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

   1 //#include "sysdeps.h"
   2 #include "h264encode.h"
   3
   4 #include <movit/util.h>
   5 #include <EGL/eglplatform.h>
   6 #include <X11/X.h>
   7 #include <X11/Xlib.h>
   8 #include <assert.h>
   9 #include <epoxy/egl.h>
  10 extern "C" {
  11 #include <libavcodec/avcodec.h>
  12 #include <libavformat/avformat.h>
  13 #include <libavutil/channel_layout.h>
  14 #include <libavutil/frame.h>
  15 #include <libavutil/rational.h>
  16 #include <libavutil/samplefmt.h>
  17 }
  18 #include <libdrm/drm_fourcc.h>
  19 #include <stdio.h>
  20 #include <stdlib.h>
  21 #include <string.h>
  22 #include <fcntl.h>
  23 #include <va/va.h>
  24 #include <va/va_drm.h>
  25 #include <va/va_drmcommon.h>
  26 #include <va/va_enc_h264.h>
  27 #include <va/va_x11.h>
  28 #include <algorithm>
  29 #include <condition_variable>
  30 #include <cstdint>
  31 #include <map>
  32 #include <memory>
  33 #include <mutex>
  34 #include <queue>
  35 #include <string>
  36 #include <thread>
  37 #include <utility>
  38
  39 #include "context.h"
  40 #include "defs.h"
  41 #include "httpd.h"
  42 #include "timebase.h"
  43
  44 using namespace std;
  45
  46 class QOpenGLContext;
  47 class QSurface;
  48
  49 #define CHECK_VASTATUS(va_status, func)                                 \
  50     if (va_status != VA_STATUS_SUCCESS) {                               \
  51         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  52         exit(1);                                                        \
  53     }
  54
  55 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  56
  57 //#include "loadsurface.h"
  58
  59 #define NAL_REF_IDC_NONE        0
  60 #define NAL_REF_IDC_LOW         1
  61 #define NAL_REF_IDC_MEDIUM      2
  62 #define NAL_REF_IDC_HIGH        3
  63
  64 #define NAL_NON_IDR             1
  65 #define NAL_IDR                 5
  66 #define NAL_SPS                 7
  67 #define NAL_PPS                 8
  68 #define NAL_SEI                 6
  69
  70 #define SLICE_TYPE_P            0
  71 #define SLICE_TYPE_B            1
  72 #define SLICE_TYPE_I            2
  73 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  74 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  75 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  76
  77
  78 #define ENTROPY_MODE_CAVLC      0
  79 #define ENTROPY_MODE_CABAC      1
  80
  81 #define PROFILE_IDC_BASELINE    66
  82 #define PROFILE_IDC_MAIN        77
  83 #define PROFILE_IDC_HIGH        100
  84
  85 #define BITSTREAM_ALLOCATE_STEPPING     4096
  86 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  87
  88 static constexpr unsigned int MaxFrameNum = (2<<16);
  89 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
  90 static constexpr unsigned int Log2MaxFrameNum = 16;
  91 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
  92 static constexpr int rc_default_modes[] = {  // Priority list of modes.
  93     VA_RC_VBR,
  94     VA_RC_CQP,
  95     VA_RC_VBR_CONSTRAINED,
  96     VA_RC_CBR,
  97     VA_RC_VCM,
  98     VA_RC_NONE,
  99 };
 100
 101 /* thread to save coded data */
 102 #define SRC_SURFACE_FREE        0
 103 #define SRC_SURFACE_IN_ENCODING 1
 104
 105 struct __bitstream {
 106     unsigned int *buffer;
 107     int bit_offset;
 108     int max_size_in_dword;
 109 };
 110 typedef struct __bitstream bitstream;
 111
 112 using namespace std;
 113
 114 class H264EncoderImpl {
 115 public:
 116         H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd);
 117         ~H264EncoderImpl();
 118         void add_audio(int64_t pts, vector<float> audio);
 119         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 120         void end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames);
 121         void shutdown();
 122
 123 private:
 124         struct storage_task {
 125                 unsigned long long display_order;
 126                 int frame_type;
 127                 vector<float> audio;
 128                 int64_t pts, dts;
 129         };
 130         struct PendingFrame {
 131                 RefCountedGLsync fence;
 132                 vector<RefCountedFrame> input_frames;
 133                 int64_t pts;
 134         };
 135
 136         void encode_thread_func();
 137         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 138         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 139                           int frame_type, int64_t pts, int64_t dts);
 140         void storage_task_thread();
 141         void storage_task_enqueue(storage_task task);
 142         void save_codeddata(storage_task task);
 143         int render_packedsequence();
 144         int render_packedpicture();
 145         void render_packedslice();
 146         int render_sequence();
 147         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 148         void sps_rbsp(bitstream *bs);
 149         void pps_rbsp(bitstream *bs);
 150         int build_packed_pic_buffer(unsigned char **header_buffer);
 151         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 152         void slice_header(bitstream *bs);
 153         int build_packed_seq_buffer(unsigned char **header_buffer);
 154         int build_packed_slice_buffer(unsigned char **header_buffer);
 155         int init_va(const string &va_display);
 156         int deinit_va();
 157         VADisplay va_open_display(const string &va_display);
 158         void va_close_display(VADisplay va_dpy);
 159         int setup_encode();
 160         int release_encode();
 161         void update_ReferenceFrames(int frame_type);
 162         int update_RefPicList(int frame_type);
 163
 164         bool is_shutdown = false;
 165         bool use_zerocopy;
 166         int drm_fd = -1;
 167
 168         thread encode_thread, storage_thread;
 169
 170         mutex storage_task_queue_mutex;
 171         condition_variable storage_task_queue_changed;
 172         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 173         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 174         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 175
 176         mutex frame_queue_mutex;
 177         condition_variable frame_queue_nonempty;
 178         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 179
 180         int current_storage_frame;
 181
 182         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 183         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
 184         QSurface *surface;
 185
 186         AVCodecContext *context_audio;
 187         HTTPD *httpd;
 188
 189         Display *x11_display = nullptr;
 190
 191         // Encoder parameters
 192         VADisplay va_dpy;
 193         VAProfile h264_profile = (VAProfile)~0;
 194         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 195         int config_attrib_num = 0, enc_packed_header_idx;
 196
 197         struct GLSurface {
 198                 VASurfaceID src_surface, ref_surface;
 199                 VABufferID coded_buf;
 200
 201                 VAImage surface_image;
 202                 GLuint y_tex, cbcr_tex;
 203
 204                 // Only if use_zerocopy == true.
 205                 EGLImage y_egl_image, cbcr_egl_image;
 206
 207                 // Only if use_zerocopy == false.
 208                 GLuint pbo;
 209                 uint8_t *y_ptr, *cbcr_ptr;
 210                 size_t y_offset, cbcr_offset;
 211         };
 212         GLSurface gl_surfaces[SURFACE_NUM];
 213
 214         VAConfigID config_id;
 215         VAContextID context_id;
 216         VAEncSequenceParameterBufferH264 seq_param;
 217         VAEncPictureParameterBufferH264 pic_param;
 218         VAEncSliceParameterBufferH264 slice_param;
 219         VAPictureH264 CurrentCurrPic;
 220         VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
 221
 222         // Static quality settings.
 223         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 224         static constexpr unsigned int num_ref_frames = 2;
 225         static constexpr int initial_qp = 15;
 226         static constexpr int minimal_qp = 0;
 227         static constexpr int intra_period = 30;
 228         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 229
 230         // Quality settings that are meant to be static, but might be overridden
 231         // by the profile.
 232         int constraint_set_flag = 0;
 233         int h264_packedheader = 0; /* support pack header? */
 234         int h264_maxref = (1<<16|1);
 235         int h264_entropy_mode = 1; /* cabac */
 236         int ip_period = 3;
 237
 238         int rc_mode = -1;
 239         unsigned int current_frame_num = 0;
 240         unsigned int numShortTerm = 0;
 241
 242         int frame_width;
 243         int frame_height;
 244         int frame_width_mbaligned;
 245         int frame_height_mbaligned;
 246 };
 247
 248 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 249 // but if we don't delete it here, we get leaks. The GStreamer implementation
 250 // does the same.
 251 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 252 {
 253     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 254     CHECK_VASTATUS(va_status, "vaRenderPicture");
 255
 256     for (int i = 0; i < num_buffers; ++i) {
 257         va_status = vaDestroyBuffer(dpy, buffers[i]);
 258         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 259     }
 260 }
 261
 262 static unsigned int
 263 va_swap32(unsigned int val)
 264 {
 265     unsigned char *pval = (unsigned char *)&val;
 266
 267     return ((pval[0] << 24)     |
 268             (pval[1] << 16)     |
 269             (pval[2] << 8)      |
 270             (pval[3] << 0));
 271 }
 272
 273 static void
 274 bitstream_start(bitstream *bs)
 275 {
 276     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 277     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 278     bs->bit_offset = 0;
 279 }
 280
 281 static void
 282 bitstream_end(bitstream *bs)
 283 {
 284     int pos = (bs->bit_offset >> 5);
 285     int bit_offset = (bs->bit_offset & 0x1f);
 286     int bit_left = 32 - bit_offset;
 287
 288     if (bit_offset) {
 289         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 290     }
 291 }
 292
 293 static void
 294 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 295 {
 296     int pos = (bs->bit_offset >> 5);
 297     int bit_offset = (bs->bit_offset & 0x1f);
 298     int bit_left = 32 - bit_offset;
 299
 300     if (!size_in_bits)
 301         return;
 302
 303     bs->bit_offset += size_in_bits;
 304
 305     if (bit_left > size_in_bits) {
 306         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 307     } else {
 308         size_in_bits -= bit_left;
 309         if (bit_left >= 32) {
 310             bs->buffer[pos] = (val >> size_in_bits);
 311         } else {
 312             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 313         }
 314         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 315
 316         if (pos + 1 == bs->max_size_in_dword) {
 317             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 318             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 319         }
 320
 321         bs->buffer[pos + 1] = val;
 322     }
 323 }
 324
 325 static void
 326 bitstream_put_ue(bitstream *bs, unsigned int val)
 327 {
 328     int size_in_bits = 0;
 329     int tmp_val = ++val;
 330
 331     while (tmp_val) {
 332         tmp_val >>= 1;
 333         size_in_bits++;
 334     }
 335
 336     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 337     bitstream_put_ui(bs, val, size_in_bits);
 338 }
 339
 340 static void
 341 bitstream_put_se(bitstream *bs, int val)
 342 {
 343     unsigned int new_val;
 344
 345     if (val <= 0)
 346         new_val = -2 * val;
 347     else
 348         new_val = 2 * val - 1;
 349
 350     bitstream_put_ue(bs, new_val);
 351 }
 352
 353 static void
 354 bitstream_byte_aligning(bitstream *bs, int bit)
 355 {
 356     int bit_offset = (bs->bit_offset & 0x7);
 357     int bit_left = 8 - bit_offset;
 358     int new_val;
 359
 360     if (!bit_offset)
 361         return;
 362
 363     assert(bit == 0 || bit == 1);
 364
 365     if (bit)
 366         new_val = (1 << bit_left) - 1;
 367     else
 368         new_val = 0;
 369
 370     bitstream_put_ui(bs, new_val, bit_left);
 371 }
 372
 373 static void
 374 rbsp_trailing_bits(bitstream *bs)
 375 {
 376     bitstream_put_ui(bs, 1, 1);
 377     bitstream_byte_aligning(bs, 0);
 378 }
 379
 380 static void nal_start_code_prefix(bitstream *bs)
 381 {
 382     bitstream_put_ui(bs, 0x00000001, 32);
 383 }
 384
 385 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 386 {
 387     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 388     bitstream_put_ui(bs, nal_ref_idc, 2);
 389     bitstream_put_ui(bs, nal_unit_type, 5);
 390 }
 391
 392 void H264EncoderImpl::sps_rbsp(bitstream *bs)
 393 {
 394     int profile_idc = PROFILE_IDC_BASELINE;
 395
 396     if (h264_profile  == VAProfileH264High)
 397         profile_idc = PROFILE_IDC_HIGH;
 398     else if (h264_profile  == VAProfileH264Main)
 399         profile_idc = PROFILE_IDC_MAIN;
 400
 401     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 402     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 403     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 404     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 405     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 406     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 407     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 408     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 409
 410     if ( profile_idc == PROFILE_IDC_HIGH) {
 411         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 412         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 413         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 414         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 415         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 416     }
 417
 418     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 419     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 420
 421     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 422         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 423     else {
 424         assert(0);
 425     }
 426
 427     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 428     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 429
 430     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 431     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 432     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 433
 434     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 435         assert(0);
 436     }
 437
 438     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 439     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 440
 441     if (seq_param.frame_cropping_flag) {
 442         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 443         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 444         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 445         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 446     }
 447
 448     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 449     if ( false ) {
 450         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 451     } else {
 452         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 453         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 454         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 455         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 456         {
 457             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 458             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 459             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 460             {
 461                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 462                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 463                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 464             }
 465         }
 466         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 467         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 468         {
 469             bitstream_put_ui(bs, 1, 32);  // FPS
 470             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 471             bitstream_put_ui(bs, 1, 1);
 472         }
 473         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 474         {
 475             // hrd_parameters
 476             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 477             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 478             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 479
 480             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 481             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 482             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 483
 484             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 485             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 486             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 487             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 488         }
 489         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 490         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 491
 492         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 493         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 494     }
 495
 496     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 497 }
 498
 499
 500 void H264EncoderImpl::pps_rbsp(bitstream *bs)
 501 {
 502     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 503     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 504
 505     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 506
 507     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 508
 509     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 510
 511     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 512     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 513
 514     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 515     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 516
 517     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 518     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 519     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 520
 521     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 522     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 523     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 524
 525     /* more_rbsp_data */
 526     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 527     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 528     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 529
 530     rbsp_trailing_bits(bs);
 531 }
 532
 533 void H264EncoderImpl::slice_header(bitstream *bs)
 534 {
 535     int first_mb_in_slice = slice_param.macroblock_address;
 536
 537     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 538     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 539     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 540     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 541
 542     /* frame_mbs_only_flag == 1 */
 543     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 544         /* FIXME: */
 545         assert(0);
 546     }
 547
 548     if (pic_param.pic_fields.bits.idr_pic_flag)
 549         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 550
 551     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 552         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 553         /* pic_order_present_flag == 0 */
 554     } else {
 555         /* FIXME: */
 556         assert(0);
 557     }
 558
 559     /* redundant_pic_cnt_present_flag == 0 */
 560     /* slice type */
 561     if (IS_P_SLICE(slice_param.slice_type)) {
 562         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 563
 564         if (slice_param.num_ref_idx_active_override_flag)
 565             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 566
 567         /* ref_pic_list_reordering */
 568         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 569     } else if (IS_B_SLICE(slice_param.slice_type)) {
 570         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 571
 572         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 573
 574         if (slice_param.num_ref_idx_active_override_flag) {
 575             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 576             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 577         }
 578
 579         /* ref_pic_list_reordering */
 580         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 581         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 582     }
 583
 584     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 585          IS_P_SLICE(slice_param.slice_type)) ||
 586         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 587          IS_B_SLICE(slice_param.slice_type))) {
 588         /* FIXME: fill weight/offset table */
 589         assert(0);
 590     }
 591
 592     /* dec_ref_pic_marking */
 593     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 594         unsigned char no_output_of_prior_pics_flag = 0;
 595         unsigned char long_term_reference_flag = 0;
 596         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 597
 598         if (pic_param.pic_fields.bits.idr_pic_flag) {
 599             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 600             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 601         } else {
 602             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 603         }
 604     }
 605
 606     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 607         !IS_I_SLICE(slice_param.slice_type))
 608         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 609
 610     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 611
 612     /* ignore for SP/SI */
 613
 614     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 615         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 616
 617         if (slice_param.disable_deblocking_filter_idc != 1) {
 618             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 619             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 620         }
 621     }
 622
 623     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 624         bitstream_byte_aligning(bs, 1);
 625     }
 626 }
 627
 628 int H264EncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 629 {
 630     bitstream bs;
 631
 632     bitstream_start(&bs);
 633     nal_start_code_prefix(&bs);
 634     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 635     pps_rbsp(&bs);
 636     bitstream_end(&bs);
 637
 638     *header_buffer = (unsigned char *)bs.buffer;
 639     return bs.bit_offset;
 640 }
 641
 642 int
 643 H264EncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 644 {
 645     bitstream bs;
 646
 647     bitstream_start(&bs);
 648     nal_start_code_prefix(&bs);
 649     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 650     sps_rbsp(&bs);
 651     bitstream_end(&bs);
 652
 653     *header_buffer = (unsigned char *)bs.buffer;
 654     return bs.bit_offset;
 655 }
 656
 657 int H264EncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 658 {
 659     bitstream bs;
 660     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 661     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 662
 663     bitstream_start(&bs);
 664     nal_start_code_prefix(&bs);
 665
 666     if (IS_I_SLICE(slice_param.slice_type)) {
 667         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 668     } else if (IS_P_SLICE(slice_param.slice_type)) {
 669         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 670     } else {
 671         assert(IS_B_SLICE(slice_param.slice_type));
 672         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 673     }
 674
 675     slice_header(&bs);
 676     bitstream_end(&bs);
 677
 678     *header_buffer = (unsigned char *)bs.buffer;
 679     return bs.bit_offset;
 680 }
 681
 682
 683 /*
 684   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 685   1) period between Frame #X and Frame #N = #X - #N
 686   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 687   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 688   4) intra_period and intra_idr_period take precedence over ip_period
 689   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 690      of I/IDR frames, see bellow examples
 691   -------------------------------------------------------------------
 692   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 693   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 694   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 695   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 696   1            1                ignored    IDR IDR IDR IDR...
 697   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 698   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 699   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 700                                               (PBB)(IBB)(PBB)(IBB)...
 701   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 702                                            IDRPPPPP IPPPPP IPPPPP...
 703   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 704                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 705                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 706                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 707                                            {IDR(PBB)(PBB)}                     (6/6/3)
 708                                            {IDR(PBB)(PBB)}.
 709 */
 710
 711 // General pts/dts strategy:
 712 //
 713 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 714 // bit tricky. We assume first of all that the frame rate never goes _above_
 715 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 716 // in at least this speed, as long at dts <= pts (the frame is not attempted
 717 // presented before it is decoded). Furthermore, we never have longer chains of
 718 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 719 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 720 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 721 // frame, with an order O <= C.)
 722 //
 723 // Many strategies are possible, but we establish these rules:
 724 //
 725 //  - Tip frames have dts = pts - (C-O)*N.
 726 //  - Non-tip frames have dts = dts_last + N.
 727 //
 728 // An example, with C=2 and N=10 and the data flow showed with arrows:
 729 //
 730 //        I  B  P  B  B  P
 731 //   pts: 30 40 50 60 70 80
 732 //        ↓  ↓     ↓
 733 //   dts: 10 30 20 60 50←40
 734 //         |  |  ↑        ↑
 735 //         `--|--'        |
 736 //             `----------'
 737 //
 738 // To show that this works fine also with irregular spacings, let's say that
 739 // the third frame is delayed a bit (something earlier was dropped). Now the
 740 // situation looks like this:
 741 //
 742 //        I  B  P  B  B   P
 743 //   pts: 30 40 80 90 100 110
 744 //        ↓  ↓     ↓
 745 //   dts: 10 30 20 90 50←40
 746 //         |  |  ↑        ↑
 747 //         `--|--'        |
 748 //             `----------'
 749 //
 750 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 751 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 752 //
 753 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 754 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 755 // a tip frame and should be given a dts based on the previous one.
 756 #define FRAME_P 0
 757 #define FRAME_B 1
 758 #define FRAME_I 2
 759 #define FRAME_IDR 7
 760 void encoding2display_order(
 761     int encoding_order, int intra_period,
 762     int intra_idr_period, int ip_period,
 763     int *displaying_order,
 764     int *frame_type, int *pts_lag)
 765 {
 766     int encoding_order_gop = 0;
 767
 768     *pts_lag = 0;
 769
 770     if (intra_period == 1) { /* all are I/IDR frames */
 771         *displaying_order = encoding_order;
 772         if (intra_idr_period == 0)
 773             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 774         else
 775             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 776         return;
 777     }
 778
 779     if (intra_period == 0)
 780         intra_idr_period = 0;
 781
 782     if (ip_period == 1) {
 783         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 784         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 785         *displaying_order = encoding_order;
 786
 787         if (encoding_order_gop == 0) { /* the first frame */
 788             *frame_type = FRAME_IDR;
 789         } else if (intra_period != 0 && /* have I frames */
 790                    encoding_order_gop >= 2 &&
 791                    (encoding_order_gop % intra_period == 0)) {
 792             *frame_type = FRAME_I;
 793         } else {
 794             *frame_type = FRAME_P;
 795         }
 796         return;
 797     }
 798
 799     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 800     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 801     *pts_lag = -1;  // Most frames are not tip frames.
 802
 803     if (encoding_order_gop == 0) { /* the first frame */
 804         *frame_type = FRAME_IDR;
 805         *displaying_order = encoding_order;
 806         // IDR frames are a special case; I honestly can't find the logic behind
 807         // why this is the right thing, but it seems to line up nicely in practice :-)
 808         *pts_lag = TIMEBASE / MAX_FPS;
 809     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 810         *frame_type = FRAME_B;
 811         *displaying_order = encoding_order - 1;
 812         if ((encoding_order_gop % ip_period) == 0) {
 813             *pts_lag = 0;  // Last B-frame.
 814         }
 815     } else if (intra_period != 0 && /* have I frames */
 816                encoding_order_gop >= 2 &&
 817                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 818         *frame_type = FRAME_I;
 819         *displaying_order = encoding_order + ip_period - 1;
 820     } else {
 821         *frame_type = FRAME_P;
 822         *displaying_order = encoding_order + ip_period - 1;
 823     }
 824 }
 825
 826
 827 static const char *rc_to_string(int rc_mode)
 828 {
 829     switch (rc_mode) {
 830     case VA_RC_NONE:
 831         return "NONE";
 832     case VA_RC_CBR:
 833         return "CBR";
 834     case VA_RC_VBR:
 835         return "VBR";
 836     case VA_RC_VCM:
 837         return "VCM";
 838     case VA_RC_CQP:
 839         return "CQP";
 840     case VA_RC_VBR_CONSTRAINED:
 841         return "VBR_CONSTRAINED";
 842     default:
 843         return "Unknown";
 844     }
 845 }
 846
 847 VADisplay H264EncoderImpl::va_open_display(const string &va_display)
 848 {
 849         if (va_display.empty()) {
 850                 x11_display = XOpenDisplay(NULL);
 851                 if (!x11_display) {
 852                         fprintf(stderr, "error: can't connect to X server!\n");
 853                         return NULL;
 854                 }
 855                 use_zerocopy = true;
 856                 return vaGetDisplay(x11_display);
 857         } else if (va_display[0] != '/') {
 858                 x11_display = XOpenDisplay(va_display.c_str());
 859                 if (!x11_display) {
 860                         fprintf(stderr, "error: can't connect to X server!\n");
 861                         return NULL;
 862                 }
 863                 use_zerocopy = true;
 864                 return vaGetDisplay(x11_display);
 865         } else {
 866                 drm_fd = open(va_display.c_str(), O_RDWR);
 867                 if (drm_fd == -1) {
 868                         perror(va_display.c_str());
 869                         return NULL;
 870                 }
 871                 use_zerocopy = false;
 872                 return vaGetDisplayDRM(drm_fd);
 873         }
 874 }
 875
 876 void H264EncoderImpl::va_close_display(VADisplay va_dpy)
 877 {
 878         if (x11_display) {
 879                 XCloseDisplay(x11_display);
 880                 x11_display = nullptr;
 881         }
 882         if (drm_fd != -1) {
 883                 close(drm_fd);
 884         }
 885 }
 886
 887 int H264EncoderImpl::init_va(const string &va_display)
 888 {
 889     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 890     VAEntrypoint *entrypoints;
 891     int num_entrypoints, slice_entrypoint;
 892     int support_encode = 0;
 893     int major_ver, minor_ver;
 894     VAStatus va_status;
 895     unsigned int i;
 896
 897     va_dpy = va_open_display(va_display);
 898     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
 899     CHECK_VASTATUS(va_status, "vaInitialize");
 900
 901     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
 902     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
 903     if (!entrypoints) {
 904         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
 905         exit(1);
 906     }
 907
 908     /* use the highest profile */
 909     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
 910         if ((h264_profile != ~0) && h264_profile != profile_list[i])
 911             continue;
 912
 913         h264_profile = profile_list[i];
 914         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
 915         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 916             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
 917                 support_encode = 1;
 918                 break;
 919             }
 920         }
 921         if (support_encode == 1)
 922             break;
 923     }
 924
 925     if (support_encode == 0) {
 926         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
 927         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/card0\n");
 928         printf("to use VA-API against DRM instead of X11.\n");
 929         exit(1);
 930     } else {
 931         switch (h264_profile) {
 932             case VAProfileH264Baseline:
 933                 ip_period = 1;
 934                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 935                 h264_entropy_mode = 0;
 936                 break;
 937             case VAProfileH264ConstrainedBaseline:
 938                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
 939                 ip_period = 1;
 940                 break;
 941
 942             case VAProfileH264Main:
 943                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
 944                 break;
 945
 946             case VAProfileH264High:
 947                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
 948                 break;
 949             default:
 950                 h264_profile = VAProfileH264Baseline;
 951                 ip_period = 1;
 952                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
 953                 break;
 954         }
 955     }
 956
 957     VAConfigAttrib attrib[VAConfigAttribTypeMax];
 958
 959     /* find out the format for the render target, and rate control mode */
 960     for (i = 0; i < VAConfigAttribTypeMax; i++)
 961         attrib[i].type = (VAConfigAttribType)i;
 962
 963     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
 964                                       &attrib[0], VAConfigAttribTypeMax);
 965     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
 966     /* check the interested configattrib */
 967     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
 968         printf("Not find desired YUV420 RT format\n");
 969         exit(1);
 970     } else {
 971         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
 972         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
 973         config_attrib_num++;
 974     }
 975
 976     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
 977         int tmp = attrib[VAConfigAttribRateControl].value;
 978
 979         if (rc_mode == -1 || !(rc_mode & tmp))  {
 980             if (rc_mode != -1) {
 981                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
 982             }
 983
 984             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
 985                 if (rc_default_modes[i] & tmp) {
 986                     rc_mode = rc_default_modes[i];
 987                     break;
 988                 }
 989             }
 990         }
 991
 992         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
 993         config_attrib[config_attrib_num].value = rc_mode;
 994         config_attrib_num++;
 995     }
 996
 997
 998     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
 999         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1000
1001         h264_packedheader = 1;
1002         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1003         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1004
1005         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1006             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1007         }
1008
1009         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1010             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1011         }
1012
1013         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1014             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1015         }
1016
1017         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1018             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1019         }
1020
1021         enc_packed_header_idx = config_attrib_num;
1022         config_attrib_num++;
1023     }
1024
1025     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1026         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1027         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1028         config_attrib_num++;
1029     }
1030
1031     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1032         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1033     }
1034
1035     free(entrypoints);
1036     return 0;
1037 }
1038
1039 int H264EncoderImpl::setup_encode()
1040 {
1041     VAStatus va_status;
1042     VASurfaceID *tmp_surfaceid;
1043     int codedbuf_size, i;
1044     static VASurfaceID src_surface[SURFACE_NUM];
1045     static VASurfaceID ref_surface[SURFACE_NUM];
1046
1047     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1048             &config_attrib[0], config_attrib_num, &config_id);
1049     CHECK_VASTATUS(va_status, "vaCreateConfig");
1050
1051     /* create source surfaces */
1052     va_status = vaCreateSurfaces(va_dpy,
1053                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1054                                  &src_surface[0], SURFACE_NUM,
1055                                  NULL, 0);
1056     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1057
1058     /* create reference surfaces */
1059     va_status = vaCreateSurfaces(va_dpy,
1060                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1061                                  &ref_surface[0], SURFACE_NUM,
1062                                  NULL, 0);
1063     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1064
1065     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1066     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1067     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1068
1069     /* Create a context for this encode pipe */
1070     va_status = vaCreateContext(va_dpy, config_id,
1071                                 frame_width_mbaligned, frame_height_mbaligned,
1072                                 VA_PROGRESSIVE,
1073                                 tmp_surfaceid, 2 * SURFACE_NUM,
1074                                 &context_id);
1075     CHECK_VASTATUS(va_status, "vaCreateContext");
1076     free(tmp_surfaceid);
1077
1078     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1079
1080     for (i = 0; i < SURFACE_NUM; i++) {
1081         /* create coded buffer once for all
1082          * other VA buffers which won't be used again after vaRenderPicture.
1083          * so APP can always vaCreateBuffer for every frame
1084          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1085          * so VA won't maintain the coded buffer
1086          */
1087         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1088                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1089         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1090     }
1091
1092     /* create OpenGL objects */
1093     //glGenFramebuffers(SURFACE_NUM, fbos);
1094
1095     for (i = 0; i < SURFACE_NUM; i++) {
1096         glGenTextures(1, &gl_surfaces[i].y_tex);
1097         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1098
1099         if (!use_zerocopy) {
1100             // Create Y image.
1101             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
1102             glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
1103
1104             // Create CbCr image.
1105             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
1106             glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
1107
1108             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1109             // buffers, due to potentially differing pitch.
1110             glGenBuffers(1, &gl_surfaces[i].pbo);
1111             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1112             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1113             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1114             gl_surfaces[i].y_offset = 0;
1115             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1116             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1117             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1118             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1119         }
1120     }
1121
1122     for (i = 0; i < SURFACE_NUM; i++) {
1123         gl_surfaces[i].src_surface = src_surface[i];
1124         gl_surfaces[i].ref_surface = ref_surface[i];
1125     }
1126
1127     return 0;
1128 }
1129
1130 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1131 //
1132 //   2 1 0 [3] 4 8 9
1133 template<class T, class C>
1134 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1135 {
1136         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1137         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1138         sort(middle, end, less_than);
1139 }
1140
1141 void H264EncoderImpl::update_ReferenceFrames(int frame_type)
1142 {
1143     int i;
1144
1145     if (frame_type == FRAME_B)
1146         return;
1147
1148     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1149     numShortTerm++;
1150     if (numShortTerm > num_ref_frames)
1151         numShortTerm = num_ref_frames;
1152     for (i=numShortTerm-1; i>0; i--)
1153         ReferenceFrames[i] = ReferenceFrames[i-1];
1154     ReferenceFrames[0] = CurrentCurrPic;
1155
1156     current_frame_num++;
1157     if (current_frame_num > MaxFrameNum)
1158         current_frame_num = 0;
1159 }
1160
1161
1162 int H264EncoderImpl::update_RefPicList(int frame_type)
1163 {
1164     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1165         return a.frame_idx > b.frame_idx;
1166     };
1167     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1168         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1169     };
1170     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1171         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1172     };
1173
1174     if (frame_type == FRAME_P) {
1175         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1176         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1177     } else if (frame_type == FRAME_B) {
1178         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1179         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1180
1181         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1182         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1183     }
1184
1185     return 0;
1186 }
1187
1188
1189 int H264EncoderImpl::render_sequence()
1190 {
1191     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1192     VAStatus va_status;
1193     VAEncMiscParameterBuffer *misc_param;
1194     VAEncMiscParameterRateControl *misc_rate_ctrl;
1195
1196     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1197     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1198     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1199     seq_param.bits_per_second = frame_bitrate;
1200
1201     seq_param.intra_period = intra_period;
1202     seq_param.intra_idr_period = intra_idr_period;
1203     seq_param.ip_period = ip_period;
1204
1205     seq_param.max_num_ref_frames = num_ref_frames;
1206     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1207     seq_param.time_scale = TIMEBASE * 2;
1208     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1209     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1210     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1211     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1212     seq_param.seq_fields.bits.chroma_format_idc = 1;
1213     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1214
1215     if (frame_width != frame_width_mbaligned ||
1216         frame_height != frame_height_mbaligned) {
1217         seq_param.frame_cropping_flag = 1;
1218         seq_param.frame_crop_left_offset = 0;
1219         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1220         seq_param.frame_crop_top_offset = 0;
1221         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1222     }
1223
1224     va_status = vaCreateBuffer(va_dpy, context_id,
1225                                VAEncSequenceParameterBufferType,
1226                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1227     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1228
1229     va_status = vaCreateBuffer(va_dpy, context_id,
1230                                VAEncMiscParameterBufferType,
1231                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1232                                1, NULL, &rc_param_buf);
1233     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1234
1235     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1236     misc_param->type = VAEncMiscParameterTypeRateControl;
1237     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1238     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1239     misc_rate_ctrl->bits_per_second = frame_bitrate;
1240     misc_rate_ctrl->target_percentage = 66;
1241     misc_rate_ctrl->window_size = 1000;
1242     misc_rate_ctrl->initial_qp = initial_qp;
1243     misc_rate_ctrl->min_qp = minimal_qp;
1244     misc_rate_ctrl->basic_unit_size = 0;
1245     vaUnmapBuffer(va_dpy, rc_param_buf);
1246
1247     render_id[0] = seq_param_buf;
1248     render_id[1] = rc_param_buf;
1249
1250     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1251
1252     return 0;
1253 }
1254
1255 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1256 {
1257     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1258     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1259     int PicOrderCntMsb, TopFieldOrderCnt;
1260
1261     if (frame_type == FRAME_IDR)
1262         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1263     else {
1264         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1265         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1266     }
1267
1268     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1269         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1270         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1271     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1272              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1273         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1274     else
1275         PicOrderCntMsb = prevPicOrderCntMsb;
1276
1277     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1278
1279     if (frame_type != FRAME_B) {
1280         PicOrderCntMsb_ref = PicOrderCntMsb;
1281         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1282     }
1283
1284     return TopFieldOrderCnt;
1285 }
1286
1287 int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1288 {
1289     VABufferID pic_param_buf;
1290     VAStatus va_status;
1291     int i = 0;
1292
1293     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1294     pic_param.CurrPic.frame_idx = current_frame_num;
1295     pic_param.CurrPic.flags = 0;
1296     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1297     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1298     CurrentCurrPic = pic_param.CurrPic;
1299
1300     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1301     for (i = numShortTerm; i < SURFACE_NUM; i++) {
1302         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1303         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1304     }
1305
1306     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1307     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1308     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1309     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1310     pic_param.frame_num = current_frame_num;
1311     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1312     pic_param.last_picture = false;  // FIXME
1313     pic_param.pic_init_qp = initial_qp;
1314
1315     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1316                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1317     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1318
1319     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1320
1321     return 0;
1322 }
1323
1324 int H264EncoderImpl::render_packedsequence()
1325 {
1326     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1327     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1328     unsigned int length_in_bits;
1329     unsigned char *packedseq_buffer = NULL;
1330     VAStatus va_status;
1331
1332     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1333
1334     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1335
1336     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1337     packedheader_param_buffer.has_emulation_bytes = 0;
1338     va_status = vaCreateBuffer(va_dpy,
1339                                context_id,
1340                                VAEncPackedHeaderParameterBufferType,
1341                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1342                                &packedseq_para_bufid);
1343     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1344
1345     va_status = vaCreateBuffer(va_dpy,
1346                                context_id,
1347                                VAEncPackedHeaderDataBufferType,
1348                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1349                                &packedseq_data_bufid);
1350     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1351
1352     render_id[0] = packedseq_para_bufid;
1353     render_id[1] = packedseq_data_bufid;
1354     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1355
1356     free(packedseq_buffer);
1357
1358     return 0;
1359 }
1360
1361
1362 int H264EncoderImpl::render_packedpicture()
1363 {
1364     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1365     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1366     unsigned int length_in_bits;
1367     unsigned char *packedpic_buffer = NULL;
1368     VAStatus va_status;
1369
1370     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1371     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1372     packedheader_param_buffer.bit_length = length_in_bits;
1373     packedheader_param_buffer.has_emulation_bytes = 0;
1374
1375     va_status = vaCreateBuffer(va_dpy,
1376                                context_id,
1377                                VAEncPackedHeaderParameterBufferType,
1378                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1379                                &packedpic_para_bufid);
1380     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1381
1382     va_status = vaCreateBuffer(va_dpy,
1383                                context_id,
1384                                VAEncPackedHeaderDataBufferType,
1385                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1386                                &packedpic_data_bufid);
1387     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1388
1389     render_id[0] = packedpic_para_bufid;
1390     render_id[1] = packedpic_data_bufid;
1391     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1392
1393     free(packedpic_buffer);
1394
1395     return 0;
1396 }
1397
1398 void H264EncoderImpl::render_packedslice()
1399 {
1400     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1401     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1402     unsigned int length_in_bits;
1403     unsigned char *packedslice_buffer = NULL;
1404     VAStatus va_status;
1405
1406     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1407     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1408     packedheader_param_buffer.bit_length = length_in_bits;
1409     packedheader_param_buffer.has_emulation_bytes = 0;
1410
1411     va_status = vaCreateBuffer(va_dpy,
1412                                context_id,
1413                                VAEncPackedHeaderParameterBufferType,
1414                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1415                                &packedslice_para_bufid);
1416     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1417
1418     va_status = vaCreateBuffer(va_dpy,
1419                                context_id,
1420                                VAEncPackedHeaderDataBufferType,
1421                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1422                                &packedslice_data_bufid);
1423     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1424
1425     render_id[0] = packedslice_para_bufid;
1426     render_id[1] = packedslice_data_bufid;
1427     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1428
1429     free(packedslice_buffer);
1430 }
1431
1432 int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1433 {
1434     VABufferID slice_param_buf;
1435     VAStatus va_status;
1436     int i;
1437
1438     update_RefPicList(frame_type);
1439
1440     /* one frame, one slice */
1441     slice_param.macroblock_address = 0;
1442     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1443     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1444     if (frame_type == FRAME_IDR) {
1445         if (encoding_frame_num != 0)
1446             ++slice_param.idr_pic_id;
1447     } else if (frame_type == FRAME_P) {
1448         int refpiclist0_max = h264_maxref & 0xffff;
1449         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1450
1451         for (i = refpiclist0_max; i < 32; i++) {
1452             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1453             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1454         }
1455     } else if (frame_type == FRAME_B) {
1456         int refpiclist0_max = h264_maxref & 0xffff;
1457         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1458
1459         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1460         for (i = refpiclist0_max; i < 32; i++) {
1461             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1462             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1463         }
1464
1465         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1466         for (i = refpiclist1_max; i < 32; i++) {
1467             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1468             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1469         }
1470     }
1471
1472     slice_param.slice_alpha_c0_offset_div2 = 0;
1473     slice_param.slice_beta_offset_div2 = 0;
1474     slice_param.direct_spatial_mv_pred_flag = 1;
1475     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1476
1477
1478     if (h264_packedheader &&
1479         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1480         render_packedslice();
1481
1482     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1483                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1484     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1485
1486     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1487
1488     return 0;
1489 }
1490
1491
1492
1493 void H264EncoderImpl::save_codeddata(storage_task task)
1494 {
1495     VACodedBufferSegment *buf_list = NULL;
1496     VAStatus va_status;
1497
1498     string data;
1499
1500     const int64_t global_delay = int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1501
1502     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1503     CHECK_VASTATUS(va_status, "vaMapBuffer");
1504     while (buf_list != NULL) {
1505         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1506         buf_list = (VACodedBufferSegment *) buf_list->next;
1507     }
1508     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1509
1510     {
1511         // Add video.
1512         AVPacket pkt;
1513         memset(&pkt, 0, sizeof(pkt));
1514         pkt.buf = nullptr;
1515         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1516         pkt.size = data.size();
1517         pkt.stream_index = 0;
1518         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1519             pkt.flags = AV_PKT_FLAG_KEY;
1520         } else {
1521             pkt.flags = 0;
1522         }
1523         //pkt.duration = 1;
1524         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1525     }
1526     // Encode and add all audio frames up to and including the pts of this video frame.
1527     for ( ;; ) {
1528         int64_t audio_pts;
1529         vector<float> audio;
1530         {
1531              unique_lock<mutex> lock(frame_queue_mutex);
1532              frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1533              if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1534              auto it = pending_audio_frames.begin();
1535              if (it->first > task.pts) break;
1536              audio_pts = it->first;
1537              audio = move(it->second);
1538              pending_audio_frames.erase(it);
1539         }
1540
1541         AVFrame *frame = avcodec_alloc_frame();
1542         frame->nb_samples = audio.size() / 2;
1543         frame->format = AV_SAMPLE_FMT_S32;
1544         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1545
1546         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1547         int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
1548         if (ret < 0) {
1549             fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1550             exit(1);
1551         }
1552         for (int i = 0; i < frame->nb_samples * 2; ++i) {
1553             if (audio[i] >= 1.0f) {
1554                 int_samples[i] = 2147483647;
1555             } else if (audio[i] <= -1.0f) {
1556                 int_samples[i] = -2147483647;
1557             } else {
1558                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1559             }
1560         }
1561
1562         AVPacket pkt;
1563         av_init_packet(&pkt);
1564         pkt.data = nullptr;
1565         pkt.size = 0;
1566         int got_output;
1567         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1568         if (got_output) {
1569             pkt.stream_index = 1;
1570             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1571         }
1572         // TODO: Delayed frames.
1573         avcodec_free_frame(&frame);
1574         av_free_packet(&pkt);
1575         if (audio_pts == task.pts) break;
1576     }
1577
1578 #if 0
1579     printf("\r      "); /* return back to startpoint */
1580     switch (encode_order % 4) {
1581         case 0:
1582             printf("|");
1583             break;
1584         case 1:
1585             printf("/");
1586             break;
1587         case 2:
1588             printf("-");
1589             break;
1590         case 3:
1591             printf("\\");
1592             break;
1593     }
1594     printf("%08lld", encode_order);
1595 #endif
1596 }
1597
1598
1599 // this is weird. but it seems to put a new frame onto the queue
1600 void H264EncoderImpl::storage_task_enqueue(storage_task task)
1601 {
1602         unique_lock<mutex> lock(storage_task_queue_mutex);
1603         storage_task_queue.push(move(task));
1604         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1605         storage_task_queue_changed.notify_all();
1606 }
1607
1608 void H264EncoderImpl::storage_task_thread()
1609 {
1610         for ( ;; ) {
1611                 storage_task current;
1612                 {
1613                         // wait until there's an encoded frame
1614                         unique_lock<mutex> lock(storage_task_queue_mutex);
1615                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1616                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1617                         current = move(storage_task_queue.front());
1618                         storage_task_queue.pop();
1619                 }
1620
1621                 VAStatus va_status;
1622
1623                 // waits for data, then saves it to disk.
1624                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1625                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1626                 save_codeddata(move(current));
1627
1628                 {
1629                         unique_lock<mutex> lock(storage_task_queue_mutex);
1630                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1631                         storage_task_queue_changed.notify_all();
1632                 }
1633         }
1634 }
1635
1636 int H264EncoderImpl::release_encode()
1637 {
1638         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1639                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1640                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1641                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1642
1643                 if (!use_zerocopy) {
1644                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1645                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1646                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1647                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1648                 }
1649                 glDeleteTextures(1, &gl_surfaces[i].y_tex);
1650                 glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
1651         }
1652
1653         vaDestroyContext(va_dpy, context_id);
1654         vaDestroyConfig(va_dpy, config_id);
1655
1656         return 0;
1657 }
1658
1659 int H264EncoderImpl::deinit_va()
1660 {
1661     vaTerminate(va_dpy);
1662
1663     va_close_display(va_dpy);
1664
1665     return 0;
1666 }
1667
1668
1669 H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
1670         : current_storage_frame(0), surface(surface), httpd(httpd)
1671 {
1672         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1673         context_audio = avcodec_alloc_context3(codec_audio);
1674         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1675         context_audio->sample_rate = OUTPUT_FREQUENCY;
1676         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1677         context_audio->channels = 2;
1678         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1679         context_audio->time_base = AVRational{1, TIMEBASE};
1680         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1681                 fprintf(stderr, "Could not open codec\n");
1682                 exit(1);
1683         }
1684
1685         frame_width = width;
1686         frame_height = height;
1687         frame_width_mbaligned = (frame_width + 15) & (~15);
1688         frame_height_mbaligned = (frame_height + 15) & (~15);
1689
1690         //print_input();
1691
1692         init_va(va_display);
1693         setup_encode();
1694
1695         // No frames are ready yet.
1696         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1697
1698         memset(&seq_param, 0, sizeof(seq_param));
1699         memset(&pic_param, 0, sizeof(pic_param));
1700         memset(&slice_param, 0, sizeof(slice_param));
1701
1702         storage_thread = thread(&H264EncoderImpl::storage_task_thread, this);
1703
1704         encode_thread = thread([this]{
1705                 //SDL_GL_MakeCurrent(window, context);
1706                 QOpenGLContext *context = create_context(this->surface);
1707                 eglBindAPI(EGL_OPENGL_API);
1708                 if (!make_current(context, this->surface)) {
1709                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1710                                 eglGetError());
1711                         exit(1);
1712                 }
1713                 encode_thread_func();
1714         });
1715 }
1716
1717 H264EncoderImpl::~H264EncoderImpl()
1718 {
1719         shutdown();
1720 }
1721
1722 bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1723 {
1724         assert(!is_shutdown);
1725         {
1726                 // Wait until this frame slot is done encoding.
1727                 unique_lock<mutex> lock(storage_task_queue_mutex);
1728                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1729                 if (storage_thread_should_quit) return false;
1730         }
1731
1732         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1733         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1734         *y_tex = surf->y_tex;
1735         *cbcr_tex = surf->cbcr_tex;
1736
1737         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1738         CHECK_VASTATUS(va_status, "vaDeriveImage");
1739
1740         if (use_zerocopy) {
1741                 VABufferInfo buf_info;
1742                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1743                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1744                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1745
1746                 // Create Y image.
1747                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1748                 EGLint y_attribs[] = {
1749                         EGL_WIDTH, frame_width,
1750                         EGL_HEIGHT, frame_height,
1751                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1752                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1753                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1754                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1755                         EGL_NONE
1756                 };
1757
1758                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1759                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1760
1761                 // Associate Y image to a texture.
1762                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1763                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1764
1765                 // Create CbCr image.
1766                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1767                 EGLint cbcr_attribs[] = {
1768                         EGL_WIDTH, frame_width,
1769                         EGL_HEIGHT, frame_height,
1770                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1771                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1772                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1773                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1774                         EGL_NONE
1775                 };
1776
1777                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1778                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1779
1780                 // Associate CbCr image to a texture.
1781                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1782                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1783         }
1784
1785         return true;
1786 }
1787
1788 void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
1789 {
1790         assert(!is_shutdown);
1791         {
1792                 unique_lock<mutex> lock(frame_queue_mutex);
1793                 pending_audio_frames[pts] = move(audio);
1794         }
1795         frame_queue_nonempty.notify_all();
1796 }
1797
1798 void H264EncoderImpl::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
1799 {
1800         assert(!is_shutdown);
1801
1802         if (!use_zerocopy) {
1803                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1804
1805                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1806                 check_error();
1807
1808                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1809                 check_error();
1810
1811                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1812                 check_error();
1813                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1814                 check_error();
1815
1816                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1817                 check_error();
1818                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1819                 check_error();
1820
1821                 glBindTexture(GL_TEXTURE_2D, 0);
1822                 check_error();
1823                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1824                 check_error();
1825
1826                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1827                 check_error();
1828                 fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1829                 check_error();
1830         }
1831
1832         {
1833                 unique_lock<mutex> lock(frame_queue_mutex);
1834                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1835                 ++current_storage_frame;
1836         }
1837         frame_queue_nonempty.notify_all();
1838 }
1839
1840 void H264EncoderImpl::shutdown()
1841 {
1842         if (is_shutdown) {
1843                 return;
1844         }
1845
1846         {
1847                 unique_lock<mutex> lock(frame_queue_mutex);
1848                 encode_thread_should_quit = true;
1849                 frame_queue_nonempty.notify_all();
1850         }
1851         encode_thread.join();
1852         {
1853                 unique_lock<mutex> lock(storage_task_queue_mutex);
1854                 storage_thread_should_quit = true;
1855                 frame_queue_nonempty.notify_all();
1856                 storage_task_queue_changed.notify_all();
1857         }
1858         storage_thread.join();
1859
1860         release_encode();
1861         deinit_va();
1862         is_shutdown = true;
1863 }
1864
1865 void H264EncoderImpl::encode_thread_func()
1866 {
1867         int64_t last_dts = -1;
1868         int gop_start_display_frame_num = 0;
1869         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1870                 PendingFrame frame;
1871                 int pts_lag;
1872                 int frame_type, display_frame_num;
1873                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1874                                        &display_frame_num, &frame_type, &pts_lag);
1875                 if (frame_type == FRAME_IDR) {
1876                         numShortTerm = 0;
1877                         current_frame_num = 0;
1878                         gop_start_display_frame_num = display_frame_num;
1879                 }
1880
1881                 {
1882                         unique_lock<mutex> lock(frame_queue_mutex);
1883                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1884                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1885                         });
1886                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1887                                 // We have queued frames that were supposed to be B-frames,
1888                                 // but will be no P-frame to encode them against. Encode them all
1889                                 // as P-frames instead. Note that this happens under the mutex,
1890                                 // but nobody else uses it at this point, since we're shutting down,
1891                                 // so there's no contention.
1892                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1893                                 return;
1894                         } else {
1895                                 frame = move(pending_video_frames[display_frame_num]);
1896                                 pending_video_frames.erase(display_frame_num);
1897                         }
1898                 }
1899
1900                 // Determine the dts of this frame.
1901                 int64_t dts;
1902                 if (pts_lag == -1) {
1903                         assert(last_dts != -1);
1904                         dts = last_dts + (TIMEBASE / MAX_FPS);
1905                 } else {
1906                         dts = frame.pts - pts_lag;
1907                 }
1908                 last_dts = dts;
1909
1910                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
1911         }
1912 }
1913
1914 void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1915 {
1916         if (pending_video_frames.empty()) {
1917                 return;
1918         }
1919
1920         for (auto &pending_frame : pending_video_frames) {
1921                 int display_frame_num = pending_frame.first;
1922                 assert(display_frame_num > 0);
1923                 PendingFrame frame = move(pending_frame.second);
1924                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1925                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1926                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
1927                 last_dts = dts;
1928         }
1929 }
1930
1931 namespace {
1932
1933 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1934 {
1935         if (src_width == dst_pitch) {
1936                 memcpy(dst, src, src_width * height);
1937         } else {
1938                 for (size_t y = 0; y < height; ++y) {
1939                         const uint8_t *sptr = src + y * src_width;
1940                         uint8_t *dptr = dst + y * dst_pitch;
1941                         memcpy(dptr, sptr, src_width);
1942                 }
1943         }
1944 }
1945
1946 }  // namespace
1947
1948 void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1949                                    int frame_type, int64_t pts, int64_t dts)
1950 {
1951         // Wait for the GPU to be done with the frame.
1952         GLenum sync_status;
1953         do {
1954                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
1955                 check_error();
1956         } while (sync_status == GL_TIMEOUT_EXPIRED);
1957         assert(sync_status != GL_WAIT_FAILED);
1958
1959         // Release back any input frames we needed to render this frame.
1960         frame.input_frames.clear();
1961
1962         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1963         VAStatus va_status;
1964
1965         if (use_zerocopy) {
1966                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1967                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1968                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1969                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1970         } else {
1971                 unsigned char *surface_p = nullptr;
1972                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
1973
1974                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
1975                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
1976
1977                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
1978                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
1979
1980                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
1981                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1982         }
1983
1984         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1985         CHECK_VASTATUS(va_status, "vaDestroyImage");
1986
1987         // Schedule the frame for encoding.
1988         VASurfaceID va_surface = surf->src_surface;
1989         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
1990         CHECK_VASTATUS(va_status, "vaBeginPicture");
1991
1992         if (frame_type == FRAME_IDR) {
1993                 render_sequence();
1994                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1995                 if (h264_packedheader) {
1996                         render_packedsequence();
1997                         render_packedpicture();
1998                 }
1999         } else {
2000                 //render_sequence();
2001                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2002         }
2003         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2004
2005         va_status = vaEndPicture(va_dpy, context_id);
2006         CHECK_VASTATUS(va_status, "vaEndPicture");
2007
2008         // so now the data is done encoding (well, async job kicked off)...
2009         // we send that to the storage thread
2010         storage_task tmp;
2011         tmp.display_order = display_frame_num;
2012         tmp.frame_type = frame_type;
2013         tmp.pts = pts;
2014         tmp.dts = dts;
2015         storage_task_enqueue(move(tmp));
2016
2017         update_ReferenceFrames(frame_type);
2018 }
2019
2020 // Proxy object.
2021 H264Encoder::H264Encoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
2022         : impl(new H264EncoderImpl(surface, va_display, width, height, httpd)) {}
2023
2024 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2025 H264Encoder::~H264Encoder() {}
2026
2027 void H264Encoder::add_audio(int64_t pts, vector<float> audio)
2028 {
2029         impl->add_audio(pts, audio);
2030 }
2031
2032 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2033 {
2034         return impl->begin_frame(y_tex, cbcr_tex);
2035 }
2036
2037 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
2038 {
2039         impl->end_frame(fence, pts, input_frames);
2040 }
2041
2042 void H264Encoder::shutdown()
2043 {
2044         impl->shutdown();
2045 }
2046
2047 // Real class.