struct x264_t
{
    /* encoder parameters */
    x264_param_t    param;

    x264_t          *thread[X264_THREAD_MAX+1];
    x264_t          *lookahead_thread[X264_LOOKAHEAD_THREAD_MAX];
    int             b_thread_active;
    int             i_thread_phase; /* which thread to use for the next frame */
    int             i_thread_idx;   /* which thread this is */
    int             i_threadslice_start; /* first row in this thread slice */
    int             i_threadslice_end; /* row after the end of this thread slice */
    int             i_threadslice_pass; /* which pass of encoding we are on */
    x264_threadpool_t *threadpool;
    x264_threadpool_t *lookaheadpool;
    x264_pthread_mutex_t mutex;
    x264_pthread_cond_t cv;

    /* bitstream output */
    struct
    {
        int         i_nal;
        int         i_nals_allocated;
        x264_nal_t  *nal;
        int         i_bitstream;    /* size of p_bitstream */
        uint8_t     *p_bitstream;   /* will hold data for all nal */
        bs_t        bs;
    } out;

    uint8_t *nal_buffer;
    int      nal_buffer_size;

    /**** thread synchronization starts here ****/

    /* frame number/poc */
    int             i_frame;
    int             i_frame_num;

    int             i_thread_frames; /* Number of different frames being encoded by threads;
                                      * 1 when sliced-threads is on. */
    int             i_nal_type;
    int             i_nal_ref_idc;

    int64_t         i_disp_fields;  /* Number of displayed fields (both coded and implied via pic_struct) */
    int             i_disp_fields_last_frame;
    int64_t         i_prev_duration; /* Duration of previous frame */
    int64_t         i_coded_fields; /* Number of coded fields (both coded and implied via pic_struct) */
    int64_t         i_cpb_delay;    /* Equal to number of fields preceding this field
                                     * since last buffering_period SEI */
    int64_t         i_coded_fields_lookahead; /* Use separate counters for lookahead */
    int64_t         i_cpb_delay_lookahead;

    int64_t         i_cpb_delay_pir_offset;
    int64_t         i_cpb_delay_pir_offset_next;

    int             b_queued_intra_refresh;
    int64_t         i_last_idr_pts;

    int             i_idr_pic_id;


    /* quantization matrix for decoding, [cqm][qp%6][coef] */
    int             (*dequant4_mf[4])[16];   /* [4][6][16] */
    int             (*dequant8_mf[4])[64];   /* [4][6][64] */
    /* quantization matrix for trellis, [cqm][qp][coef] */
    int             (*unquant4_mf[4])[16];   /* [4][52][16] */
    int             (*unquant8_mf[4])[64];   /* [4][52][64] */
    /* quantization matrix for deadzone */
    udctcoef        (*quant4_mf[4])[16];     /* [4][52][16] */
    udctcoef        (*quant8_mf[4])[64];     /* [4][52][64] */
    udctcoef        (*quant4_bias[4])[16];   /* [4][52][16] */
    udctcoef        (*quant8_bias[4])[64];   /* [4][52][64] */
    udctcoef        (*quant4_bias0[4])[16];  /* [4][52][16] */
    udctcoef        (*quant8_bias0[4])[64];  /* [4][52][64] */
    udctcoef        (*nr_offset_emergency)[4][64];


    /* mv/ref cost arrays. */
    uint16_t *cost_mv[QP_MAX+1];
    uint16_t *cost_mv_fpel[QP_MAX+1][4];

    const uint8_t   *chroma_qp_table; /* includes both the nonlinear luma->chroma mapping and chroma_qp_offset */

    /* Slice header */
    x264_slice_header_t sh;

    /* SPS / PPS */
    x264_sps_t      sps[1];
    x264_pps_t      pps[1];

    /* Slice header backup, for SEI_DEC_REF_PIC_MARKING */
    int b_sh_backup;
    x264_slice_header_t sh_backup;

    /* cabac context */
    x264_cabac_t    cabac;

    struct
    {
        /* Frames to be encoded (whose types have been decided) */
        x264_frame_t **current;
        /* Unused frames: 0 = fenc, 1 = fdec */
        x264_frame_t **unused[2];

        /* Unused blank frames (for duplicates) */
        x264_frame_t **blank_unused;

        /* frames used for reference + sentinels */
        x264_frame_t *reference[X264_REF_MAX+2];

        int i_last_keyframe;       /* Frame number of the last keyframe */
        int i_last_idr;            /* Frame number of the last IDR (not RP)*/
        int i_poc_last_open_gop;   /* Poc of the I frame of the last open-gop. The value
                                    * is only assigned during the period between that
                                    * I frame and the next P or I frame, else -1 */

        int i_input;    /* Number of input frames already accepted */

        int i_max_dpb;  /* Number of frames allocated in the decoded picture buffer */
        int i_max_ref0;
        int i_max_ref1;
        int i_delay;    /* Number of frames buffered for B reordering */
        int     i_bframe_delay;
        int64_t i_bframe_delay_time;
        int64_t i_first_pts;
        int64_t i_prev_reordered_pts[2];
        int64_t i_largest_pts;
        int64_t i_second_largest_pts;
        int b_have_lowres;  /* Whether 1/2 resolution luma planes are being used */
        int b_have_sub8x8_esa;
    } frames;

    /* current frame being encoded */
    x264_frame_t    *fenc;

    /* frame being reconstructed */
    x264_frame_t    *fdec;

    /* references lists */
    int             i_ref[2];
    x264_frame_t    *fref[2][X264_REF_MAX+3];
    x264_frame_t    *fref_nearest[2];
    int             b_ref_reorder[2];

    /* hrd */
    int initial_cpb_removal_delay;
    int initial_cpb_removal_delay_offset;
    int64_t i_reordered_pts_delay;

    /* Current MB DCT coeffs */
    struct
    {
        ALIGNED_16( dctcoef luma16x16_dc[3][16] );
        ALIGNED_16( dctcoef chroma_dc[2][8] );
        // FIXME share memory?
        ALIGNED_16( dctcoef luma8x8[12][64] );
        ALIGNED_16( dctcoef luma4x4[16*3][16] );
    } dct;

    /* MB table and cache for current frame/mb */
    struct
    {
        int     i_mb_width;
        int     i_mb_height;
        int     i_mb_count;                 /* number of mbs in a frame */

        /* Chroma subsampling */
        int     chroma_h_shift;
        int     chroma_v_shift;

        /* Strides */
        int     i_mb_stride;
        int     i_b8_stride;
        int     i_b4_stride;
        int     left_b8[2];
        int     left_b4[2];

        /* Current index */
        int     i_mb_x;
        int     i_mb_y;
        int     i_mb_xy;
        int     i_b8_xy;
        int     i_b4_xy;

        /* Search parameters */
        int     i_me_method;
        int     i_subpel_refine;
        int     b_chroma_me;
        int     b_trellis;
        int     b_noise_reduction;
        int     b_dct_decimate;
        int     i_psy_rd; /* Psy RD strength--fixed point value*/
        int     i_psy_trellis; /* Psy trellis strength--fixed point value*/

        int     b_interlaced;
        int     b_adaptive_mbaff; /* MBAFF+subme 0 requires non-adaptive MBAFF i.e. all field mbs */

        /* Allowed qpel MV range to stay within the picture + emulated edge pixels */
        int     mv_min[2];
        int     mv_max[2];
        int     mv_miny_row[3]; /* 0 == top progressive, 1 == bot progressive, 2 == interlaced */
        int     mv_maxy_row[3];
        /* Subpel MV range for motion search.
         * same mv_min/max but includes levels' i_mv_range. */
        int     mv_min_spel[2];
        int     mv_max_spel[2];
        int     mv_miny_spel_row[3];
        int     mv_maxy_spel_row[3];
        /* Fullpel MV range for motion search */
        int     mv_min_fpel[2];
        int     mv_max_fpel[2];
        int     mv_miny_fpel_row[3];
        int     mv_maxy_fpel_row[3];

        /* neighboring MBs */
        unsigned int i_neighbour;
        unsigned int i_neighbour8[4];       /* neighbours of each 8x8 or 4x4 block that are available */
        unsigned int i_neighbour4[16];      /* at the time the block is coded */
        unsigned int i_neighbour_intra;     /* for constrained intra pred */
        unsigned int i_neighbour_frame;     /* ignoring slice boundaries */
        int     i_mb_type_top;
        int     i_mb_type_left[2];
        int     i_mb_type_topleft;
        int     i_mb_type_topright;
        int     i_mb_prev_xy;
        int     i_mb_left_xy[2];
        int     i_mb_top_xy;
        int     i_mb_topleft_xy;
        int     i_mb_topright_xy;
        int     i_mb_top_y;
        int     i_mb_topleft_y;
        int     i_mb_topright_y;
        const x264_left_table_t *left_index_table;
        int     i_mb_top_mbpair_xy;
        int     topleft_partition;
        int     b_allow_skip;
        int     field_decoding_flag;

        /**** thread synchronization ends here ****/
        /* subsequent variables are either thread-local or constant,
         * and won't be copied from one thread to another */

        /* mb table */
        int8_t  *type;                      /* mb type */
        uint8_t *partition;                 /* mb partition */
        int8_t  *qp;                        /* mb qp */
        int16_t *cbp;                       /* mb cbp: 0x0?: luma, 0x?0: chroma, 0x100: luma dc, 0x0200 and 0x0400: chroma dc  (all set for PCM)*/
        int8_t  (*intra4x4_pred_mode)[8];   /* intra4x4 pred mode. for non I4x4 set to I_PRED_4x4_DC(2) */
                                            /* actually has only 7 entries; set to 8 for write-combining optimizations */
        uint8_t (*non_zero_count)[16*3];    /* nzc. for I_PCM set to 16 */
        int8_t  *chroma_pred_mode;          /* chroma_pred_mode. cabac only. for non intra I_PRED_CHROMA_DC(0) */
        int16_t (*mv[2])[2];                /* mb mv. set to 0 for intra mb */
        uint8_t (*mvd[2])[8][2];            /* absolute value of mb mv difference with predict, clipped to [0,33]. set to 0 if intra. cabac only */
        int8_t   *ref[2];                   /* mb ref. set to -1 if non used (intra or Lx only) */
        int16_t (*mvr[2][X264_REF_MAX*2])[2];/* 16x16 mv for each possible ref */
        int8_t  *skipbp;                    /* block pattern for SKIP or DIRECT (sub)mbs. B-frames + cabac only */
        int8_t  *mb_transform_size;         /* transform_size_8x8_flag of each mb */
        uint16_t *slice_table;              /* sh->first_mb of the slice that the indexed mb is part of
                                             * NOTE: this will fail on resolutions above 2^16 MBs... */
        uint8_t *field;

         /* buffer for weighted versions of the reference frames */
        pixel *p_weight_buf[X264_REF_MAX];

        /* current value */
        int     i_type;
        int     i_partition;
        ALIGNED_4( uint8_t i_sub_partition[4] );
        int     b_transform_8x8;

        int     i_cbp_luma;
        int     i_cbp_chroma;

        int     i_intra16x16_pred_mode;
        int     i_chroma_pred_mode;

        /* skip flags for i4x4 and i8x8
         * 0 = encode as normal.
         * 1 (non-RD only) = the DCT is still in h->dct, restore fdec and skip reconstruction.
         * 2 (RD only) = the DCT has since been overwritten by RD; restore that too. */
        int i_skip_intra;
        /* skip flag for motion compensation */
        /* if we've already done MC, we don't need to do it again */
        int b_skip_mc;
        /* set to true if we are re-encoding a macroblock. */
        int b_reencode_mb;
        int ip_offset; /* Used by PIR to offset the quantizer of intra-refresh blocks. */
        int b_deblock_rdo;
        int b_overflow; /* If CAVLC had a level code overflow during bitstream writing. */

        struct
        {
            /* space for p_fenc and p_fdec */
#define FENC_STRIDE 16
#define FDEC_STRIDE 32
            ALIGNED_16( pixel fenc_buf[48*FENC_STRIDE] );
            ALIGNED_16( pixel fdec_buf[52*FDEC_STRIDE] );

            /* i4x4 and i8x8 backup data, for skipping the encode stage when possible */
            ALIGNED_16( pixel i4x4_fdec_buf[16*16] );
            ALIGNED_16( pixel i8x8_fdec_buf[16*16] );
            ALIGNED_16( dctcoef i8x8_dct_buf[3][64] );
            ALIGNED_16( dctcoef i4x4_dct_buf[15][16] );
            uint32_t i4x4_nnz_buf[4];
            uint32_t i8x8_nnz_buf[4];
            int i4x4_cbp;
            int i8x8_cbp;

            /* Psy trellis DCT data */
            ALIGNED_16( dctcoef fenc_dct8[4][64] );
            ALIGNED_16( dctcoef fenc_dct4[16][16] );

            /* Psy RD SATD/SA8D scores cache */
            ALIGNED_16( uint64_t fenc_hadamard_cache[9] );
            ALIGNED_16( uint32_t fenc_satd_cache[32] );

            /* pointer over mb of the frame to be compressed */
            pixel *p_fenc[3]; /* y,u,v */
            /* pointer to the actual source frame, not a block copy */
            pixel *p_fenc_plane[3];

            /* pointer over mb of the frame to be reconstructed  */
            pixel *p_fdec[3];

            /* pointer over mb of the references */
            int i_fref[2];
            /* [12]: yN, yH, yV, yHV, (NV12 ? uv : I444 ? (uN, uH, uV, uHV, vN, ...)) */
            pixel *p_fref[2][X264_REF_MAX*2][12];
            pixel *p_fref_w[X264_REF_MAX*2];  /* weighted fullpel luma */
            uint16_t *p_integral[2][X264_REF_MAX];

            /* fref stride */
            int     i_stride[3];
        } pic;

        /* cache */
        struct
        {
            /* real intra4x4_pred_mode if I_4X4 or I_8X8, I_PRED_4x4_DC if mb available, -1 if not */
            ALIGNED_8( int8_t intra4x4_pred_mode[X264_SCAN8_LUMA_SIZE] );

            /* i_non_zero_count if available else 0x80 */
            ALIGNED_16( uint8_t non_zero_count[X264_SCAN8_SIZE] );

            /* -1 if unused, -2 if unavailable */
            ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );

            /* 0 if not available */
            ALIGNED_16( int16_t mv[2][X264_SCAN8_LUMA_SIZE][2] );
            ALIGNED_8( uint8_t mvd[2][X264_SCAN8_LUMA_SIZE][2] );

            /* 1 if SKIP or DIRECT. set only for B-frames + CABAC */
            ALIGNED_4( int8_t skip[X264_SCAN8_LUMA_SIZE] );

            ALIGNED_4( int16_t direct_mv[2][4][2] );
            ALIGNED_4( int8_t  direct_ref[2][4] );
            int     direct_partition;
            ALIGNED_4( int16_t pskip_mv[2] );

            /* number of neighbors (top and left) that used 8x8 dct */
            int     i_neighbour_transform_size;
            int     i_neighbour_skip;

            /* neighbor CBPs */
            int     i_cbp_top;
            int     i_cbp_left;

            /* extra data required for mbaff in mv prediction */
            int16_t topright_mv[2][3][2];
            int8_t  topright_ref[2][3];

            /* current mb deblock strength */
            uint8_t (*deblock_strength)[8][4];
        } cache;

        /* */
        int     i_qp;       /* current qp */
        int     i_chroma_qp;
        int     i_last_qp;  /* last qp */
        int     i_last_dqp; /* last delta qp */
        int     b_variable_qp; /* whether qp is allowed to vary per macroblock */
        int     b_lossless;
        int     b_direct_auto_read; /* take stats for --direct auto from the 2pass log */
        int     b_direct_auto_write; /* analyse direct modes, to use and/or save */

        /* lambda values */
        int     i_trellis_lambda2[2][2]; /* [luma,chroma][inter,intra] */
        int     i_psy_rd_lambda;
        int     i_chroma_lambda2_offset;

        /* B_direct and weighted prediction */
        int16_t dist_scale_factor_buf[2][2][X264_REF_MAX*2][4];
        int16_t (*dist_scale_factor)[4];
        int8_t bipred_weight_buf[2][2][X264_REF_MAX*2][4];
        int8_t (*bipred_weight)[4];
        /* maps fref1[0]'s ref indices into the current list0 */
#define map_col_to_list0(col) h->mb.map_col_to_list0[(col)+2]
        int8_t  map_col_to_list0[X264_REF_MAX+2];
        int ref_blind_dupe; /* The index of the blind reference frame duplicate. */
        int8_t deblock_ref_table[X264_REF_MAX*2+2];
#define deblock_ref_table(x) h->mb.deblock_ref_table[(x)+2]
    } mb;

    /* rate control encoding only */
    x264_ratecontrol_t *rc;

    /* stats */
    struct
    {
        /* Current frame stats */
        x264_frame_stat_t frame;

        /* Cumulated stats */

        /* per slice info */
        int     i_frame_count[3];
        int64_t i_frame_size[3];
        double  f_frame_qp[3];
        int     i_consecutive_bframes[X264_BFRAME_MAX+1];
        /* */
        double  f_ssd_global[3];
        double  f_psnr_average[3];
        double  f_psnr_mean_y[3];
        double  f_psnr_mean_u[3];
        double  f_psnr_mean_v[3];
        double  f_ssim_mean_y[3];
        double  f_frame_duration[3];
        /* */
        int64_t i_mb_count[3][19];
        int64_t i_mb_partition[2][17];
        int64_t i_mb_count_8x8dct[2];
        int64_t i_mb_count_ref[2][2][X264_REF_MAX*2];
        int64_t i_mb_cbp[6];
        int64_t i_mb_pred_mode[4][13];
        int64_t i_mb_field[3];
        /* */
        int     i_direct_score[2];
        int     i_direct_frames[2];
        /* num p-frames weighted */
        int     i_wpred[2];

    } stat;

    /* 0 = luma 4x4, 1 = luma 8x8, 2 = chroma 4x4, 3 = chroma 8x8 */
    udctcoef (*nr_offset)[64];
    uint32_t (*nr_residual_sum)[64];
    uint32_t *nr_count;

    ALIGNED_16( udctcoef nr_offset_denoise[4][64] );
    ALIGNED_16( uint32_t nr_residual_sum_buf[2][4][64] );
    uint32_t nr_count_buf[2][4];

    uint8_t luma2chroma_pixel[7]; /* Subsampled pixel size */

    /* Buffers that are allocated per-thread even in sliced threads. */
    void *scratch_buffer; /* for any temporary storage that doesn't want repeated malloc */
    void *scratch_buffer2; /* if the first one's already in use */
    pixel *intra_border_backup[5][3]; /* bottom pixels of the previous mb row, used for intra prediction after the framebuffer has been deblocked */
    /* Deblock strength values are stored for each 4x4 partition. In MBAFF
     * there are four extra values that need to be stored, located in [4][i]. */
    uint8_t (*deblock_strength[2])[2][8][4];

    /* CPU functions dependents */
    x264_predict_t      predict_16x16[4+3];
    x264_predict8x8_t   predict_8x8[9+3];
    x264_predict_t      predict_4x4[9+3];
    x264_predict_t      predict_chroma[4+3];
    x264_predict_t      predict_8x8c[4+3];
    x264_predict_t      predict_8x16c[4+3];
    x264_predict_8x8_filter_t predict_8x8_filter;

    x264_pixel_function_t pixf;
    x264_mc_functions_t   mc;
    x264_dct_function_t   dctf;
    x264_zigzag_function_t zigzagf;
    x264_zigzag_function_t zigzagf_interlaced;
    x264_zigzag_function_t zigzagf_progressive;
    x264_quant_function_t quantf;
    x264_deblock_function_t loopf;
    x264_bitstream_function_t bsf;

#if HAVE_VISUALIZE
    struct visualize_t *visualize;
#endif
    x264_lookahead_t *lookahead;
};




typedef struct
{
    int i_id;

    int i_profile_idc;
    int i_level_idc;

    int b_constraint_set0;
    int b_constraint_set1;
    int b_constraint_set2;
    int b_constraint_set3;

    int i_log2_max_frame_num;

    int i_poc_type;
    /* poc 0 */
    int i_log2_max_poc_lsb;

    int i_num_ref_frames;
    int b_gaps_in_frame_num_value_allowed;
    int i_mb_width;
    int i_mb_height;
    int b_frame_mbs_only;
    int b_mb_adaptive_frame_field;
    int b_direct8x8_inference;

    int b_crop;
    struct
    {
        int i_left;
        int i_right;
        int i_top;
        int i_bottom;
    } crop;

    int b_vui;
  struct
    {
        int b_aspect_ratio_info_present;
        int i_sar_width;
        int i_sar_height;

        int b_overscan_info_present;
        int b_overscan_info;

        int b_signal_type_present;
        int i_vidformat;
        int b_fullrange;
        int b_color_description_present;
        int i_colorprim;
        int i_transfer;
        int i_colmatrix;

        int b_chroma_loc_info_present;
        int i_chroma_loc_top;
        int i_chroma_loc_bottom;

        int b_timing_info_present;
        uint32_t i_num_units_in_tick;
        uint32_t i_time_scale;
        int b_fixed_frame_rate;

        int b_nal_hrd_parameters_present;
        int b_vcl_hrd_parameters_present;

        struct
        {
            int i_cpb_cnt;
            int i_bit_rate_scale;
            int i_cpb_size_scale;
            int i_bit_rate_value;
            int i_cpb_size_value;
            int i_bit_rate_unscaled;
            int i_cpb_size_unscaled;
            int b_cbr_hrd;

            int i_initial_cpb_removal_delay_length;
            int i_cpb_removal_delay_length;
            int i_dpb_output_delay_length;
            int i_time_offset_length;
        } hrd;

        int b_pic_struct_present;
        int b_bitstream_restriction;
        int b_motion_vectors_over_pic_boundaries;
        int i_max_bytes_per_pic_denom;
        int i_max_bits_per_mb_denom;
        int i_log2_max_mv_length_horizontal;
        int i_log2_max_mv_length_vertical;
        int i_num_reorder_frames;
        int i_max_dec_frame_buffering;

        /* FIXME to complete */
    } vui;

    int b_qpprime_y_zero_transform_bypass;
    int i_chroma_format_idc;

} x264_sps_t;

typedef struct
{
    int i_id;
    int i_sps_id;

    int b_cabac;

    int b_pic_order;
    int i_num_slice_groups;

    int i_num_ref_idx_l0_default_active;
    int i_num_ref_idx_l1_default_active;

    int b_weighted_pred;
    int b_weighted_bipred;

    int i_pic_init_qp;
    int i_pic_init_qs;

    int i_chroma_qp_index_offset;

    int b_deblocking_filter_control;
    int b_constrained_intra_pred;
    int b_redundant_pic_cnt;

    int b_transform_8x8_mode;

    int i_cqm_preset;
    const uint8_t *scaling_list[8]; /* could be 12, but we don't allow separate Cb/Cr lists */

} x264_pps_t;