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