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