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 "flags.h"
  42 #include "httpd.h"
  43 #include "timebase.h"
  44
  45 using namespace std;
  46
  47 class QOpenGLContext;
  48 class QSurface;
  49
  50 #define CHECK_VASTATUS(va_status, func)                                 \
  51     if (va_status != VA_STATUS_SUCCESS) {                               \
  52         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  53         exit(1);                                                        \
  54     }
  55
  56 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
  57
  58 //#include "loadsurface.h"
  59
  60 #define NAL_REF_IDC_NONE        0
  61 #define NAL_REF_IDC_LOW         1
  62 #define NAL_REF_IDC_MEDIUM      2
  63 #define NAL_REF_IDC_HIGH        3
  64
  65 #define NAL_NON_IDR             1
  66 #define NAL_IDR                 5
  67 #define NAL_SPS                 7
  68 #define NAL_PPS                 8
  69 #define NAL_SEI                 6
  70
  71 #define SLICE_TYPE_P            0
  72 #define SLICE_TYPE_B            1
  73 #define SLICE_TYPE_I            2
  74 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  75 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  76 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  77
  78
  79 #define ENTROPY_MODE_CAVLC      0
  80 #define ENTROPY_MODE_CABAC      1
  81
  82 #define PROFILE_IDC_BASELINE    66
  83 #define PROFILE_IDC_MAIN        77
  84 #define PROFILE_IDC_HIGH        100
  85
  86 #define BITSTREAM_ALLOCATE_STEPPING     4096
  87 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  88 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  89 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  90
  91 static constexpr unsigned int MaxFrameNum = (2<<16);
  92 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
  93 static constexpr unsigned int Log2MaxFrameNum = 16;
  94 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
  95 static constexpr int rc_default_modes[] = {  // Priority list of modes.
  96     VA_RC_VBR,
  97     VA_RC_CQP,
  98     VA_RC_VBR_CONSTRAINED,
  99     VA_RC_CBR,
 100     VA_RC_VCM,
 101     VA_RC_NONE,
 102 };
 103
 104 /* thread to save coded data */
 105 #define SRC_SURFACE_FREE        0
 106 #define SRC_SURFACE_IN_ENCODING 1
 107
 108 struct __bitstream {
 109     unsigned int *buffer;
 110     int bit_offset;
 111     int max_size_in_dword;
 112 };
 113 typedef struct __bitstream bitstream;
 114
 115 using namespace std;
 116
 117 // H.264 video comes out in encoding order (e.g. with two B-frames:
 118 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
 119 // come in the right order. Since we do everything, including waiting
 120 // for the frames to come out of OpenGL, in encoding order, we need
 121 // a reordering buffer for uncompressed frames so that they come out
 122 // correctly. We go the super-lazy way of not making it understand
 123 // anything about the true order (which introduces some extra latency,
 124 // though); we know that for N B-frames we need at most (N-1) frames
 125 // in the reorder buffer, and can just sort on that.
 126 //
 127 // The class also deals with keeping a freelist as needed.
 128 class FrameReorderer {
 129 public:
 130         FrameReorderer(unsigned queue_length, int width, int height);
 131
 132         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
 133         // Does _not_ take ownership of data; a copy is taken if needed.
 134         // The returned pointer is valid until the next call to reorder_frame, or destruction.
 135         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
 136         pair<int64_t, const uint8_t *> reorder_frame(int64_t pts, const uint8_t *data);
 137
 138         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
 139         pair<int64_t, const uint8_t *> get_first_frame();
 140
 141         bool empty() const { return frames.empty(); }
 142
 143 private:
 144         unsigned queue_length;
 145         int width, height;
 146
 147         priority_queue<pair<int64_t, uint8_t *>> frames;
 148         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
 149
 150         // Owns all the pointers. Normally, freelist and frames could do this themselves,
 151         // except priority_queue doesn't work well with movable-only types.
 152         vector<unique_ptr<uint8_t[]>> owner;
 153 };
 154
 155 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
 156     : queue_length(queue_length), width(width), height(height)
 157 {
 158         for (unsigned i = 0; i < queue_length; ++i) {
 159                 owner.emplace_back(new uint8_t[width * height * 2]);
 160                 freelist.push(owner.back().get());
 161         }
 162 }
 163
 164 pair<int64_t, const uint8_t *> FrameReorderer::reorder_frame(int64_t pts, const uint8_t *data)
 165 {
 166         if (queue_length == 0) {
 167                 return make_pair(pts, data);
 168         }
 169
 170         assert(!freelist.empty());
 171         uint8_t *storage = freelist.top();
 172         freelist.pop();
 173         memcpy(storage, data, width * height * 2);
 174         frames.emplace(-pts, storage);  // Invert pts to get smallest first.
 175
 176         if (frames.size() >= queue_length) {
 177                 return get_first_frame();
 178         } else {
 179                 return make_pair(-1, nullptr);
 180         }
 181 }
 182
 183 pair<int64_t, const uint8_t *> FrameReorderer::get_first_frame()
 184 {
 185         assert(!frames.empty());
 186         pair<int64_t, uint8_t *> storage = frames.top();
 187         frames.pop();
 188         int64_t pts = storage.first;
 189         freelist.push(storage.second);
 190         return make_pair(-pts, storage.second);  // Re-invert pts (see reorder_frame()).
 191 }
 192
 193 class H264EncoderImpl {
 194 public:
 195         H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd);
 196         ~H264EncoderImpl();
 197         void add_audio(int64_t pts, vector<float> audio);
 198         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 199         RefCountedGLsync end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames);
 200         void shutdown();
 201         void open_output_file(const std::string &filename);
 202         void close_output_file();
 203
 204 private:
 205         struct storage_task {
 206                 unsigned long long display_order;
 207                 int frame_type;
 208                 vector<float> audio;
 209                 int64_t pts, dts;
 210         };
 211         struct PendingFrame {
 212                 RefCountedGLsync fence;
 213                 vector<RefCountedFrame> input_frames;
 214                 int64_t pts;
 215         };
 216
 217         // So we never get negative dts.
 218         int64_t global_delay() const {
 219                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
 220         }
 221
 222         void encode_thread_func();
 223         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 224         void add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data);
 225         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 226                           int frame_type, int64_t pts, int64_t dts);
 227         void storage_task_thread();
 228         void encode_audio(const vector<float> &audio,
 229                           int64_t audio_pts,
 230                           AVCodecContext *ctx);
 231         void storage_task_enqueue(storage_task task);
 232         void save_codeddata(storage_task task);
 233         int render_packedsequence();
 234         int render_packedpicture();
 235         void render_packedslice();
 236         int render_sequence();
 237         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 238         void sps_rbsp(bitstream *bs);
 239         void pps_rbsp(bitstream *bs);
 240         int build_packed_pic_buffer(unsigned char **header_buffer);
 241         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 242         void slice_header(bitstream *bs);
 243         int build_packed_seq_buffer(unsigned char **header_buffer);
 244         int build_packed_slice_buffer(unsigned char **header_buffer);
 245         int init_va(const string &va_display);
 246         int deinit_va();
 247         void enable_zerocopy_if_possible();
 248         VADisplay va_open_display(const string &va_display);
 249         void va_close_display(VADisplay va_dpy);
 250         int setup_encode();
 251         int release_encode();
 252         void update_ReferenceFrames(int frame_type);
 253         int update_RefPicList(int frame_type);
 254
 255         bool is_shutdown = false;
 256         bool use_zerocopy;
 257         int drm_fd = -1;
 258
 259         thread encode_thread, storage_thread;
 260
 261         mutex storage_task_queue_mutex;
 262         condition_variable storage_task_queue_changed;
 263         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 264         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 265         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 266
 267         mutex frame_queue_mutex;
 268         condition_variable frame_queue_nonempty;
 269         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 270
 271         int current_storage_frame;
 272
 273         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 274         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
 275         QSurface *surface;
 276
 277         AVCodecContext *context_audio;
 278         AVFrame *audio_frame = nullptr;
 279         HTTPD *httpd;
 280         unique_ptr<FrameReorderer> reorderer;
 281
 282         Display *x11_display = nullptr;
 283
 284         // Encoder parameters
 285         VADisplay va_dpy;
 286         VAProfile h264_profile = (VAProfile)~0;
 287         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 288         int config_attrib_num = 0, enc_packed_header_idx;
 289
 290         struct GLSurface {
 291                 VASurfaceID src_surface, ref_surface;
 292                 VABufferID coded_buf;
 293
 294                 VAImage surface_image;
 295                 GLuint y_tex, cbcr_tex;
 296
 297                 // Only if use_zerocopy == true.
 298                 EGLImage y_egl_image, cbcr_egl_image;
 299
 300                 // Only if use_zerocopy == false.
 301                 GLuint pbo;
 302                 uint8_t *y_ptr, *cbcr_ptr;
 303                 size_t y_offset, cbcr_offset;
 304         };
 305         GLSurface gl_surfaces[SURFACE_NUM];
 306
 307         VAConfigID config_id;
 308         VAContextID context_id;
 309         VAEncSequenceParameterBufferH264 seq_param;
 310         VAEncPictureParameterBufferH264 pic_param;
 311         VAEncSliceParameterBufferH264 slice_param;
 312         VAPictureH264 CurrentCurrPic;
 313         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 314
 315         // Static quality settings.
 316         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 317         static constexpr unsigned int num_ref_frames = 2;
 318         static constexpr int initial_qp = 15;
 319         static constexpr int minimal_qp = 0;
 320         static constexpr int intra_period = 30;
 321         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 322
 323         // Quality settings that are meant to be static, but might be overridden
 324         // by the profile.
 325         int constraint_set_flag = 0;
 326         int h264_packedheader = 0; /* support pack header? */
 327         int h264_maxref = (1<<16|1);
 328         int h264_entropy_mode = 1; /* cabac */
 329         int ip_period = 3;
 330
 331         int rc_mode = -1;
 332         unsigned int current_frame_num = 0;
 333         unsigned int numShortTerm = 0;
 334
 335         int frame_width;
 336         int frame_height;
 337         int frame_width_mbaligned;
 338         int frame_height_mbaligned;
 339
 340         unique_ptr<Mux> file_mux;  // To local disk.
 341 };
 342
 343 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 344 // but if we don't delete it here, we get leaks. The GStreamer implementation
 345 // does the same.
 346 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 347 {
 348     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 349     CHECK_VASTATUS(va_status, "vaRenderPicture");
 350
 351     for (int i = 0; i < num_buffers; ++i) {
 352         va_status = vaDestroyBuffer(dpy, buffers[i]);
 353         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 354     }
 355 }
 356
 357 static unsigned int
 358 va_swap32(unsigned int val)
 359 {
 360     unsigned char *pval = (unsigned char *)&val;
 361
 362     return ((pval[0] << 24)     |
 363             (pval[1] << 16)     |
 364             (pval[2] << 8)      |
 365             (pval[3] << 0));
 366 }
 367
 368 static void
 369 bitstream_start(bitstream *bs)
 370 {
 371     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 372     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 373     bs->bit_offset = 0;
 374 }
 375
 376 static void
 377 bitstream_end(bitstream *bs)
 378 {
 379     int pos = (bs->bit_offset >> 5);
 380     int bit_offset = (bs->bit_offset & 0x1f);
 381     int bit_left = 32 - bit_offset;
 382
 383     if (bit_offset) {
 384         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 385     }
 386 }
 387
 388 static void
 389 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 390 {
 391     int pos = (bs->bit_offset >> 5);
 392     int bit_offset = (bs->bit_offset & 0x1f);
 393     int bit_left = 32 - bit_offset;
 394
 395     if (!size_in_bits)
 396         return;
 397
 398     bs->bit_offset += size_in_bits;
 399
 400     if (bit_left > size_in_bits) {
 401         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 402     } else {
 403         size_in_bits -= bit_left;
 404         if (bit_left >= 32) {
 405             bs->buffer[pos] = (val >> size_in_bits);
 406         } else {
 407             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 408         }
 409         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 410
 411         if (pos + 1 == bs->max_size_in_dword) {
 412             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 413             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 414         }
 415
 416         bs->buffer[pos + 1] = val;
 417     }
 418 }
 419
 420 static void
 421 bitstream_put_ue(bitstream *bs, unsigned int val)
 422 {
 423     int size_in_bits = 0;
 424     int tmp_val = ++val;
 425
 426     while (tmp_val) {
 427         tmp_val >>= 1;
 428         size_in_bits++;
 429     }
 430
 431     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 432     bitstream_put_ui(bs, val, size_in_bits);
 433 }
 434
 435 static void
 436 bitstream_put_se(bitstream *bs, int val)
 437 {
 438     unsigned int new_val;
 439
 440     if (val <= 0)
 441         new_val = -2 * val;
 442     else
 443         new_val = 2 * val - 1;
 444
 445     bitstream_put_ue(bs, new_val);
 446 }
 447
 448 static void
 449 bitstream_byte_aligning(bitstream *bs, int bit)
 450 {
 451     int bit_offset = (bs->bit_offset & 0x7);
 452     int bit_left = 8 - bit_offset;
 453     int new_val;
 454
 455     if (!bit_offset)
 456         return;
 457
 458     assert(bit == 0 || bit == 1);
 459
 460     if (bit)
 461         new_val = (1 << bit_left) - 1;
 462     else
 463         new_val = 0;
 464
 465     bitstream_put_ui(bs, new_val, bit_left);
 466 }
 467
 468 static void
 469 rbsp_trailing_bits(bitstream *bs)
 470 {
 471     bitstream_put_ui(bs, 1, 1);
 472     bitstream_byte_aligning(bs, 0);
 473 }
 474
 475 static void nal_start_code_prefix(bitstream *bs)
 476 {
 477     bitstream_put_ui(bs, 0x00000001, 32);
 478 }
 479
 480 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 481 {
 482     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 483     bitstream_put_ui(bs, nal_ref_idc, 2);
 484     bitstream_put_ui(bs, nal_unit_type, 5);
 485 }
 486
 487 void H264EncoderImpl::sps_rbsp(bitstream *bs)
 488 {
 489     int profile_idc = PROFILE_IDC_BASELINE;
 490
 491     if (h264_profile  == VAProfileH264High)
 492         profile_idc = PROFILE_IDC_HIGH;
 493     else if (h264_profile  == VAProfileH264Main)
 494         profile_idc = PROFILE_IDC_MAIN;
 495
 496     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 497     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 498     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 499     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 500     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 501     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 502     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 503     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 504
 505     if ( profile_idc == PROFILE_IDC_HIGH) {
 506         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 507         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 508         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 509         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 510         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 511     }
 512
 513     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 514     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 515
 516     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 517         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 518     else {
 519         assert(0);
 520     }
 521
 522     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 523     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 524
 525     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 526     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 527     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 528
 529     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 530         assert(0);
 531     }
 532
 533     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 534     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 535
 536     if (seq_param.frame_cropping_flag) {
 537         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 538         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 539         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 540         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 541     }
 542
 543     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 544     if ( false ) {
 545         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 546     } else {
 547         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 548         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 549         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 550         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 551         {
 552             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 553             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 554             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 555             {
 556                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 557                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 558                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 559             }
 560         }
 561         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 562         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 563         {
 564             bitstream_put_ui(bs, 1, 32);  // FPS
 565             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 566             bitstream_put_ui(bs, 1, 1);
 567         }
 568         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 569         {
 570             // hrd_parameters
 571             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 572             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 573             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 574
 575             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 576             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 577             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 578
 579             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 580             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 581             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 582             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 583         }
 584         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 585         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 586
 587         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 588         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 589     }
 590
 591     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 592 }
 593
 594
 595 void H264EncoderImpl::pps_rbsp(bitstream *bs)
 596 {
 597     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 598     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 599
 600     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 601
 602     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 603
 604     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 605
 606     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 607     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 608
 609     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 610     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 611
 612     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 613     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 614     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 615
 616     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 617     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 618     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 619
 620     /* more_rbsp_data */
 621     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 622     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 623     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 624
 625     rbsp_trailing_bits(bs);
 626 }
 627
 628 void H264EncoderImpl::slice_header(bitstream *bs)
 629 {
 630     int first_mb_in_slice = slice_param.macroblock_address;
 631
 632     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 633     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 634     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 635     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 636
 637     /* frame_mbs_only_flag == 1 */
 638     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 639         /* FIXME: */
 640         assert(0);
 641     }
 642
 643     if (pic_param.pic_fields.bits.idr_pic_flag)
 644         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 645
 646     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 647         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 648         /* pic_order_present_flag == 0 */
 649     } else {
 650         /* FIXME: */
 651         assert(0);
 652     }
 653
 654     /* redundant_pic_cnt_present_flag == 0 */
 655     /* slice type */
 656     if (IS_P_SLICE(slice_param.slice_type)) {
 657         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 658
 659         if (slice_param.num_ref_idx_active_override_flag)
 660             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 661
 662         /* ref_pic_list_reordering */
 663         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 664     } else if (IS_B_SLICE(slice_param.slice_type)) {
 665         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 666
 667         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 668
 669         if (slice_param.num_ref_idx_active_override_flag) {
 670             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 671             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 672         }
 673
 674         /* ref_pic_list_reordering */
 675         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 676         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 677     }
 678
 679     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 680          IS_P_SLICE(slice_param.slice_type)) ||
 681         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 682          IS_B_SLICE(slice_param.slice_type))) {
 683         /* FIXME: fill weight/offset table */
 684         assert(0);
 685     }
 686
 687     /* dec_ref_pic_marking */
 688     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 689         unsigned char no_output_of_prior_pics_flag = 0;
 690         unsigned char long_term_reference_flag = 0;
 691         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 692
 693         if (pic_param.pic_fields.bits.idr_pic_flag) {
 694             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 695             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 696         } else {
 697             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 698         }
 699     }
 700
 701     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 702         !IS_I_SLICE(slice_param.slice_type))
 703         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 704
 705     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 706
 707     /* ignore for SP/SI */
 708
 709     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 710         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 711
 712         if (slice_param.disable_deblocking_filter_idc != 1) {
 713             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 714             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 715         }
 716     }
 717
 718     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 719         bitstream_byte_aligning(bs, 1);
 720     }
 721 }
 722
 723 int H264EncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 724 {
 725     bitstream bs;
 726
 727     bitstream_start(&bs);
 728     nal_start_code_prefix(&bs);
 729     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 730     pps_rbsp(&bs);
 731     bitstream_end(&bs);
 732
 733     *header_buffer = (unsigned char *)bs.buffer;
 734     return bs.bit_offset;
 735 }
 736
 737 int
 738 H264EncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 739 {
 740     bitstream bs;
 741
 742     bitstream_start(&bs);
 743     nal_start_code_prefix(&bs);
 744     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 745     sps_rbsp(&bs);
 746     bitstream_end(&bs);
 747
 748     *header_buffer = (unsigned char *)bs.buffer;
 749     return bs.bit_offset;
 750 }
 751
 752 int H264EncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 753 {
 754     bitstream bs;
 755     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 756     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 757
 758     bitstream_start(&bs);
 759     nal_start_code_prefix(&bs);
 760
 761     if (IS_I_SLICE(slice_param.slice_type)) {
 762         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 763     } else if (IS_P_SLICE(slice_param.slice_type)) {
 764         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 765     } else {
 766         assert(IS_B_SLICE(slice_param.slice_type));
 767         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 768     }
 769
 770     slice_header(&bs);
 771     bitstream_end(&bs);
 772
 773     *header_buffer = (unsigned char *)bs.buffer;
 774     return bs.bit_offset;
 775 }
 776
 777
 778 /*
 779   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 780   1) period between Frame #X and Frame #N = #X - #N
 781   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 782   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 783   4) intra_period and intra_idr_period take precedence over ip_period
 784   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 785      of I/IDR frames, see bellow examples
 786   -------------------------------------------------------------------
 787   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 788   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 789   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 790   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 791   1            1                ignored    IDR IDR IDR IDR...
 792   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 793   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 794   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 795                                               (PBB)(IBB)(PBB)(IBB)...
 796   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 797                                            IDRPPPPP IPPPPP IPPPPP...
 798   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 799                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 800                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 801                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 802                                            {IDR(PBB)(PBB)}                     (6/6/3)
 803                                            {IDR(PBB)(PBB)}.
 804 */
 805
 806 // General pts/dts strategy:
 807 //
 808 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 809 // bit tricky. We assume first of all that the frame rate never goes _above_
 810 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 811 // in at least this speed, as long at dts <= pts (the frame is not attempted
 812 // presented before it is decoded). Furthermore, we never have longer chains of
 813 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 814 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 815 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 816 // frame, with an order O <= C.)
 817 //
 818 // Many strategies are possible, but we establish these rules:
 819 //
 820 //  - Tip frames have dts = pts - (C-O)*N.
 821 //  - Non-tip frames have dts = dts_last + N.
 822 //
 823 // An example, with C=2 and N=10 and the data flow showed with arrows:
 824 //
 825 //        I  B  P  B  B  P
 826 //   pts: 30 40 50 60 70 80
 827 //        ↓  ↓     ↓
 828 //   dts: 10 30 20 60 50←40
 829 //         |  |  ↑        ↑
 830 //         `--|--'        |
 831 //             `----------'
 832 //
 833 // To show that this works fine also with irregular spacings, let's say that
 834 // the third frame is delayed a bit (something earlier was dropped). Now the
 835 // situation looks like this:
 836 //
 837 //        I  B  P  B  B   P
 838 //   pts: 30 40 80 90 100 110
 839 //        ↓  ↓     ↓
 840 //   dts: 10 30 20 90 50←40
 841 //         |  |  ↑        ↑
 842 //         `--|--'        |
 843 //             `----------'
 844 //
 845 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 846 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 847 //
 848 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 849 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 850 // a tip frame and should be given a dts based on the previous one.
 851 #define FRAME_P 0
 852 #define FRAME_B 1
 853 #define FRAME_I 2
 854 #define FRAME_IDR 7
 855 void encoding2display_order(
 856     int encoding_order, int intra_period,
 857     int intra_idr_period, int ip_period,
 858     int *displaying_order,
 859     int *frame_type, int *pts_lag)
 860 {
 861     int encoding_order_gop = 0;
 862
 863     *pts_lag = 0;
 864
 865     if (intra_period == 1) { /* all are I/IDR frames */
 866         *displaying_order = encoding_order;
 867         if (intra_idr_period == 0)
 868             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 869         else
 870             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 871         return;
 872     }
 873
 874     if (intra_period == 0)
 875         intra_idr_period = 0;
 876
 877     if (ip_period == 1) {
 878         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 879         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 880         *displaying_order = encoding_order;
 881
 882         if (encoding_order_gop == 0) { /* the first frame */
 883             *frame_type = FRAME_IDR;
 884         } else if (intra_period != 0 && /* have I frames */
 885                    encoding_order_gop >= 2 &&
 886                    (encoding_order_gop % intra_period == 0)) {
 887             *frame_type = FRAME_I;
 888         } else {
 889             *frame_type = FRAME_P;
 890         }
 891         return;
 892     }
 893
 894     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 895     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 896     *pts_lag = -1;  // Most frames are not tip frames.
 897
 898     if (encoding_order_gop == 0) { /* the first frame */
 899         *frame_type = FRAME_IDR;
 900         *displaying_order = encoding_order;
 901         // IDR frames are a special case; I honestly can't find the logic behind
 902         // why this is the right thing, but it seems to line up nicely in practice :-)
 903         *pts_lag = TIMEBASE / MAX_FPS;
 904     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 905         *frame_type = FRAME_B;
 906         *displaying_order = encoding_order - 1;
 907         if ((encoding_order_gop % ip_period) == 0) {
 908             *pts_lag = 0;  // Last B-frame.
 909         }
 910     } else if (intra_period != 0 && /* have I frames */
 911                encoding_order_gop >= 2 &&
 912                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 913         *frame_type = FRAME_I;
 914         *displaying_order = encoding_order + ip_period - 1;
 915     } else {
 916         *frame_type = FRAME_P;
 917         *displaying_order = encoding_order + ip_period - 1;
 918     }
 919 }
 920
 921
 922 static const char *rc_to_string(int rc_mode)
 923 {
 924     switch (rc_mode) {
 925     case VA_RC_NONE:
 926         return "NONE";
 927     case VA_RC_CBR:
 928         return "CBR";
 929     case VA_RC_VBR:
 930         return "VBR";
 931     case VA_RC_VCM:
 932         return "VCM";
 933     case VA_RC_CQP:
 934         return "CQP";
 935     case VA_RC_VBR_CONSTRAINED:
 936         return "VBR_CONSTRAINED";
 937     default:
 938         return "Unknown";
 939     }
 940 }
 941
 942 void H264EncoderImpl::enable_zerocopy_if_possible()
 943 {
 944         if (global_flags.uncompressed_video_to_http) {
 945                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --uncompressed_video_to_http.\n");
 946                 use_zerocopy = false;
 947         } else {
 948                 use_zerocopy = true;
 949         }
 950 }
 951
 952 VADisplay H264EncoderImpl::va_open_display(const string &va_display)
 953 {
 954         if (va_display.empty()) {
 955                 x11_display = XOpenDisplay(NULL);
 956                 if (!x11_display) {
 957                         fprintf(stderr, "error: can't connect to X server!\n");
 958                         return NULL;
 959                 }
 960                 enable_zerocopy_if_possible();
 961                 return vaGetDisplay(x11_display);
 962         } else if (va_display[0] != '/') {
 963                 x11_display = XOpenDisplay(va_display.c_str());
 964                 if (!x11_display) {
 965                         fprintf(stderr, "error: can't connect to X server!\n");
 966                         return NULL;
 967                 }
 968                 enable_zerocopy_if_possible();
 969                 return vaGetDisplay(x11_display);
 970         } else {
 971                 drm_fd = open(va_display.c_str(), O_RDWR);
 972                 if (drm_fd == -1) {
 973                         perror(va_display.c_str());
 974                         return NULL;
 975                 }
 976                 use_zerocopy = false;
 977                 return vaGetDisplayDRM(drm_fd);
 978         }
 979 }
 980
 981 void H264EncoderImpl::va_close_display(VADisplay va_dpy)
 982 {
 983         if (x11_display) {
 984                 XCloseDisplay(x11_display);
 985                 x11_display = nullptr;
 986         }
 987         if (drm_fd != -1) {
 988                 close(drm_fd);
 989         }
 990 }
 991
 992 int H264EncoderImpl::init_va(const string &va_display)
 993 {
 994     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 995     VAEntrypoint *entrypoints;
 996     int num_entrypoints, slice_entrypoint;
 997     int support_encode = 0;
 998     int major_ver, minor_ver;
 999     VAStatus va_status;
1000     unsigned int i;
1001
1002     va_dpy = va_open_display(va_display);
1003     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1004     CHECK_VASTATUS(va_status, "vaInitialize");
1005
1006     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1007     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1008     if (!entrypoints) {
1009         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1010         exit(1);
1011     }
1012
1013     /* use the highest profile */
1014     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1015         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1016             continue;
1017
1018         h264_profile = profile_list[i];
1019         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1020         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1021             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1022                 support_encode = 1;
1023                 break;
1024             }
1025         }
1026         if (support_encode == 1)
1027             break;
1028     }
1029
1030     if (support_encode == 0) {
1031         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1032         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1033         printf("to use VA-API against DRM instead of X11.\n");
1034         exit(1);
1035     } else {
1036         switch (h264_profile) {
1037             case VAProfileH264Baseline:
1038                 ip_period = 1;
1039                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1040                 h264_entropy_mode = 0;
1041                 break;
1042             case VAProfileH264ConstrainedBaseline:
1043                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1044                 ip_period = 1;
1045                 break;
1046
1047             case VAProfileH264Main:
1048                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1049                 break;
1050
1051             case VAProfileH264High:
1052                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1053                 break;
1054             default:
1055                 h264_profile = VAProfileH264Baseline;
1056                 ip_period = 1;
1057                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1058                 break;
1059         }
1060     }
1061
1062     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1063
1064     /* find out the format for the render target, and rate control mode */
1065     for (i = 0; i < VAConfigAttribTypeMax; i++)
1066         attrib[i].type = (VAConfigAttribType)i;
1067
1068     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1069                                       &attrib[0], VAConfigAttribTypeMax);
1070     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1071     /* check the interested configattrib */
1072     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1073         printf("Not find desired YUV420 RT format\n");
1074         exit(1);
1075     } else {
1076         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1077         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1078         config_attrib_num++;
1079     }
1080
1081     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1082         int tmp = attrib[VAConfigAttribRateControl].value;
1083
1084         if (rc_mode == -1 || !(rc_mode & tmp))  {
1085             if (rc_mode != -1) {
1086                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1087             }
1088
1089             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1090                 if (rc_default_modes[i] & tmp) {
1091                     rc_mode = rc_default_modes[i];
1092                     break;
1093                 }
1094             }
1095         }
1096
1097         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1098         config_attrib[config_attrib_num].value = rc_mode;
1099         config_attrib_num++;
1100     }
1101
1102
1103     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1104         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1105
1106         h264_packedheader = 1;
1107         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1108         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1109
1110         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1111             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1112         }
1113
1114         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1115             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1116         }
1117
1118         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1119             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1120         }
1121
1122         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1123             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1124         }
1125
1126         enc_packed_header_idx = config_attrib_num;
1127         config_attrib_num++;
1128     }
1129
1130     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1131         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1132         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1133         config_attrib_num++;
1134     }
1135
1136     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1137         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1138     }
1139
1140     free(entrypoints);
1141     return 0;
1142 }
1143
1144 int H264EncoderImpl::setup_encode()
1145 {
1146     VAStatus va_status;
1147     VASurfaceID *tmp_surfaceid;
1148     int codedbuf_size, i;
1149     static VASurfaceID src_surface[SURFACE_NUM];
1150     static VASurfaceID ref_surface[SURFACE_NUM];
1151
1152     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1153             &config_attrib[0], config_attrib_num, &config_id);
1154     CHECK_VASTATUS(va_status, "vaCreateConfig");
1155
1156     /* create source surfaces */
1157     va_status = vaCreateSurfaces(va_dpy,
1158                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1159                                  &src_surface[0], SURFACE_NUM,
1160                                  NULL, 0);
1161     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1162
1163     /* create reference surfaces */
1164     va_status = vaCreateSurfaces(va_dpy,
1165                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1166                                  &ref_surface[0], SURFACE_NUM,
1167                                  NULL, 0);
1168     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1169
1170     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1171     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1172     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1173
1174     /* Create a context for this encode pipe */
1175     va_status = vaCreateContext(va_dpy, config_id,
1176                                 frame_width_mbaligned, frame_height_mbaligned,
1177                                 VA_PROGRESSIVE,
1178                                 tmp_surfaceid, 2 * SURFACE_NUM,
1179                                 &context_id);
1180     CHECK_VASTATUS(va_status, "vaCreateContext");
1181     free(tmp_surfaceid);
1182
1183     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1184
1185     for (i = 0; i < SURFACE_NUM; i++) {
1186         /* create coded buffer once for all
1187          * other VA buffers which won't be used again after vaRenderPicture.
1188          * so APP can always vaCreateBuffer for every frame
1189          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1190          * so VA won't maintain the coded buffer
1191          */
1192         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1193                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1194         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1195     }
1196
1197     /* create OpenGL objects */
1198     //glGenFramebuffers(SURFACE_NUM, fbos);
1199
1200     for (i = 0; i < SURFACE_NUM; i++) {
1201         glGenTextures(1, &gl_surfaces[i].y_tex);
1202         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1203
1204         if (!use_zerocopy) {
1205             // Create Y image.
1206             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
1207             glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
1208
1209             // Create CbCr image.
1210             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
1211             glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
1212
1213             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1214             // buffers, due to potentially differing pitch.
1215             glGenBuffers(1, &gl_surfaces[i].pbo);
1216             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1217             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1218             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1219             gl_surfaces[i].y_offset = 0;
1220             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1221             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1222             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1223             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1224         }
1225     }
1226
1227     for (i = 0; i < SURFACE_NUM; i++) {
1228         gl_surfaces[i].src_surface = src_surface[i];
1229         gl_surfaces[i].ref_surface = ref_surface[i];
1230     }
1231
1232     return 0;
1233 }
1234
1235 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1236 //
1237 //   2 1 0 [3] 4 8 9
1238 template<class T, class C>
1239 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1240 {
1241         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1242         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1243         sort(middle, end, less_than);
1244 }
1245
1246 void H264EncoderImpl::update_ReferenceFrames(int frame_type)
1247 {
1248     int i;
1249
1250     if (frame_type == FRAME_B)
1251         return;
1252
1253     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1254     numShortTerm++;
1255     if (numShortTerm > num_ref_frames)
1256         numShortTerm = num_ref_frames;
1257     for (i=numShortTerm-1; i>0; i--)
1258         ReferenceFrames[i] = ReferenceFrames[i-1];
1259     ReferenceFrames[0] = CurrentCurrPic;
1260
1261     current_frame_num++;
1262     if (current_frame_num > MaxFrameNum)
1263         current_frame_num = 0;
1264 }
1265
1266
1267 int H264EncoderImpl::update_RefPicList(int frame_type)
1268 {
1269     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1270         return a.frame_idx > b.frame_idx;
1271     };
1272     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1273         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1274     };
1275     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1276         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1277     };
1278
1279     if (frame_type == FRAME_P) {
1280         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1281         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1282     } else if (frame_type == FRAME_B) {
1283         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1284         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1285
1286         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1287         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1288     }
1289
1290     return 0;
1291 }
1292
1293
1294 int H264EncoderImpl::render_sequence()
1295 {
1296     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1297     VAStatus va_status;
1298     VAEncMiscParameterBuffer *misc_param;
1299     VAEncMiscParameterRateControl *misc_rate_ctrl;
1300
1301     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1302     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1303     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1304     seq_param.bits_per_second = frame_bitrate;
1305
1306     seq_param.intra_period = intra_period;
1307     seq_param.intra_idr_period = intra_idr_period;
1308     seq_param.ip_period = ip_period;
1309
1310     seq_param.max_num_ref_frames = num_ref_frames;
1311     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1312     seq_param.time_scale = TIMEBASE * 2;
1313     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1314     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1315     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1316     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1317     seq_param.seq_fields.bits.chroma_format_idc = 1;
1318     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1319
1320     if (frame_width != frame_width_mbaligned ||
1321         frame_height != frame_height_mbaligned) {
1322         seq_param.frame_cropping_flag = 1;
1323         seq_param.frame_crop_left_offset = 0;
1324         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1325         seq_param.frame_crop_top_offset = 0;
1326         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1327     }
1328
1329     va_status = vaCreateBuffer(va_dpy, context_id,
1330                                VAEncSequenceParameterBufferType,
1331                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1332     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1333
1334     va_status = vaCreateBuffer(va_dpy, context_id,
1335                                VAEncMiscParameterBufferType,
1336                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1337                                1, NULL, &rc_param_buf);
1338     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1339
1340     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1341     misc_param->type = VAEncMiscParameterTypeRateControl;
1342     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1343     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1344     misc_rate_ctrl->bits_per_second = frame_bitrate;
1345     misc_rate_ctrl->target_percentage = 66;
1346     misc_rate_ctrl->window_size = 1000;
1347     misc_rate_ctrl->initial_qp = initial_qp;
1348     misc_rate_ctrl->min_qp = minimal_qp;
1349     misc_rate_ctrl->basic_unit_size = 0;
1350     vaUnmapBuffer(va_dpy, rc_param_buf);
1351
1352     render_id[0] = seq_param_buf;
1353     render_id[1] = rc_param_buf;
1354
1355     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1356
1357     return 0;
1358 }
1359
1360 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1361 {
1362     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1363     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1364     int PicOrderCntMsb, TopFieldOrderCnt;
1365
1366     if (frame_type == FRAME_IDR)
1367         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1368     else {
1369         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1370         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1371     }
1372
1373     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1374         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1375         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1376     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1377              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1378         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1379     else
1380         PicOrderCntMsb = prevPicOrderCntMsb;
1381
1382     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1383
1384     if (frame_type != FRAME_B) {
1385         PicOrderCntMsb_ref = PicOrderCntMsb;
1386         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1387     }
1388
1389     return TopFieldOrderCnt;
1390 }
1391
1392 int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1393 {
1394     VABufferID pic_param_buf;
1395     VAStatus va_status;
1396     int i = 0;
1397
1398     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1399     pic_param.CurrPic.frame_idx = current_frame_num;
1400     pic_param.CurrPic.flags = 0;
1401     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1402     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1403     CurrentCurrPic = pic_param.CurrPic;
1404
1405     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1406     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1407         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1408         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1409     }
1410
1411     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1412     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1413     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1414     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1415     pic_param.frame_num = current_frame_num;
1416     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1417     pic_param.last_picture = false;  // FIXME
1418     pic_param.pic_init_qp = initial_qp;
1419
1420     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1421                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1422     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1423
1424     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1425
1426     return 0;
1427 }
1428
1429 int H264EncoderImpl::render_packedsequence()
1430 {
1431     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1432     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1433     unsigned int length_in_bits;
1434     unsigned char *packedseq_buffer = NULL;
1435     VAStatus va_status;
1436
1437     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1438
1439     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1440
1441     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1442     packedheader_param_buffer.has_emulation_bytes = 0;
1443     va_status = vaCreateBuffer(va_dpy,
1444                                context_id,
1445                                VAEncPackedHeaderParameterBufferType,
1446                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1447                                &packedseq_para_bufid);
1448     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1449
1450     va_status = vaCreateBuffer(va_dpy,
1451                                context_id,
1452                                VAEncPackedHeaderDataBufferType,
1453                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1454                                &packedseq_data_bufid);
1455     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1456
1457     render_id[0] = packedseq_para_bufid;
1458     render_id[1] = packedseq_data_bufid;
1459     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1460
1461     free(packedseq_buffer);
1462
1463     return 0;
1464 }
1465
1466
1467 int H264EncoderImpl::render_packedpicture()
1468 {
1469     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1470     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1471     unsigned int length_in_bits;
1472     unsigned char *packedpic_buffer = NULL;
1473     VAStatus va_status;
1474
1475     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1476     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1477     packedheader_param_buffer.bit_length = length_in_bits;
1478     packedheader_param_buffer.has_emulation_bytes = 0;
1479
1480     va_status = vaCreateBuffer(va_dpy,
1481                                context_id,
1482                                VAEncPackedHeaderParameterBufferType,
1483                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1484                                &packedpic_para_bufid);
1485     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1486
1487     va_status = vaCreateBuffer(va_dpy,
1488                                context_id,
1489                                VAEncPackedHeaderDataBufferType,
1490                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1491                                &packedpic_data_bufid);
1492     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1493
1494     render_id[0] = packedpic_para_bufid;
1495     render_id[1] = packedpic_data_bufid;
1496     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1497
1498     free(packedpic_buffer);
1499
1500     return 0;
1501 }
1502
1503 void H264EncoderImpl::render_packedslice()
1504 {
1505     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1506     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1507     unsigned int length_in_bits;
1508     unsigned char *packedslice_buffer = NULL;
1509     VAStatus va_status;
1510
1511     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1512     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1513     packedheader_param_buffer.bit_length = length_in_bits;
1514     packedheader_param_buffer.has_emulation_bytes = 0;
1515
1516     va_status = vaCreateBuffer(va_dpy,
1517                                context_id,
1518                                VAEncPackedHeaderParameterBufferType,
1519                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1520                                &packedslice_para_bufid);
1521     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1522
1523     va_status = vaCreateBuffer(va_dpy,
1524                                context_id,
1525                                VAEncPackedHeaderDataBufferType,
1526                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1527                                &packedslice_data_bufid);
1528     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1529
1530     render_id[0] = packedslice_para_bufid;
1531     render_id[1] = packedslice_data_bufid;
1532     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1533
1534     free(packedslice_buffer);
1535 }
1536
1537 int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1538 {
1539     VABufferID slice_param_buf;
1540     VAStatus va_status;
1541     int i;
1542
1543     update_RefPicList(frame_type);
1544
1545     /* one frame, one slice */
1546     slice_param.macroblock_address = 0;
1547     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1548     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1549     if (frame_type == FRAME_IDR) {
1550         if (encoding_frame_num != 0)
1551             ++slice_param.idr_pic_id;
1552     } else if (frame_type == FRAME_P) {
1553         int refpiclist0_max = h264_maxref & 0xffff;
1554         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1555
1556         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1557             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1558             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1559         }
1560     } else if (frame_type == FRAME_B) {
1561         int refpiclist0_max = h264_maxref & 0xffff;
1562         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1563
1564         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1565         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1566             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1567             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1568         }
1569
1570         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1571         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1572             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1573             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1574         }
1575     }
1576
1577     slice_param.slice_alpha_c0_offset_div2 = 0;
1578     slice_param.slice_beta_offset_div2 = 0;
1579     slice_param.direct_spatial_mv_pred_flag = 1;
1580     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1581
1582
1583     if (h264_packedheader &&
1584         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1585         render_packedslice();
1586
1587     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1588                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1589     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1590
1591     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1592
1593     return 0;
1594 }
1595
1596
1597
1598 void H264EncoderImpl::save_codeddata(storage_task task)
1599 {
1600         VACodedBufferSegment *buf_list = NULL;
1601         VAStatus va_status;
1602
1603         string data;
1604
1605         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1606         CHECK_VASTATUS(va_status, "vaMapBuffer");
1607         while (buf_list != NULL) {
1608                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1609                 buf_list = (VACodedBufferSegment *) buf_list->next;
1610         }
1611         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1612
1613         {
1614                 // Add video.
1615                 AVPacket pkt;
1616                 memset(&pkt, 0, sizeof(pkt));
1617                 pkt.buf = nullptr;
1618                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1619                 pkt.size = data.size();
1620                 pkt.stream_index = 0;
1621                 if (task.frame_type == FRAME_IDR) {
1622                         pkt.flags = AV_PKT_FLAG_KEY;
1623                 } else {
1624                         pkt.flags = 0;
1625                 }
1626                 //pkt.duration = 1;
1627                 if (file_mux) {
1628                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1629                 }
1630                 if (!global_flags.uncompressed_video_to_http) {
1631                         httpd->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1632                 }
1633         }
1634         // Encode and add all audio frames up to and including the pts of this video frame.
1635         for ( ;; ) {
1636                 int64_t audio_pts;
1637                 vector<float> audio;
1638                 {
1639                         unique_lock<mutex> lock(frame_queue_mutex);
1640                         frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1641                         if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1642                         auto it = pending_audio_frames.begin();
1643                         if (it->first > task.pts) break;
1644                         audio_pts = it->first;
1645                         audio = move(it->second);
1646                         pending_audio_frames.erase(it);
1647                 }
1648
1649                 encode_audio(audio, audio_pts, context_audio);
1650
1651                 if (audio_pts == task.pts) break;
1652         }
1653 }
1654
1655 void H264EncoderImpl::encode_audio(
1656         const vector<float> &audio,
1657         int64_t audio_pts,
1658         AVCodecContext *ctx)
1659 {
1660         audio_frame->nb_samples = audio.size() / 2;
1661         audio_frame->channel_layout = AV_CH_LAYOUT_STEREO;
1662
1663         unique_ptr<float[]> planar_samples;
1664         unique_ptr<int32_t[]> int_samples;
1665
1666         if (ctx->sample_fmt == AV_SAMPLE_FMT_FLTP) {
1667                 audio_frame->format = AV_SAMPLE_FMT_FLTP;
1668                 planar_samples.reset(new float[audio.size()]);
1669                 avcodec_fill_audio_frame(audio_frame, 2, AV_SAMPLE_FMT_FLTP, (const uint8_t*)planar_samples.get(), audio.size() * sizeof(float), 0);
1670                 for (int i = 0; i < audio_frame->nb_samples; ++i) {
1671                         planar_samples[i] = audio[i * 2 + 0];
1672                         planar_samples[i + audio_frame->nb_samples] = audio[i * 2 + 1];
1673                 }
1674         } else {
1675                 assert(ctx->sample_fmt == AV_SAMPLE_FMT_S32);
1676                 int_samples.reset(new int32_t[audio.size()]);
1677                 int ret = avcodec_fill_audio_frame(audio_frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
1678                 if (ret < 0) {
1679                         fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1680                         exit(1);
1681                 }
1682                 for (int i = 0; i < audio_frame->nb_samples * 2; ++i) {
1683                         if (audio[i] >= 1.0f) {
1684                                 int_samples[i] = 2147483647;
1685                         } else if (audio[i] <= -1.0f) {
1686                                 int_samples[i] = -2147483647;
1687                         } else {
1688                                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1689                         }
1690                 }
1691         }
1692
1693         AVPacket pkt;
1694         av_init_packet(&pkt);
1695         pkt.data = nullptr;
1696         pkt.size = 0;
1697         int got_output = 0;
1698         avcodec_encode_audio2(context_audio, &pkt, audio_frame, &got_output);
1699         if (got_output) {
1700                 pkt.stream_index = 1;
1701                 pkt.flags = AV_PKT_FLAG_KEY;
1702                 if (file_mux) {
1703                         file_mux->add_packet(pkt, audio_pts + global_delay(), audio_pts + global_delay());
1704                 }
1705                 httpd->add_packet(pkt, audio_pts + global_delay(), audio_pts + global_delay());
1706         }
1707         // TODO: Delayed frames.
1708         av_frame_unref(audio_frame);
1709         av_free_packet(&pkt);
1710 }
1711
1712 // this is weird. but it seems to put a new frame onto the queue
1713 void H264EncoderImpl::storage_task_enqueue(storage_task task)
1714 {
1715         unique_lock<mutex> lock(storage_task_queue_mutex);
1716         storage_task_queue.push(move(task));
1717         storage_task_queue_changed.notify_all();
1718 }
1719
1720 void H264EncoderImpl::storage_task_thread()
1721 {
1722         for ( ;; ) {
1723                 storage_task current;
1724                 {
1725                         // wait until there's an encoded frame
1726                         unique_lock<mutex> lock(storage_task_queue_mutex);
1727                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1728                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1729                         current = move(storage_task_queue.front());
1730                         storage_task_queue.pop();
1731                 }
1732
1733                 VAStatus va_status;
1734
1735                 // waits for data, then saves it to disk.
1736                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1737                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1738                 save_codeddata(move(current));
1739
1740                 {
1741                         unique_lock<mutex> lock(storage_task_queue_mutex);
1742                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1743                         storage_task_queue_changed.notify_all();
1744                 }
1745         }
1746 }
1747
1748 int H264EncoderImpl::release_encode()
1749 {
1750         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1751                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1752                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1753                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1754
1755                 if (!use_zerocopy) {
1756                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1757                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1758                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1759                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1760                 }
1761                 glDeleteTextures(1, &gl_surfaces[i].y_tex);
1762                 glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
1763         }
1764
1765         vaDestroyContext(va_dpy, context_id);
1766         vaDestroyConfig(va_dpy, config_id);
1767
1768         return 0;
1769 }
1770
1771 int H264EncoderImpl::deinit_va()
1772 {
1773     vaTerminate(va_dpy);
1774
1775     va_close_display(va_dpy);
1776
1777     return 0;
1778 }
1779
1780 namespace {
1781
1782 void init_audio_encoder(const string &codec_name, int bit_rate, AVCodecContext **ctx)
1783 {
1784         AVCodec *codec_audio = avcodec_find_encoder_by_name(codec_name.c_str());
1785         if (codec_audio == nullptr) {
1786                 fprintf(stderr, "ERROR: Could not find codec '%s'\n", codec_name.c_str());
1787                 exit(1);
1788         }
1789
1790         AVCodecContext *context_audio = avcodec_alloc_context3(codec_audio);
1791         context_audio->bit_rate = bit_rate;
1792         context_audio->sample_rate = OUTPUT_FREQUENCY;
1793
1794         // Choose sample format; we currently only support these two
1795         // (see encode_audio), so we're a bit picky.
1796         const AVSampleFormat *ptr = codec_audio->sample_fmts;
1797         for ( ; *ptr != -1; ++ptr) {
1798                 if (*ptr == AV_SAMPLE_FMT_FLTP || *ptr == AV_SAMPLE_FMT_S32) {
1799                         context_audio->sample_fmt = *ptr;
1800                         break;
1801                 }
1802         }
1803         if (*ptr == -1) {
1804                 fprintf(stderr, "ERROR: Audio codec does not support fltp or s32 sample formats\n");
1805                 exit(1);
1806         }
1807
1808         context_audio->channels = 2;
1809         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1810         context_audio->time_base = AVRational{1, TIMEBASE};
1811         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1812                 fprintf(stderr, "Could not open codec '%s'\n", codec_name.c_str());
1813                 exit(1);
1814         }
1815
1816         *ctx = context_audio;
1817 }
1818
1819 }  // namespace
1820
1821 H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
1822         : current_storage_frame(0), surface(surface), httpd(httpd)
1823 {
1824         init_audio_encoder(AUDIO_OUTPUT_CODEC_NAME, AUDIO_OUTPUT_BIT_RATE, &context_audio);
1825
1826         audio_frame = av_frame_alloc();
1827
1828         frame_width = width;
1829         frame_height = height;
1830         frame_width_mbaligned = (frame_width + 15) & (~15);
1831         frame_height_mbaligned = (frame_height + 15) & (~15);
1832
1833         //print_input();
1834
1835         if (global_flags.uncompressed_video_to_http) {
1836                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1837         }
1838
1839         init_va(va_display);
1840         setup_encode();
1841
1842         // No frames are ready yet.
1843         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1844
1845         memset(&seq_param, 0, sizeof(seq_param));
1846         memset(&pic_param, 0, sizeof(pic_param));
1847         memset(&slice_param, 0, sizeof(slice_param));
1848
1849         storage_thread = thread(&H264EncoderImpl::storage_task_thread, this);
1850
1851         encode_thread = thread([this]{
1852                 //SDL_GL_MakeCurrent(window, context);
1853                 QOpenGLContext *context = create_context(this->surface);
1854                 eglBindAPI(EGL_OPENGL_API);
1855                 if (!make_current(context, this->surface)) {
1856                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1857                                 eglGetError());
1858                         exit(1);
1859                 }
1860                 encode_thread_func();
1861         });
1862 }
1863
1864 H264EncoderImpl::~H264EncoderImpl()
1865 {
1866         shutdown();
1867         av_frame_free(&audio_frame);
1868
1869         // TODO: Destroy context.
1870 }
1871
1872 bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1873 {
1874         assert(!is_shutdown);
1875         {
1876                 // Wait until this frame slot is done encoding.
1877                 unique_lock<mutex> lock(storage_task_queue_mutex);
1878                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1879                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1880                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1881                 }
1882                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1883                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1884                 if (storage_thread_should_quit) return false;
1885         }
1886
1887         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1888         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1889         *y_tex = surf->y_tex;
1890         *cbcr_tex = surf->cbcr_tex;
1891
1892         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1893         CHECK_VASTATUS(va_status, "vaDeriveImage");
1894
1895         if (use_zerocopy) {
1896                 VABufferInfo buf_info;
1897                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1898                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1899                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1900
1901                 // Create Y image.
1902                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1903                 EGLint y_attribs[] = {
1904                         EGL_WIDTH, frame_width,
1905                         EGL_HEIGHT, frame_height,
1906                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1907                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1908                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1909                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1910                         EGL_NONE
1911                 };
1912
1913                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1914                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1915
1916                 // Associate Y image to a texture.
1917                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1918                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1919
1920                 // Create CbCr image.
1921                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1922                 EGLint cbcr_attribs[] = {
1923                         EGL_WIDTH, frame_width,
1924                         EGL_HEIGHT, frame_height,
1925                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1926                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1927                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1928                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1929                         EGL_NONE
1930                 };
1931
1932                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1933                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1934
1935                 // Associate CbCr image to a texture.
1936                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1937                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1938         }
1939
1940         return true;
1941 }
1942
1943 void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
1944 {
1945         assert(!is_shutdown);
1946         {
1947                 unique_lock<mutex> lock(frame_queue_mutex);
1948                 pending_audio_frames[pts] = move(audio);
1949         }
1950         frame_queue_nonempty.notify_all();
1951 }
1952
1953 RefCountedGLsync H264EncoderImpl::end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames)
1954 {
1955         assert(!is_shutdown);
1956
1957         if (!use_zerocopy) {
1958                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1959
1960                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1961                 check_error();
1962
1963                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1964                 check_error();
1965
1966                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1967                 check_error();
1968                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1969                 check_error();
1970
1971                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1972                 check_error();
1973                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1974                 check_error();
1975
1976                 glBindTexture(GL_TEXTURE_2D, 0);
1977                 check_error();
1978                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1979                 check_error();
1980
1981                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1982                 check_error();
1983         }
1984
1985         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1986         check_error();
1987         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1988         check_error();
1989
1990         {
1991                 unique_lock<mutex> lock(frame_queue_mutex);
1992                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1993                 ++current_storage_frame;
1994         }
1995         frame_queue_nonempty.notify_all();
1996         return fence;
1997 }
1998
1999 void H264EncoderImpl::shutdown()
2000 {
2001         if (is_shutdown) {
2002                 return;
2003         }
2004
2005         {
2006                 unique_lock<mutex> lock(frame_queue_mutex);
2007                 encode_thread_should_quit = true;
2008                 frame_queue_nonempty.notify_all();
2009         }
2010         encode_thread.join();
2011         {
2012                 unique_lock<mutex> lock(storage_task_queue_mutex);
2013                 storage_thread_should_quit = true;
2014                 frame_queue_nonempty.notify_all();
2015                 storage_task_queue_changed.notify_all();
2016         }
2017         storage_thread.join();
2018
2019         release_encode();
2020         deinit_va();
2021         is_shutdown = true;
2022 }
2023
2024 void H264EncoderImpl::open_output_file(const std::string &filename)
2025 {
2026         AVFormatContext *avctx = avformat_alloc_context();
2027         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
2028         assert(filename.size() < sizeof(avctx->filename) - 1);
2029         strcpy(avctx->filename, filename.c_str());
2030
2031         string url = "file:" + filename;
2032         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
2033         if (ret < 0) {
2034                 char tmp[AV_ERROR_MAX_STRING_SIZE];
2035                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
2036                 exit(1);
2037         }
2038
2039         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, TIMEBASE));
2040 }
2041
2042 void H264EncoderImpl::close_output_file()
2043 {
2044         file_mux.reset();
2045 }
2046
2047 void H264EncoderImpl::encode_thread_func()
2048 {
2049         int64_t last_dts = -1;
2050         int gop_start_display_frame_num = 0;
2051         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
2052                 PendingFrame frame;
2053                 int pts_lag;
2054                 int frame_type, display_frame_num;
2055                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
2056                                        &display_frame_num, &frame_type, &pts_lag);
2057                 if (frame_type == FRAME_IDR) {
2058                         numShortTerm = 0;
2059                         current_frame_num = 0;
2060                         gop_start_display_frame_num = display_frame_num;
2061                 }
2062
2063                 {
2064                         unique_lock<mutex> lock(frame_queue_mutex);
2065                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
2066                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
2067                         });
2068                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
2069                                 // We have queued frames that were supposed to be B-frames,
2070                                 // but will be no P-frame to encode them against. Encode them all
2071                                 // as P-frames instead. Note that this happens under the mutex,
2072                                 // but nobody else uses it at this point, since we're shutting down,
2073                                 // so there's no contention.
2074                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
2075                                 return;
2076                         } else {
2077                                 frame = move(pending_video_frames[display_frame_num]);
2078                                 pending_video_frames.erase(display_frame_num);
2079                         }
2080                 }
2081
2082                 // Determine the dts of this frame.
2083                 int64_t dts;
2084                 if (pts_lag == -1) {
2085                         assert(last_dts != -1);
2086                         dts = last_dts + (TIMEBASE / MAX_FPS);
2087                 } else {
2088                         dts = frame.pts - pts_lag;
2089                 }
2090                 last_dts = dts;
2091
2092                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
2093         }
2094 }
2095
2096 void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2097 {
2098         if (pending_video_frames.empty()) {
2099                 return;
2100         }
2101
2102         for (auto &pending_frame : pending_video_frames) {
2103                 int display_frame_num = pending_frame.first;
2104                 assert(display_frame_num > 0);
2105                 PendingFrame frame = move(pending_frame.second);
2106                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2107                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2108                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
2109                 last_dts = dts;
2110         }
2111
2112         if (global_flags.uncompressed_video_to_http) {
2113                 // Add frames left in reorderer.
2114                 while (!reorderer->empty()) {
2115                         pair<int64_t, const uint8_t *> output_frame = reorderer->get_first_frame();
2116                         add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
2117                 }
2118         }
2119 }
2120
2121 void H264EncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data)
2122 {
2123         AVPacket pkt;
2124         memset(&pkt, 0, sizeof(pkt));
2125         pkt.buf = nullptr;
2126         pkt.data = const_cast<uint8_t *>(data);
2127         pkt.size = frame_width * frame_height * 2;
2128         pkt.stream_index = 0;
2129         pkt.flags = AV_PKT_FLAG_KEY;
2130         httpd->add_packet(pkt, pts, pts);
2131 }
2132
2133 namespace {
2134
2135 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2136 {
2137         if (src_width == dst_pitch) {
2138                 memcpy(dst, src, src_width * height);
2139         } else {
2140                 for (size_t y = 0; y < height; ++y) {
2141                         const uint8_t *sptr = src + y * src_width;
2142                         uint8_t *dptr = dst + y * dst_pitch;
2143                         memcpy(dptr, sptr, src_width);
2144                 }
2145         }
2146 }
2147
2148 }  // namespace
2149
2150 void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2151                                    int frame_type, int64_t pts, int64_t dts)
2152 {
2153         // Wait for the GPU to be done with the frame.
2154         GLenum sync_status;
2155         do {
2156                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2157                 check_error();
2158         } while (sync_status == GL_TIMEOUT_EXPIRED);
2159         assert(sync_status != GL_WAIT_FAILED);
2160
2161         // Release back any input frames we needed to render this frame.
2162         frame.input_frames.clear();
2163
2164         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2165         VAStatus va_status;
2166
2167         if (use_zerocopy) {
2168                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2169                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2170                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2171                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2172         } else {
2173                 unsigned char *surface_p = nullptr;
2174                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2175
2176                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2177                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2178
2179                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2180                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2181
2182                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2183                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2184
2185                 if (global_flags.uncompressed_video_to_http) {
2186                         // Add uncompressed video. (Note that pts == dts here.)
2187                         // Delay needs to match audio.
2188                         pair<int64_t, const uint8_t *> output_frame = reorderer->reorder_frame(pts + global_delay(), reinterpret_cast<uint8_t *>(surf->y_ptr));
2189                         if (output_frame.second != nullptr) {
2190                                 add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
2191                         }
2192                 }
2193         }
2194
2195         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2196         CHECK_VASTATUS(va_status, "vaDestroyImage");
2197
2198         // Schedule the frame for encoding.
2199         VASurfaceID va_surface = surf->src_surface;
2200         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2201         CHECK_VASTATUS(va_status, "vaBeginPicture");
2202
2203         if (frame_type == FRAME_IDR) {
2204                 render_sequence();
2205                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2206                 if (h264_packedheader) {
2207                         render_packedsequence();
2208                         render_packedpicture();
2209                 }
2210         } else {
2211                 //render_sequence();
2212                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2213         }
2214         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2215
2216         va_status = vaEndPicture(va_dpy, context_id);
2217         CHECK_VASTATUS(va_status, "vaEndPicture");
2218
2219         // so now the data is done encoding (well, async job kicked off)...
2220         // we send that to the storage thread
2221         storage_task tmp;
2222         tmp.display_order = display_frame_num;
2223         tmp.frame_type = frame_type;
2224         tmp.pts = pts;
2225         tmp.dts = dts;
2226         storage_task_enqueue(move(tmp));
2227
2228         update_ReferenceFrames(frame_type);
2229 }
2230
2231 // Proxy object.
2232 H264Encoder::H264Encoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
2233         : impl(new H264EncoderImpl(surface, va_display, width, height, httpd)) {}
2234
2235 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2236 H264Encoder::~H264Encoder() {}
2237
2238 void H264Encoder::add_audio(int64_t pts, vector<float> audio)
2239 {
2240         impl->add_audio(pts, audio);
2241 }
2242
2243 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2244 {
2245         return impl->begin_frame(y_tex, cbcr_tex);
2246 }
2247
2248 RefCountedGLsync H264Encoder::end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames)
2249 {
2250         return impl->end_frame(pts, input_frames);
2251 }
2252
2253 void H264Encoder::shutdown()
2254 {
2255         impl->shutdown();
2256 }
2257
2258 void H264Encoder::open_output_file(const std::string &filename)
2259 {
2260         impl->open_output_file(filename);
2261 }
2262
2263 void H264Encoder::close_output_file()
2264 {
2265         impl->close_output_file();
2266 }