_NONE
1862         };
1863
1864         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1865         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1866
1867         // Associate Y image to a texture.
1868         glBindTexture(GL_TEXTURE_2D, *y_tex);
1869         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1870
1871         // Create CbCr image.
1872         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1873         EGLint cbcr_attribs[] = {
1874                 EGL_WIDTH, frame_width,
1875                 EGL_HEIGHT, frame_height,
1876                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1877                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1878                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1879                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1880                 EGL_NONE
1881         };
1882
1883         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1884         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1885
1886         // Associate CbCr image to a texture.
1887         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1888         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1889
1890         return true;
1891 }
1892
1893 void H264Encoder::add_audio(int64_t pts, vector<float> audio)
1894 {
1895         {
1896                 unique_lock<mutex> lock(frame_queue_mutex);
1897                 pending_audio_frames[pts] = move(audio);
1898         }
1899         frame_queue_nonempty.notify_all();
1900 }
1901
1902 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
1903 {
1904         {
1905                 unique_lock<mutex> lock(frame_queue_mutex);
1906                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1907                 ++current_storage_frame;
1908         }
1909         frame_queue_nonempty.notify_all();
1910 }
1911
1912 void H264Encoder::copy_thread_func()
1913 {
1914         int64_t last_dts = -1;
1915         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1916                 PendingFrame frame;
1917                 int pts_lag;
1918                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1919                                        &current_frame_display, &current_frame_type, &pts_lag);
1920                 if (current_frame_type == FRAME_IDR) {
1921                         numShortTerm = 0;
1922                         current_frame_num = 0;
1923                         current_IDR_display = current_frame_display;
1924                 }
1925
1926                 {
1927                         unique_lock<mutex> lock(frame_queue_mutex);
1928                         frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; });
1929                         if (copy_thread_should_quit) {
1930                                 return;
1931                         } else {
1932                                 frame = move(pending_video_frames[current_frame_display]);
1933                                 pending_video_frames.erase(current_frame_display);
1934                         }
1935                 }
1936
1937                 // Determine the dts of this frame.
1938                 int64_t dts;
1939                 if (pts_lag == -1) {
1940                         assert(last_dts != -1);
1941                         dts = last_dts + (TIMEBASE / MAX_FPS);
1942                 } else {
1943                         dts = frame.pts - pts_lag;
1944                 }
1945                 last_dts = dts;
1946
1947                 encode_frame(frame, encoding_frame_num, frame.pts, dts);
1948         }
1949 }
1950
1951 void H264Encoder::encode_frame(H264Encoder::PendingFrame frame, int encoding_frame_num, int64_t pts, int64_t dts)
1952 {
1953         // Wait for the GPU to be done with the frame.
1954         glClientWaitSync(frame.fence.get(), 0, 0);
1955
1956         // Release back any input frames we needed to render this frame.
1957         frame.input_frames.clear();
1958
1959         // Unmap the image.
1960         GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
1961         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1962         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1963         VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1964         CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1965         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1966         CHECK_VASTATUS(va_status, "vaDestroyImage");
1967
1968         VASurfaceID surface = surf->src_surface;
1969
1970         // Schedule the frame for encoding.
1971         va_status = vaBeginPicture(va_dpy, context_id, surface);
1972         CHECK_VASTATUS(va_status, "vaBeginPicture");
1973
1974         if (current_frame_type == FRAME_IDR) {
1975                 render_sequence();
1976                 render_picture();
1977                 if (h264_packedheader) {
1978                         render_packedsequence();
1979                         render_packedpicture();
1980                 }
1981         } else {
1982                 //render_sequence();
1983                 render_picture();
1984         }
1985         render_slice(encoding_frame_num);
1986
1987         va_status = vaEndPicture(va_dpy, context_id);
1988         CHECK_VASTATUS(va_status, "vaEndPicture");
1989
1990         // so now the data is done encoding (well, async job kicked off)...
1991         // we send that to the storage thread
1992         storage_task tmp;
1993         tmp.display_order = current_frame_display;
1994         tmp.frame_type = current_frame_type;
1995         tmp.pts = pts;
1996         tmp.dts = dts;
1997         storage_task_enqueue(move(tmp));
1998
1999         update_ReferenceFrames();
2000 }