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