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

   1 //#include "sysdeps.h"
   2 #include "h264encode.h"
   3
   4 #include <EGL/eglplatform.h>
   5 #include <X11/X.h>
   6 #include <X11/Xlib.h>
   7 #include <assert.h>
   8 #include <epoxy/egl.h>
   9 extern "C" {
  10 #include <libavcodec/avcodec.h>
  11 #include <libavformat/avformat.h>
  12 #include <libavutil/channel_layout.h>
  13 #include <libavutil/frame.h>
  14 #include <libavutil/rational.h>
  15 #include <libavutil/samplefmt.h>
  16 }
  17 #include <libdrm/drm_fourcc.h>
  18 #include <stdio.h>
  19 #include <stdlib.h>
  20 #include <string.h>
  21 #include <va/va.h>
  22 #include <va/va_drmcommon.h>
  23 #include <va/va_enc_h264.h>
  24 #include <va/va_x11.h>
  25 #include <condition_variable>
  26 #include <cstdint>
  27 #include <map>
  28 #include <memory>
  29 #include <mutex>
  30 #include <queue>
  31 #include <string>
  32 #include <thread>
  33 #include <utility>
  34
  35 #include "context.h"
  36 #include "defs.h"
  37 #include "httpd.h"
  38 #include "timebase.h"
  39
  40 using namespace std;
  41
  42 class QOpenGLContext;
  43 class QSurface;
  44
  45 #define CHECK_VASTATUS(va_status, func)                                 \
  46     if (va_status != VA_STATUS_SUCCESS) {                               \
  47         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  48         exit(1);                                                        \
  49     }
  50
  51 //#include "loadsurface.h"
  52
  53 #define NAL_REF_IDC_NONE        0
  54 #define NAL_REF_IDC_LOW         1
  55 #define NAL_REF_IDC_MEDIUM      2
  56 #define NAL_REF_IDC_HIGH        3
  57
  58 #define NAL_NON_IDR             1
  59 #define NAL_IDR                 5
  60 #define NAL_SPS                 7
  61 #define NAL_PPS                 8
  62 #define NAL_SEI                 6
  63
  64 #define SLICE_TYPE_P            0
  65 #define SLICE_TYPE_B            1
  66 #define SLICE_TYPE_I            2
  67 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
  68 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
  69 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
  70
  71
  72 #define ENTROPY_MODE_CAVLC      0
  73 #define ENTROPY_MODE_CABAC      1
  74
  75 #define PROFILE_IDC_BASELINE    66
  76 #define PROFILE_IDC_MAIN        77
  77 #define PROFILE_IDC_HIGH        100
  78
  79 #define BITSTREAM_ALLOCATE_STEPPING     4096
  80 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  81
  82 static constexpr unsigned int MaxFrameNum = (2<<16);
  83 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
  84 static constexpr unsigned int Log2MaxFrameNum = 16;
  85 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
  86 static constexpr int rc_default_modes[] = {  // Priority list of modes.
  87     VA_RC_VBR,
  88     VA_RC_CQP,
  89     VA_RC_VBR_CONSTRAINED,
  90     VA_RC_CBR,
  91     VA_RC_VCM,
  92     VA_RC_NONE,
  93 };
  94
  95 /* thread to save coded data */
  96 #define SRC_SURFACE_FREE        0
  97 #define SRC_SURFACE_IN_ENCODING 1
  98
  99 struct __bitstream {
 100     unsigned int *buffer;
 101     int bit_offset;
 102     int max_size_in_dword;
 103 };
 104 typedef struct __bitstream bitstream;
 105
 106 using namespace std;
 107
 108 class H264EncoderImpl {
 109 public:
 110         H264EncoderImpl(QSurface *surface, int width, int height, HTTPD *httpd);
 111         ~H264EncoderImpl();
 112         void add_audio(int64_t pts, std::vector<float> audio);  // Needs to come before end_frame() of same pts.
 113         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
 114         void end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames);
 115
 116 private:
 117         struct storage_task {
 118                 unsigned long long display_order;
 119                 int frame_type;
 120                 std::vector<float> audio;
 121                 int64_t pts, dts;
 122         };
 123         struct PendingFrame {
 124                 RefCountedGLsync fence;
 125                 std::vector<RefCountedFrame> input_frames;
 126                 int64_t pts;
 127         };
 128
 129         void encode_thread_func();
 130         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 131         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 132                           int frame_type, int64_t pts, int64_t dts);
 133         void storage_task_thread();
 134         void storage_task_enqueue(storage_task task);
 135         void save_codeddata(storage_task task);
 136         int render_packedsequence();
 137         int render_packedpicture();
 138         void render_packedslice();
 139         int render_sequence();
 140         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 141         void sps_rbsp(bitstream *bs);
 142         void pps_rbsp(bitstream *bs);
 143         int build_packed_pic_buffer(unsigned char **header_buffer);
 144         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 145         void slice_header(bitstream *bs);
 146         int build_packed_seq_buffer(unsigned char **header_buffer);
 147         int build_packed_slice_buffer(unsigned char **header_buffer);
 148         int init_va();
 149         int deinit_va();
 150         VADisplay va_open_display(void);
 151         void va_close_display(VADisplay va_dpy);
 152         int setup_encode();
 153         int release_encode();
 154         void update_ReferenceFrames(int frame_type);
 155         int update_RefPicList(int frame_type);
 156
 157         std::thread encode_thread, storage_thread;
 158
 159         std::mutex storage_task_queue_mutex;
 160         std::condition_variable storage_task_queue_changed;
 161         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 162         std::queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 163         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 164
 165         std::mutex frame_queue_mutex;
 166         std::condition_variable frame_queue_nonempty;
 167         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 168
 169         int current_storage_frame;
 170
 171         std::map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 172         std::map<int64_t, std::vector<float>> pending_audio_frames;  // under frame_queue_mutex
 173         QSurface *surface;
 174
 175         AVCodecContext *context_audio;
 176         HTTPD *httpd;
 177
 178         Display *x11_display;
 179         Window x11_window;
 180
 181         // Encoder parameters
 182         VADisplay va_dpy;
 183         VAProfile h264_profile = (VAProfile)~0;
 184         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 185         int config_attrib_num = 0, enc_packed_header_idx;
 186
 187         struct GLSurface {
 188                 VASurfaceID src_surface, ref_surface;
 189                 VABufferID coded_buf;
 190
 191                 VAImage surface_image;
 192                 GLuint y_tex, cbcr_tex;
 193                 EGLImage y_egl_image, cbcr_egl_image;
 194         };
 195         GLSurface gl_surfaces[SURFACE_NUM];
 196
 197         VAConfigID config_id;
 198         VAContextID context_id;
 199         VAEncSequenceParameterBufferH264 seq_param;
 200         VAEncPictureParameterBufferH264 pic_param;
 201         VAEncSliceParameterBufferH264 slice_param;
 202         VAPictureH264 CurrentCurrPic;
 203         VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
 204
 205         // Static quality settings.
 206         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 207         static constexpr unsigned int num_ref_frames = 2;
 208         static constexpr int initial_qp = 15;
 209         static constexpr int minimal_qp = 0;
 210         static constexpr int intra_period = 30;
 211         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 212
 213         // Quality settings that are meant to be static, but might be overridden
 214         // by the profile.
 215         int constraint_set_flag = 0;
 216         int h264_packedheader = 0; /* support pack header? */
 217         int h264_maxref = (1<<16|1);
 218         int h264_entropy_mode = 1; /* cabac */
 219         int ip_period = 3;
 220
 221         int rc_mode = -1;
 222         unsigned int current_frame_num = 0;
 223         unsigned int numShortTerm = 0;
 224
 225         int frame_width;
 226         int frame_height;
 227         int frame_width_mbaligned;
 228         int frame_height_mbaligned;
 229 };
 230
 231
 232 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
 233 // but if we don't delete it here, we get leaks. The GStreamer implementation
 234 // does the same.
 235 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
 236 {
 237     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
 238     CHECK_VASTATUS(va_status, "vaRenderPicture");
 239
 240     for (int i = 0; i < num_buffers; ++i) {
 241         va_status = vaDestroyBuffer(dpy, buffers[i]);
 242         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 243     }
 244 }
 245
 246 static unsigned int
 247 va_swap32(unsigned int val)
 248 {
 249     unsigned char *pval = (unsigned char *)&val;
 250
 251     return ((pval[0] << 24)     |
 252             (pval[1] << 16)     |
 253             (pval[2] << 8)      |
 254             (pval[3] << 0));
 255 }
 256
 257 static void
 258 bitstream_start(bitstream *bs)
 259 {
 260     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
 261     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
 262     bs->bit_offset = 0;
 263 }
 264
 265 static void
 266 bitstream_end(bitstream *bs)
 267 {
 268     int pos = (bs->bit_offset >> 5);
 269     int bit_offset = (bs->bit_offset & 0x1f);
 270     int bit_left = 32 - bit_offset;
 271
 272     if (bit_offset) {
 273         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
 274     }
 275 }
 276
 277 static void
 278 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
 279 {
 280     int pos = (bs->bit_offset >> 5);
 281     int bit_offset = (bs->bit_offset & 0x1f);
 282     int bit_left = 32 - bit_offset;
 283
 284     if (!size_in_bits)
 285         return;
 286
 287     bs->bit_offset += size_in_bits;
 288
 289     if (bit_left > size_in_bits) {
 290         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
 291     } else {
 292         size_in_bits -= bit_left;
 293         bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
 294         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
 295
 296         if (pos + 1 == bs->max_size_in_dword) {
 297             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
 298             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
 299         }
 300
 301         bs->buffer[pos + 1] = val;
 302     }
 303 }
 304
 305 static void
 306 bitstream_put_ue(bitstream *bs, unsigned int val)
 307 {
 308     int size_in_bits = 0;
 309     int tmp_val = ++val;
 310
 311     while (tmp_val) {
 312         tmp_val >>= 1;
 313         size_in_bits++;
 314     }
 315
 316     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
 317     bitstream_put_ui(bs, val, size_in_bits);
 318 }
 319
 320 static void
 321 bitstream_put_se(bitstream *bs, int val)
 322 {
 323     unsigned int new_val;
 324
 325     if (val <= 0)
 326         new_val = -2 * val;
 327     else
 328         new_val = 2 * val - 1;
 329
 330     bitstream_put_ue(bs, new_val);
 331 }
 332
 333 static void
 334 bitstream_byte_aligning(bitstream *bs, int bit)
 335 {
 336     int bit_offset = (bs->bit_offset & 0x7);
 337     int bit_left = 8 - bit_offset;
 338     int new_val;
 339
 340     if (!bit_offset)
 341         return;
 342
 343     assert(bit == 0 || bit == 1);
 344
 345     if (bit)
 346         new_val = (1 << bit_left) - 1;
 347     else
 348         new_val = 0;
 349
 350     bitstream_put_ui(bs, new_val, bit_left);
 351 }
 352
 353 static void
 354 rbsp_trailing_bits(bitstream *bs)
 355 {
 356     bitstream_put_ui(bs, 1, 1);
 357     bitstream_byte_aligning(bs, 0);
 358 }
 359
 360 static void nal_start_code_prefix(bitstream *bs)
 361 {
 362     bitstream_put_ui(bs, 0x00000001, 32);
 363 }
 364
 365 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
 366 {
 367     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
 368     bitstream_put_ui(bs, nal_ref_idc, 2);
 369     bitstream_put_ui(bs, nal_unit_type, 5);
 370 }
 371
 372 void H264EncoderImpl::sps_rbsp(bitstream *bs)
 373 {
 374     int profile_idc = PROFILE_IDC_BASELINE;
 375
 376     if (h264_profile  == VAProfileH264High)
 377         profile_idc = PROFILE_IDC_HIGH;
 378     else if (h264_profile  == VAProfileH264Main)
 379         profile_idc = PROFILE_IDC_MAIN;
 380
 381     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
 382     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
 383     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
 384     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
 385     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
 386     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
 387     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
 388     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 389
 390     if ( profile_idc == PROFILE_IDC_HIGH) {
 391         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */
 392         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
 393         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
 394         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
 395         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
 396     }
 397
 398     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
 399     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
 400
 401     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
 402         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
 403     else {
 404         assert(0);
 405     }
 406
 407     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
 408     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
 409
 410     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
 411     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
 412     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
 413
 414     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 415         assert(0);
 416     }
 417
 418     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
 419     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
 420
 421     if (seq_param.frame_cropping_flag) {
 422         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
 423         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
 424         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
 425         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
 426     }
 427
 428     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
 429     if ( false ) {
 430         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
 431     } else {
 432         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
 433         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
 434         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
 435         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
 436         {
 437             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
 438             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
 439             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
 440             {
 441                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
 442                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
 443                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
 444             }
 445         }
 446         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
 447         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
 448         {
 449             bitstream_put_ui(bs, 1, 32);  // FPS
 450             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
 451             bitstream_put_ui(bs, 1, 1);
 452         }
 453         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
 454         {
 455             // hrd_parameters
 456             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
 457             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
 458             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
 459
 460             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
 461             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
 462             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
 463
 464             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
 465             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
 466             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
 467             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
 468         }
 469         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
 470         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */
 471
 472         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
 473         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
 474     }
 475
 476     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
 477 }
 478
 479
 480 void H264EncoderImpl::pps_rbsp(bitstream *bs)
 481 {
 482     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
 483     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
 484
 485     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
 486
 487     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
 488
 489     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
 490
 491     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
 492     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
 493
 494     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
 495     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
 496
 497     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
 498     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
 499     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
 500
 501     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
 502     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
 503     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
 504
 505     /* more_rbsp_data */
 506     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
 507     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
 508     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
 509
 510     rbsp_trailing_bits(bs);
 511 }
 512
 513 void H264EncoderImpl::slice_header(bitstream *bs)
 514 {
 515     int first_mb_in_slice = slice_param.macroblock_address;
 516
 517     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
 518     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
 519     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
 520     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
 521
 522     /* frame_mbs_only_flag == 1 */
 523     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
 524         /* FIXME: */
 525         assert(0);
 526     }
 527
 528     if (pic_param.pic_fields.bits.idr_pic_flag)
 529         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
 530
 531     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
 532         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
 533         /* pic_order_present_flag == 0 */
 534     } else {
 535         /* FIXME: */
 536         assert(0);
 537     }
 538
 539     /* redundant_pic_cnt_present_flag == 0 */
 540     /* slice type */
 541     if (IS_P_SLICE(slice_param.slice_type)) {
 542         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
 543
 544         if (slice_param.num_ref_idx_active_override_flag)
 545             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 546
 547         /* ref_pic_list_reordering */
 548         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 549     } else if (IS_B_SLICE(slice_param.slice_type)) {
 550         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
 551
 552         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
 553
 554         if (slice_param.num_ref_idx_active_override_flag) {
 555             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
 556             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
 557         }
 558
 559         /* ref_pic_list_reordering */
 560         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
 561         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
 562     }
 563
 564     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
 565          IS_P_SLICE(slice_param.slice_type)) ||
 566         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
 567          IS_B_SLICE(slice_param.slice_type))) {
 568         /* FIXME: fill weight/offset table */
 569         assert(0);
 570     }
 571
 572     /* dec_ref_pic_marking */
 573     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
 574         unsigned char no_output_of_prior_pics_flag = 0;
 575         unsigned char long_term_reference_flag = 0;
 576         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
 577
 578         if (pic_param.pic_fields.bits.idr_pic_flag) {
 579             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
 580             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
 581         } else {
 582             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
 583         }
 584     }
 585
 586     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
 587         !IS_I_SLICE(slice_param.slice_type))
 588         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
 589
 590     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
 591
 592     /* ignore for SP/SI */
 593
 594     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
 595         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
 596
 597         if (slice_param.disable_deblocking_filter_idc != 1) {
 598             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
 599             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
 600         }
 601     }
 602
 603     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
 604         bitstream_byte_aligning(bs, 1);
 605     }
 606 }
 607
 608 int H264EncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
 609 {
 610     bitstream bs;
 611
 612     bitstream_start(&bs);
 613     nal_start_code_prefix(&bs);
 614     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
 615     pps_rbsp(&bs);
 616     bitstream_end(&bs);
 617
 618     *header_buffer = (unsigned char *)bs.buffer;
 619     return bs.bit_offset;
 620 }
 621
 622 int
 623 H264EncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
 624 {
 625     bitstream bs;
 626
 627     bitstream_start(&bs);
 628     nal_start_code_prefix(&bs);
 629     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
 630     sps_rbsp(&bs);
 631     bitstream_end(&bs);
 632
 633     *header_buffer = (unsigned char *)bs.buffer;
 634     return bs.bit_offset;
 635 }
 636
 637 int H264EncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
 638 {
 639     bitstream bs;
 640     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
 641     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
 642
 643     bitstream_start(&bs);
 644     nal_start_code_prefix(&bs);
 645
 646     if (IS_I_SLICE(slice_param.slice_type)) {
 647         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
 648     } else if (IS_P_SLICE(slice_param.slice_type)) {
 649         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
 650     } else {
 651         assert(IS_B_SLICE(slice_param.slice_type));
 652         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
 653     }
 654
 655     slice_header(&bs);
 656     bitstream_end(&bs);
 657
 658     *header_buffer = (unsigned char *)bs.buffer;
 659     return bs.bit_offset;
 660 }
 661
 662
 663 /*
 664   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
 665   1) period between Frame #X and Frame #N = #X - #N
 666   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
 667   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
 668   4) intra_period and intra_idr_period take precedence over ip_period
 669   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods
 670      of I/IDR frames, see bellow examples
 671   -------------------------------------------------------------------
 672   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
 673   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
 674   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
 675   1            0                ignored    IDRIIIIIII...      (No IDR any more)
 676   1            1                ignored    IDR IDR IDR IDR...
 677   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
 678   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
 679   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
 680                                               (PBB)(IBB)(PBB)(IBB)...
 681   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
 682                                            IDRPPPPP IPPPPP IPPPPP...
 683   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
 684                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
 685                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
 686                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
 687                                            {IDR(PBB)(PBB)}                     (6/6/3)
 688                                            {IDR(PBB)(PBB)}.
 689 */
 690
 691 // General pts/dts strategy:
 692 //
 693 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
 694 // bit tricky. We assume first of all that the frame rate never goes _above_
 695 // MAX_FPS, which gives us a frame period N. The decoder can always decode
 696 // in at least this speed, as long at dts <= pts (the frame is not attempted
 697 // presented before it is decoded). Furthermore, we never have longer chains of
 698 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
 699 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
 700 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
 701 // frame, with an order O <= C.)
 702 //
 703 // Many strategies are possible, but we establish these rules:
 704 //
 705 //  - Tip frames have dts = pts - (C-O)*N.
 706 //  - Non-tip frames have dts = dts_last + N.
 707 //
 708 // An example, with C=2 and N=10 and the data flow showed with arrows:
 709 //
 710 //        I  B  P  B  B  P
 711 //   pts: 30 40 50 60 70 80
 712 //        ↓  ↓     ↓
 713 //   dts: 10 30 20 60 50←40
 714 //         |  |  ↑        ↑
 715 //         `--|--'        |
 716 //             `----------'
 717 //
 718 // To show that this works fine also with irregular spacings, let's say that
 719 // the third frame is delayed a bit (something earlier was dropped). Now the
 720 // situation looks like this:
 721 //
 722 //        I  B  P  B  B   P
 723 //   pts: 30 40 80 90 100 110
 724 //        ↓  ↓     ↓
 725 //   dts: 10 30 20 90 50←40
 726 //         |  |  ↑        ↑
 727 //         `--|--'        |
 728 //             `----------'
 729 //
 730 // The resetting on every tip frame makes sure dts never ends up lagging a lot
 731 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
 732 //
 733 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
 734 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
 735 // a tip frame and should be given a dts based on the previous one.
 736 #define FRAME_P 0
 737 #define FRAME_B 1
 738 #define FRAME_I 2
 739 #define FRAME_IDR 7
 740 void encoding2display_order(
 741     int encoding_order, int intra_period,
 742     int intra_idr_period, int ip_period,
 743     int *displaying_order,
 744     int *frame_type, int *pts_lag)
 745 {
 746     int encoding_order_gop = 0;
 747
 748     *pts_lag = 0;
 749
 750     if (intra_period == 1) { /* all are I/IDR frames */
 751         *displaying_order = encoding_order;
 752         if (intra_idr_period == 0)
 753             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
 754         else
 755             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
 756         return;
 757     }
 758
 759     if (intra_period == 0)
 760         intra_idr_period = 0;
 761
 762     if (ip_period == 1) {
 763         // No B-frames, sequence is like IDR PPPPP IPPPPP.
 764         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
 765         *displaying_order = encoding_order;
 766
 767         if (encoding_order_gop == 0) { /* the first frame */
 768             *frame_type = FRAME_IDR;
 769         } else if (intra_period != 0 && /* have I frames */
 770                    encoding_order_gop >= 2 &&
 771                    (encoding_order_gop % intra_period == 0)) {
 772             *frame_type = FRAME_I;
 773         } else {
 774             *frame_type = FRAME_P;
 775         }
 776         return;
 777     }
 778
 779     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
 780     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
 781     *pts_lag = -1;  // Most frames are not tip frames.
 782
 783     if (encoding_order_gop == 0) { /* the first frame */
 784         *frame_type = FRAME_IDR;
 785         *displaying_order = encoding_order;
 786         // IDR frames are a special case; I honestly can't find the logic behind
 787         // why this is the right thing, but it seems to line up nicely in practice :-)
 788         *pts_lag = TIMEBASE / MAX_FPS;
 789     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
 790         *frame_type = FRAME_B;
 791         *displaying_order = encoding_order - 1;
 792         if ((encoding_order_gop % ip_period) == 0) {
 793             *pts_lag = 0;  // Last B-frame.
 794         }
 795     } else if (intra_period != 0 && /* have I frames */
 796                encoding_order_gop >= 2 &&
 797                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
 798         *frame_type = FRAME_I;
 799         *displaying_order = encoding_order + ip_period - 1;
 800     } else {
 801         *frame_type = FRAME_P;
 802         *displaying_order = encoding_order + ip_period - 1;
 803     }
 804 }
 805
 806
 807 static const char *rc_to_string(int rc_mode)
 808 {
 809     switch (rc_mode) {
 810     case VA_RC_NONE:
 811         return "NONE";
 812     case VA_RC_CBR:
 813         return "CBR";
 814     case VA_RC_VBR:
 815         return "VBR";
 816     case VA_RC_VCM:
 817         return "VCM";
 818     case VA_RC_CQP:
 819         return "CQP";
 820     case VA_RC_VBR_CONSTRAINED:
 821         return "VBR_CONSTRAINED";
 822     default:
 823         return "Unknown";
 824     }
 825 }
 826
 827 #if 0
 828 static int process_cmdline(int argc, char *argv[])
 829 {
 830     char c;
 831     const struct option long_opts[] = {
 832         {"help", no_argument, NULL, 0 },
 833         {"bitrate", required_argument, NULL, 1 },
 834         {"minqp", required_argument, NULL, 2 },
 835         {"initialqp", required_argument, NULL, 3 },
 836         {"intra_period", required_argument, NULL, 4 },
 837         {"idr_period", required_argument, NULL, 5 },
 838         {"ip_period", required_argument, NULL, 6 },
 839         {"rcmode", required_argument, NULL, 7 },
 840         {"srcyuv", required_argument, NULL, 9 },
 841         {"recyuv", required_argument, NULL, 10 },
 842         {"fourcc", required_argument, NULL, 11 },
 843         {"syncmode", no_argument, NULL, 12 },
 844         {"enablePSNR", no_argument, NULL, 13 },
 845         {"prit", required_argument, NULL, 14 },
 846         {"priv", required_argument, NULL, 15 },
 847         {"framecount", required_argument, NULL, 16 },
 848         {"entropy", required_argument, NULL, 17 },
 849         {"profile", required_argument, NULL, 18 },
 850         {NULL, no_argument, NULL, 0 }};
 851     int long_index;
 852
 853     while ((c =getopt_long_only(argc, argv, "w:h:n:f:o:?", long_opts, &long_index)) != EOF) {
 854         switch (c) {
 855         case 'w':
 856             frame_width = atoi(optarg);
 857             break;
 858         case 'h':
 859             frame_height = atoi(optarg);
 860             break;
 861         case 'n':
 862         case 'f':
 863             frame_rate = atoi(optarg);
 864             break;
 865         case 'o':
 866             coded_fn = strdup(optarg);
 867             break;
 868         case 0:
 869             print_help();
 870             exit(0);
 871         case 1:
 872             frame_bitrate = atoi(optarg);
 873             break;
 874         case 2:
 875             minimal_qp = atoi(optarg);
 876             break;
 877         case 3:
 878             initial_qp = atoi(optarg);
 879             break;
 880         case 4:
 881             intra_period = atoi(optarg);
 882             break;
 883         case 5:
 884             intra_idr_period = atoi(optarg);
 885             break;
 886         case 6:
 887             ip_period = atoi(optarg);
 888             break;
 889         case 7:
 890             rc_mode = string_to_rc(optarg);
 891             if (rc_mode < 0) {
 892                 print_help();
 893                 exit(1);
 894             }
 895             break;
 896         case 9:
 897             srcyuv_fn = strdup(optarg);
 898             break;
 899         case 11:
 900             srcyuv_fourcc = string_to_fourcc(optarg);
 901             if (srcyuv_fourcc <= 0) {
 902                 print_help();
 903                 exit(1);
 904             }
 905             break;
 906         case 13:
 907             calc_psnr = 1;
 908             break;
 909         case 17:
 910             h264_entropy_mode = atoi(optarg) ? 1: 0;
 911             break;
 912         case 18:
 913             if (strncmp(optarg, "BP", 2) == 0)
 914                 h264_profile = VAProfileH264Baseline;
 915             else if (strncmp(optarg, "MP", 2) == 0)
 916                 h264_profile = VAProfileH264Main;
 917             else if (strncmp(optarg, "HP", 2) == 0)
 918                 h264_profile = VAProfileH264High;
 919             else
 920                 h264_profile = (VAProfile)0;
 921             break;
 922         case ':':
 923         case '?':
 924             print_help();
 925             exit(0);
 926         }
 927     }
 928
 929     if (ip_period < 1) {
 930         printf(" ip_period must be greater than 0\n");
 931         exit(0);
 932     }
 933     if (intra_period != 1 && intra_period % ip_period != 0) {
 934         printf(" intra_period must be a multiplier of ip_period\n");
 935         exit(0);
 936     }
 937     if (intra_period != 0 && intra_idr_period % intra_period != 0) {
 938         printf(" intra_idr_period must be a multiplier of intra_period\n");
 939         exit(0);
 940     }
 941
 942     if (frame_bitrate == 0)
 943         frame_bitrate = frame_width * frame_height * 12 * MAX_FPS / 50;
 944
 945     if (coded_fn == NULL) {
 946         struct stat buf;
 947         if (stat("/tmp", &buf) == 0)
 948             coded_fn = strdup("/tmp/test.264");
 949         else if (stat("/sdcard", &buf) == 0)
 950             coded_fn = strdup("/sdcard/test.264");
 951         else
 952             coded_fn = strdup("./test.264");
 953     }
 954
 955
 956     frame_width_mbaligned = (frame_width + 15) & (~15);
 957     frame_height_mbaligned = (frame_height + 15) & (~15);
 958     if (frame_width != frame_width_mbaligned ||
 959         frame_height != frame_height_mbaligned) {
 960         printf("Source frame is %dx%d and will code clip to %dx%d with crop\n",
 961                frame_width, frame_height,
 962                frame_width_mbaligned, frame_height_mbaligned
 963                );
 964     }
 965
 966     return 0;
 967 }
 968 #endif
 969
 970 VADisplay H264EncoderImpl::va_open_display(void)
 971 {
 972     x11_display = XOpenDisplay(NULL);
 973     if (!x11_display) {
 974         fprintf(stderr, "error: can't connect to X server!\n");
 975         return NULL;
 976     }
 977     return vaGetDisplay(x11_display);
 978 }
 979
 980 void H264EncoderImpl::va_close_display(VADisplay va_dpy)
 981 {
 982     if (!x11_display)
 983         return;
 984
 985     if (x11_window) {
 986         XUnmapWindow(x11_display, x11_window);
 987         XDestroyWindow(x11_display, x11_window);
 988         x11_window = None;
 989     }
 990     XCloseDisplay(x11_display);
 991     x11_display = NULL;
 992 }
 993
 994 int H264EncoderImpl::init_va()
 995 {
 996     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
 997     VAEntrypoint *entrypoints;
 998     int num_entrypoints, slice_entrypoint;
 999     int support_encode = 0;
1000     int major_ver, minor_ver;
1001     VAStatus va_status;
1002     unsigned int i;
1003
1004     va_dpy = va_open_display();
1005     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1006     CHECK_VASTATUS(va_status, "vaInitialize");
1007
1008     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1009     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1010     if (!entrypoints) {
1011         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1012         exit(1);
1013     }
1014
1015     /* use the highest profile */
1016     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1017         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1018             continue;
1019
1020         h264_profile = profile_list[i];
1021         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1022         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1023             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1024                 support_encode = 1;
1025                 break;
1026             }
1027         }
1028         if (support_encode == 1)
1029             break;
1030     }
1031
1032     if (support_encode == 0) {
1033         printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
1034         exit(1);
1035     } else {
1036         switch (h264_profile) {
1037             case VAProfileH264Baseline:
1038                 ip_period = 1;
1039                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1040                 h264_entropy_mode = 0;
1041                 break;
1042             case VAProfileH264ConstrainedBaseline:
1043                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1044                 ip_period = 1;
1045                 break;
1046
1047             case VAProfileH264Main:
1048                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1049                 break;
1050
1051             case VAProfileH264High:
1052                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1053                 break;
1054             default:
1055                 h264_profile = VAProfileH264Baseline;
1056                 ip_period = 1;
1057                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1058                 break;
1059         }
1060     }
1061
1062     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1063
1064     /* find out the format for the render target, and rate control mode */
1065     for (i = 0; i < VAConfigAttribTypeMax; i++)
1066         attrib[i].type = (VAConfigAttribType)i;
1067
1068     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1069                                       &attrib[0], VAConfigAttribTypeMax);
1070     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1071     /* check the interested configattrib */
1072     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1073         printf("Not find desired YUV420 RT format\n");
1074         exit(1);
1075     } else {
1076         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1077         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1078         config_attrib_num++;
1079     }
1080
1081     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1082         int tmp = attrib[VAConfigAttribRateControl].value;
1083
1084         if (rc_mode == -1 || !(rc_mode & tmp))  {
1085             if (rc_mode != -1) {
1086                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1087             }
1088
1089             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1090                 if (rc_default_modes[i] & tmp) {
1091                     rc_mode = rc_default_modes[i];
1092                     break;
1093                 }
1094             }
1095         }
1096
1097         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1098         config_attrib[config_attrib_num].value = rc_mode;
1099         config_attrib_num++;
1100     }
1101
1102
1103     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1104         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1105
1106         h264_packedheader = 1;
1107         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1108         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1109
1110         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1111             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1112         }
1113
1114         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1115             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1116         }
1117
1118         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1119             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1120         }
1121
1122         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1123             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1124         }
1125
1126         enc_packed_header_idx = config_attrib_num;
1127         config_attrib_num++;
1128     }
1129
1130     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1131         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1132         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1133         config_attrib_num++;
1134     }
1135
1136     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1137         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1138     }
1139
1140     free(entrypoints);
1141     return 0;
1142 }
1143
1144 int H264EncoderImpl::setup_encode()
1145 {
1146     VAStatus va_status;
1147     VASurfaceID *tmp_surfaceid;
1148     int codedbuf_size, i;
1149     static VASurfaceID src_surface[SURFACE_NUM];
1150     static VASurfaceID ref_surface[SURFACE_NUM];
1151
1152     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1153             &config_attrib[0], config_attrib_num, &config_id);
1154     CHECK_VASTATUS(va_status, "vaCreateConfig");
1155
1156     /* create source surfaces */
1157     va_status = vaCreateSurfaces(va_dpy,
1158                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1159                                  &src_surface[0], SURFACE_NUM,
1160                                  NULL, 0);
1161     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1162
1163     /* create reference surfaces */
1164     va_status = vaCreateSurfaces(va_dpy,
1165                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1166                                  &ref_surface[0], SURFACE_NUM,
1167                                  NULL, 0);
1168     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1169
1170     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1171     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1172     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1173
1174     /* Create a context for this encode pipe */
1175     va_status = vaCreateContext(va_dpy, config_id,
1176                                 frame_width_mbaligned, frame_height_mbaligned,
1177                                 VA_PROGRESSIVE,
1178                                 tmp_surfaceid, 2 * SURFACE_NUM,
1179                                 &context_id);
1180     CHECK_VASTATUS(va_status, "vaCreateContext");
1181     free(tmp_surfaceid);
1182
1183     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1184
1185     for (i = 0; i < SURFACE_NUM; i++) {
1186         /* create coded buffer once for all
1187          * other VA buffers which won't be used again after vaRenderPicture.
1188          * so APP can always vaCreateBuffer for every frame
1189          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1190          * so VA won't maintain the coded buffer
1191          */
1192         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1193                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1194         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1195     }
1196
1197     /* create OpenGL objects */
1198     //glGenFramebuffers(SURFACE_NUM, fbos);
1199
1200     for (i = 0; i < SURFACE_NUM; i++) {
1201         glGenTextures(1, &gl_surfaces[i].y_tex);
1202         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1203     }
1204
1205     for (i = 0; i < SURFACE_NUM; i++) {
1206         gl_surfaces[i].src_surface = src_surface[i];
1207         gl_surfaces[i].ref_surface = ref_surface[i];
1208     }
1209
1210     return 0;
1211 }
1212
1213
1214
1215 #define partition(ref, field, key, ascending)   \
1216     while (i <= j) {                            \
1217         if (ascending) {                        \
1218             while (ref[i].field < key)          \
1219                 i++;                            \
1220             while (ref[j].field > key)          \
1221                 j--;                            \
1222         } else {                                \
1223             while (ref[i].field > key)          \
1224                 i++;                            \
1225             while (ref[j].field < key)          \
1226                 j--;                            \
1227         }                                       \
1228         if (i <= j) {                           \
1229             tmp = ref[i];                       \
1230             ref[i] = ref[j];                    \
1231             ref[j] = tmp;                       \
1232             i++;                                \
1233             j--;                                \
1234         }                                       \
1235     }                                           \
1236
1237 static void sort_one(VAPictureH264 ref[], int left, int right,
1238                      int ascending, int frame_idx)
1239 {
1240     int i = left, j = right;
1241     unsigned int key;
1242     VAPictureH264 tmp;
1243
1244     if (frame_idx) {
1245         key = ref[(left + right) / 2].frame_idx;
1246         partition(ref, frame_idx, key, ascending);
1247     } else {
1248         key = ref[(left + right) / 2].TopFieldOrderCnt;
1249         partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
1250     }
1251
1252     /* recursion */
1253     if (left < j)
1254         sort_one(ref, left, j, ascending, frame_idx);
1255
1256     if (i < right)
1257         sort_one(ref, i, right, ascending, frame_idx);
1258 }
1259
1260 static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
1261                      int partition_ascending, int list0_ascending, int list1_ascending)
1262 {
1263     int i = left, j = right;
1264     VAPictureH264 tmp;
1265
1266     if (frame_idx) {
1267         partition(ref, frame_idx, key, partition_ascending);
1268     } else {
1269         partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
1270     }
1271
1272
1273     sort_one(ref, left, i-1, list0_ascending, frame_idx);
1274     sort_one(ref, j+1, right, list1_ascending, frame_idx);
1275 }
1276
1277 void H264EncoderImpl::update_ReferenceFrames(int frame_type)
1278 {
1279     int i;
1280
1281     if (frame_type == FRAME_B)
1282         return;
1283
1284     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1285     numShortTerm++;
1286     if (numShortTerm > num_ref_frames)
1287         numShortTerm = num_ref_frames;
1288     for (i=numShortTerm-1; i>0; i--)
1289         ReferenceFrames[i] = ReferenceFrames[i-1];
1290     ReferenceFrames[0] = CurrentCurrPic;
1291
1292     current_frame_num++;
1293     if (current_frame_num > MaxFrameNum)
1294         current_frame_num = 0;
1295 }
1296
1297
1298 int H264EncoderImpl::update_RefPicList(int frame_type)
1299 {
1300     unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
1301
1302     if (frame_type == FRAME_P) {
1303         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1304         sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
1305     }
1306
1307     if (frame_type == FRAME_B) {
1308         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1309         sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
1310                  1, 0, 1);
1311
1312         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1313         sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
1314                  0, 1, 0);
1315     }
1316
1317     return 0;
1318 }
1319
1320
1321 int H264EncoderImpl::render_sequence()
1322 {
1323     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1324     VAStatus va_status;
1325     VAEncMiscParameterBuffer *misc_param;
1326     VAEncMiscParameterRateControl *misc_rate_ctrl;
1327
1328     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1329     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1330     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1331     seq_param.bits_per_second = frame_bitrate;
1332
1333     seq_param.intra_period = intra_period;
1334     seq_param.intra_idr_period = intra_idr_period;
1335     seq_param.ip_period = ip_period;
1336
1337     seq_param.max_num_ref_frames = num_ref_frames;
1338     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1339     seq_param.time_scale = TIMEBASE * 2;
1340     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1341     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1342     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1343     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1344     seq_param.seq_fields.bits.chroma_format_idc = 1;
1345     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1346
1347     if (frame_width != frame_width_mbaligned ||
1348         frame_height != frame_height_mbaligned) {
1349         seq_param.frame_cropping_flag = 1;
1350         seq_param.frame_crop_left_offset = 0;
1351         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1352         seq_param.frame_crop_top_offset = 0;
1353         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1354     }
1355
1356     va_status = vaCreateBuffer(va_dpy, context_id,
1357                                VAEncSequenceParameterBufferType,
1358                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1359     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1360
1361     va_status = vaCreateBuffer(va_dpy, context_id,
1362                                VAEncMiscParameterBufferType,
1363                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1364                                1, NULL, &rc_param_buf);
1365     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1366
1367     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1368     misc_param->type = VAEncMiscParameterTypeRateControl;
1369     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1370     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1371     misc_rate_ctrl->bits_per_second = frame_bitrate;
1372     misc_rate_ctrl->target_percentage = 66;
1373     misc_rate_ctrl->window_size = 1000;
1374     misc_rate_ctrl->initial_qp = initial_qp;
1375     misc_rate_ctrl->min_qp = minimal_qp;
1376     misc_rate_ctrl->basic_unit_size = 0;
1377     vaUnmapBuffer(va_dpy, rc_param_buf);
1378
1379     render_id[0] = seq_param_buf;
1380     render_id[1] = rc_param_buf;
1381
1382     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1383
1384     return 0;
1385 }
1386
1387 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1388 {
1389     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1390     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1391     int PicOrderCntMsb, TopFieldOrderCnt;
1392
1393     if (frame_type == FRAME_IDR)
1394         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1395     else {
1396         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1397         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1398     }
1399
1400     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1401         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1402         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1403     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1404              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1405         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1406     else
1407         PicOrderCntMsb = prevPicOrderCntMsb;
1408
1409     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1410
1411     if (frame_type != FRAME_B) {
1412         PicOrderCntMsb_ref = PicOrderCntMsb;
1413         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1414     }
1415
1416     return TopFieldOrderCnt;
1417 }
1418
1419 int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1420 {
1421     VABufferID pic_param_buf;
1422     VAStatus va_status;
1423     int i = 0;
1424
1425     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1426     pic_param.CurrPic.frame_idx = current_frame_num;
1427     pic_param.CurrPic.flags = 0;
1428     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1429     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1430     CurrentCurrPic = pic_param.CurrPic;
1431
1432     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1433     for (i = numShortTerm; i < SURFACE_NUM; i++) {
1434         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1435         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1436     }
1437
1438     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1439     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1440     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1441     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1442     pic_param.frame_num = current_frame_num;
1443     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1444     pic_param.last_picture = false;  // FIXME
1445     pic_param.pic_init_qp = initial_qp;
1446
1447     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1448                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1449     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1450
1451     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1452
1453     return 0;
1454 }
1455
1456 int H264EncoderImpl::render_packedsequence()
1457 {
1458     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1459     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1460     unsigned int length_in_bits;
1461     unsigned char *packedseq_buffer = NULL;
1462     VAStatus va_status;
1463
1464     length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1465
1466     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1467
1468     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1469     packedheader_param_buffer.has_emulation_bytes = 0;
1470     va_status = vaCreateBuffer(va_dpy,
1471                                context_id,
1472                                VAEncPackedHeaderParameterBufferType,
1473                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1474                                &packedseq_para_bufid);
1475     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1476
1477     va_status = vaCreateBuffer(va_dpy,
1478                                context_id,
1479                                VAEncPackedHeaderDataBufferType,
1480                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1481                                &packedseq_data_bufid);
1482     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1483
1484     render_id[0] = packedseq_para_bufid;
1485     render_id[1] = packedseq_data_bufid;
1486     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1487
1488     free(packedseq_buffer);
1489
1490     return 0;
1491 }
1492
1493
1494 int H264EncoderImpl::render_packedpicture()
1495 {
1496     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1497     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1498     unsigned int length_in_bits;
1499     unsigned char *packedpic_buffer = NULL;
1500     VAStatus va_status;
1501
1502     length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1503     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1504     packedheader_param_buffer.bit_length = length_in_bits;
1505     packedheader_param_buffer.has_emulation_bytes = 0;
1506
1507     va_status = vaCreateBuffer(va_dpy,
1508                                context_id,
1509                                VAEncPackedHeaderParameterBufferType,
1510                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1511                                &packedpic_para_bufid);
1512     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1513
1514     va_status = vaCreateBuffer(va_dpy,
1515                                context_id,
1516                                VAEncPackedHeaderDataBufferType,
1517                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1518                                &packedpic_data_bufid);
1519     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1520
1521     render_id[0] = packedpic_para_bufid;
1522     render_id[1] = packedpic_data_bufid;
1523     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1524
1525     free(packedpic_buffer);
1526
1527     return 0;
1528 }
1529
1530 void H264EncoderImpl::render_packedslice()
1531 {
1532     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1533     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1534     unsigned int length_in_bits;
1535     unsigned char *packedslice_buffer = NULL;
1536     VAStatus va_status;
1537
1538     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1539     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1540     packedheader_param_buffer.bit_length = length_in_bits;
1541     packedheader_param_buffer.has_emulation_bytes = 0;
1542
1543     va_status = vaCreateBuffer(va_dpy,
1544                                context_id,
1545                                VAEncPackedHeaderParameterBufferType,
1546                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1547                                &packedslice_para_bufid);
1548     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1549
1550     va_status = vaCreateBuffer(va_dpy,
1551                                context_id,
1552                                VAEncPackedHeaderDataBufferType,
1553                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1554                                &packedslice_data_bufid);
1555     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1556
1557     render_id[0] = packedslice_para_bufid;
1558     render_id[1] = packedslice_data_bufid;
1559     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1560
1561     free(packedslice_buffer);
1562 }
1563
1564 int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1565 {
1566     VABufferID slice_param_buf;
1567     VAStatus va_status;
1568     int i;
1569
1570     update_RefPicList(frame_type);
1571
1572     /* one frame, one slice */
1573     slice_param.macroblock_address = 0;
1574     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1575     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1576     if (frame_type == FRAME_IDR) {
1577         if (encoding_frame_num != 0)
1578             ++slice_param.idr_pic_id;
1579     } else if (frame_type == FRAME_P) {
1580         int refpiclist0_max = h264_maxref & 0xffff;
1581         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1582
1583         for (i = refpiclist0_max; i < 32; i++) {
1584             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1585             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1586         }
1587     } else if (frame_type == FRAME_B) {
1588         int refpiclist0_max = h264_maxref & 0xffff;
1589         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1590
1591         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1592         for (i = refpiclist0_max; i < 32; i++) {
1593             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1594             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1595         }
1596
1597         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1598         for (i = refpiclist1_max; i < 32; i++) {
1599             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1600             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1601         }
1602     }
1603
1604     slice_param.slice_alpha_c0_offset_div2 = 0;
1605     slice_param.slice_beta_offset_div2 = 0;
1606     slice_param.direct_spatial_mv_pred_flag = 1;
1607     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1608
1609
1610     if (h264_packedheader &&
1611         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1612         render_packedslice();
1613
1614     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1615                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1616     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1617
1618     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1619
1620     return 0;
1621 }
1622
1623
1624
1625 void H264EncoderImpl::save_codeddata(storage_task task)
1626 {
1627     VACodedBufferSegment *buf_list = NULL;
1628     VAStatus va_status;
1629
1630     string data;
1631
1632     const int64_t global_delay = (ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1633
1634     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1635     CHECK_VASTATUS(va_status, "vaMapBuffer");
1636     while (buf_list != NULL) {
1637         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1638         buf_list = (VACodedBufferSegment *) buf_list->next;
1639     }
1640     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1641
1642     {
1643         // Add video.
1644         AVPacket pkt;
1645         memset(&pkt, 0, sizeof(pkt));
1646         pkt.buf = nullptr;
1647         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1648         pkt.size = data.size();
1649         pkt.stream_index = 0;
1650         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1651             pkt.flags = AV_PKT_FLAG_KEY;
1652         } else {
1653             pkt.flags = 0;
1654         }
1655         //pkt.duration = 1;
1656         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1657     }
1658     // Encode and add all audio frames up to and including the pts of this video frame.
1659     for ( ;; ) {
1660         int64_t audio_pts;
1661         vector<float> audio;
1662         {
1663              unique_lock<mutex> lock(frame_queue_mutex);
1664              frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1665              if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1666              auto it = pending_audio_frames.begin();
1667              if (it->first > task.pts) break;
1668              audio_pts = it->first;
1669              audio = move(it->second);
1670              pending_audio_frames.erase(it);
1671         }
1672
1673         AVFrame *frame = avcodec_alloc_frame();
1674         frame->nb_samples = audio.size() / 2;
1675         frame->format = AV_SAMPLE_FMT_S32;
1676         frame->channel_layout = AV_CH_LAYOUT_STEREO;
1677
1678         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1679         int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
1680         if (ret < 0) {
1681             fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1682             exit(1);
1683         }
1684         for (int i = 0; i < frame->nb_samples * 2; ++i) {
1685             if (audio[i] >= 1.0f) {
1686                 int_samples[i] = 2147483647;
1687             } else if (audio[i] <= -1.0f) {
1688                 int_samples[i] = -2147483647;
1689             } else {
1690                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1691             }
1692         }
1693
1694         AVPacket pkt;
1695         av_init_packet(&pkt);
1696         pkt.data = nullptr;
1697         pkt.size = 0;
1698         int got_output;
1699         avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
1700         if (got_output) {
1701             pkt.stream_index = 1;
1702             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1703         }
1704         // TODO: Delayed frames.
1705         avcodec_free_frame(&frame);
1706         av_free_packet(&pkt);
1707         if (audio_pts == task.pts) break;
1708     }
1709
1710 #if 0
1711     printf("\r      "); /* return back to startpoint */
1712     switch (encode_order % 4) {
1713         case 0:
1714             printf("|");
1715             break;
1716         case 1:
1717             printf("/");
1718             break;
1719         case 2:
1720             printf("-");
1721             break;
1722         case 3:
1723             printf("\\");
1724             break;
1725     }
1726     printf("%08lld", encode_order);
1727 #endif
1728 }
1729
1730
1731 // this is weird. but it seems to put a new frame onto the queue
1732 void H264EncoderImpl::storage_task_enqueue(storage_task task)
1733 {
1734         unique_lock<mutex> lock(storage_task_queue_mutex);
1735         storage_task_queue.push(move(task));
1736         srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1737         storage_task_queue_changed.notify_all();
1738 }
1739
1740 void H264EncoderImpl::storage_task_thread()
1741 {
1742         for ( ;; ) {
1743                 storage_task current;
1744                 {
1745                         // wait until there's an encoded frame
1746                         unique_lock<mutex> lock(storage_task_queue_mutex);
1747                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1748                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1749                         current = move(storage_task_queue.front());
1750                         storage_task_queue.pop();
1751                 }
1752
1753                 VAStatus va_status;
1754
1755                 // waits for data, then saves it to disk.
1756                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1757                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1758                 save_codeddata(move(current));
1759
1760                 {
1761                         unique_lock<mutex> lock(storage_task_queue_mutex);
1762                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1763                         storage_task_queue_changed.notify_all();
1764                 }
1765         }
1766 }
1767
1768 int H264EncoderImpl::release_encode()
1769 {
1770     int i;
1771
1772     for (i = 0; i < SURFACE_NUM; i++) {
1773         vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1774         vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1775         vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1776     }
1777
1778     vaDestroyContext(va_dpy, context_id);
1779     vaDestroyConfig(va_dpy, config_id);
1780
1781     return 0;
1782 }
1783
1784 int H264EncoderImpl::deinit_va()
1785 {
1786     vaTerminate(va_dpy);
1787
1788     va_close_display(va_dpy);
1789
1790     return 0;
1791 }
1792
1793
1794 H264EncoderImpl::H264EncoderImpl(QSurface *surface, int width, int height, HTTPD *httpd)
1795         : current_storage_frame(0), surface(surface), httpd(httpd)
1796 {
1797         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1798         context_audio = avcodec_alloc_context3(codec_audio);
1799         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1800         context_audio->sample_rate = OUTPUT_FREQUENCY;
1801         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1802         context_audio->channels = 2;
1803         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1804         context_audio->time_base = AVRational{1, TIMEBASE};
1805         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1806                 fprintf(stderr, "Could not open codec\n");
1807                 exit(1);
1808         }
1809
1810         frame_width = width;
1811         frame_height = height;
1812         frame_width_mbaligned = (frame_width + 15) & (~15);
1813         frame_height_mbaligned = (frame_height + 15) & (~15);
1814
1815         //print_input();
1816
1817         init_va();
1818         setup_encode();
1819
1820         // No frames are ready yet.
1821         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1822
1823         memset(&seq_param, 0, sizeof(seq_param));
1824         memset(&pic_param, 0, sizeof(pic_param));
1825         memset(&slice_param, 0, sizeof(slice_param));
1826
1827         storage_thread = thread(&H264EncoderImpl::storage_task_thread, this);
1828
1829         encode_thread = thread([this]{
1830                 //SDL_GL_MakeCurrent(window, context);
1831                 QOpenGLContext *context = create_context(this->surface);
1832                 eglBindAPI(EGL_OPENGL_API);
1833                 if (!make_current(context, this->surface)) {
1834                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1835                                 eglGetError());
1836                         exit(1);
1837                 }
1838                 encode_thread_func();
1839         });
1840 }
1841
1842 H264EncoderImpl::~H264EncoderImpl()
1843 {
1844         {
1845                 unique_lock<mutex> lock(frame_queue_mutex);
1846                 encode_thread_should_quit = true;
1847                 frame_queue_nonempty.notify_all();
1848         }
1849         encode_thread.join();
1850         {
1851                 unique_lock<mutex> lock(storage_task_queue_mutex);
1852                 storage_thread_should_quit = true;
1853                 frame_queue_nonempty.notify_all();
1854                 storage_task_queue_changed.notify_all();
1855         }
1856         storage_thread.join();
1857
1858         release_encode();
1859         deinit_va();
1860 }
1861
1862 bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1863 {
1864         {
1865                 // Wait until this frame slot is done encoding.
1866                 unique_lock<mutex> lock(storage_task_queue_mutex);
1867                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1868                 if (storage_thread_should_quit) return false;
1869         }
1870
1871         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1872         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1873         *y_tex = surf->y_tex;
1874         *cbcr_tex = surf->cbcr_tex;
1875
1876         VASurfaceID surface = surf->src_surface;
1877         VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
1878         CHECK_VASTATUS(va_status, "vaDeriveImage");
1879
1880         VABufferInfo buf_info;
1881         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1882         va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1883         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1884
1885         // Create Y image.
1886         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1887         EGLint y_attribs[] = {
1888                 EGL_WIDTH, frame_width,
1889                 EGL_HEIGHT, frame_height,
1890                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1891                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1892                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1893                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1894                 EGL_NONE
1895         };
1896
1897         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1898         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1899
1900         // Associate Y image to a texture.
1901         glBindTexture(GL_TEXTURE_2D, *y_tex);
1902         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1903
1904         // Create CbCr image.
1905         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1906         EGLint cbcr_attribs[] = {
1907                 EGL_WIDTH, frame_width,
1908                 EGL_HEIGHT, frame_height,
1909                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1910                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1911                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1912                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1913                 EGL_NONE
1914         };
1915
1916         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1917         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1918
1919         // Associate CbCr image to a texture.
1920         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1921         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1922
1923         return true;
1924 }
1925
1926 void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
1927 {
1928         {
1929                 unique_lock<mutex> lock(frame_queue_mutex);
1930                 pending_audio_frames[pts] = move(audio);
1931         }
1932         frame_queue_nonempty.notify_all();
1933 }
1934
1935 void H264EncoderImpl::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
1936 {
1937         {
1938                 unique_lock<mutex> lock(frame_queue_mutex);
1939                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1940                 ++current_storage_frame;
1941         }
1942         frame_queue_nonempty.notify_all();
1943 }
1944
1945 void H264EncoderImpl::encode_thread_func()
1946 {
1947         int64_t last_dts = -1;
1948         int gop_start_display_frame_num = 0;
1949         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1950                 PendingFrame frame;
1951                 int pts_lag;
1952                 int frame_type, display_frame_num;
1953                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1954                                        &display_frame_num, &frame_type, &pts_lag);
1955                 if (frame_type == FRAME_IDR) {
1956                         numShortTerm = 0;
1957                         current_frame_num = 0;
1958                         gop_start_display_frame_num = display_frame_num;
1959                 }
1960
1961                 {
1962                         unique_lock<mutex> lock(frame_queue_mutex);
1963                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1964                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1965                         });
1966                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1967                                 // We have queued frames that were supposed to be B-frames,
1968                                 // but will be no P-frame to encode them against. Encode them all
1969                                 // as P-frames instead. Note that this happens under the mutex,
1970                                 // but nobody else uses it at this point, since we're shutting down,
1971                                 // so there's no contention.
1972                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1973                                 return;
1974                         } else {
1975                                 frame = move(pending_video_frames[display_frame_num]);
1976                                 pending_video_frames.erase(display_frame_num);
1977                         }
1978                 }
1979
1980                 // Determine the dts of this frame.
1981                 int64_t dts;
1982                 if (pts_lag == -1) {
1983                         assert(last_dts != -1);
1984                         dts = last_dts + (TIMEBASE / MAX_FPS);
1985                 } else {
1986                         dts = frame.pts - pts_lag;
1987                 }
1988                 last_dts = dts;
1989
1990                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
1991         }
1992 }
1993
1994 void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1995 {
1996         if (pending_video_frames.empty()) {
1997                 return;
1998         }
1999
2000         for (auto &pending_frame : pending_video_frames) {
2001                 int display_frame_num = pending_frame.first;
2002                 assert(display_frame_num > 0);
2003                 PendingFrame frame = move(pending_frame.second);
2004                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2005                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2006                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
2007                 last_dts = dts;
2008         }
2009 }
2010
2011 void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2012                                    int frame_type, int64_t pts, int64_t dts)
2013 {
2014         // Wait for the GPU to be done with the frame.
2015         glClientWaitSync(frame.fence.get(), 0, 0);
2016
2017         // Release back any input frames we needed to render this frame.
2018         frame.input_frames.clear();
2019
2020         // Unmap the image.
2021         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2022         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2023         eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2024         VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2025         CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2026         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2027         CHECK_VASTATUS(va_status, "vaDestroyImage");
2028
2029         VASurfaceID surface = surf->src_surface;
2030
2031         // Schedule the frame for encoding.
2032         va_status = vaBeginPicture(va_dpy, context_id, surface);
2033         CHECK_VASTATUS(va_status, "vaBeginPicture");
2034
2035         if (frame_type == FRAME_IDR) {
2036                 render_sequence();
2037                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2038                 if (h264_packedheader) {
2039                         render_packedsequence();
2040                         render_packedpicture();
2041                 }
2042         } else {
2043                 //render_sequence();
2044                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2045         }
2046         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2047
2048         va_status = vaEndPicture(va_dpy, context_id);
2049         CHECK_VASTATUS(va_status, "vaEndPicture");
2050
2051         // so now the data is done encoding (well, async job kicked off)...
2052         // we send that to the storage thread
2053         storage_task tmp;
2054         tmp.display_order = display_frame_num;
2055         tmp.frame_type = frame_type;
2056         tmp.pts = pts;
2057         tmp.dts = dts;
2058         storage_task_enqueue(move(tmp));
2059
2060         update_ReferenceFrames(frame_type);
2061 }
2062
2063 // Proxy object.
2064 H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
2065         : impl(new H264EncoderImpl(surface, width, height, httpd)) {}
2066
2067 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2068 H264Encoder::~H264Encoder() {}
2069
2070 void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
2071 {
2072         impl->add_audio(pts, audio);
2073 }
2074
2075 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2076 {
2077         return impl->begin_frame(y_tex, cbcr_tex);
2078 }
2079
2080 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
2081 {
2082         impl->end_frame(fence, pts, input_frames);
2083 }
2084
2085 // Real class.