t264.h

T264是中国的视频编码自由组织合力开发的264编解码程序

    int32_t pic_init_qp_minus26;
    int32_t pic_init_qs_minus26;
    int32_t chroma_qp_index_offset;
    int32_t deblocking_filter_control_present_flag;
} T264_pic_set_t;

typedef struct
{
    int32_t first_mb_in_slice;
    int32_t slice_type;
    int32_t pic_id;
    int32_t frame_num;
    int32_t idr_pic_id;
    /* pic_oder_cnt_type == 0*/
    int32_t pic_order_cnt_lsb;
    int32_t delta_pic_order_cnt_bottom;
    int32_t direct_spatial_mv_pred_flag;
    /*P frame*/
    int32_t num_ref_idx_active_override_flag;
    int32_t slice_qp_delta;
    int32_t ref_pic_list_reordering_flag_l0;
    int32_t no_output_of_prior_pics_flag;
    int32_t long_term_reference_flag;
    int32_t adaptive_ref_pic_marking_mode_flag;
	/* for CABAC */
	int32_t cabac_init_idc;
} T264_slice_t;

typedef struct
{
    int32_t ver;
    int32_t flags;
} T264_custom_set_t;

typedef struct
{
    void (*proc)(T264_t* t, void* data, int32_t state);
    void (*close)(T264_t* t, void* data);
    void* handle;
    int32_t ret;    // ret value, optional
} T264_plugin_t;

typedef struct
{
    int8_t  refno;
    int16_t x;
    int16_t y;
} T264_vector_t;

// this struct should keep 16 bytes align
typedef struct
{
    // intra 4x4 mode
    uint8_t     mode_i4x4[16];
    // inter sub partition size(if no sub partition -1)
    uint8_t     submb_part[16];
    // inter 4x4 block mv
    T264_vector_t vec[2][4 * 4];
	// for CABAC, mv delta
	int16_t		mvd[2][4*4][2];
    // non zero count
    uint8_t     nnz[16 + 4 + 4];
	// sad
	uint32_t    sad;

    // I16x16, I4x4, P_MODE, ....
    uint8_t     mb_mode;
    uint8_t     mb_mode_uv;
    // intra 16x16 mode
    uint8_t     mode_i16x16;
    // inter
    uint8_t     mb_part;
    // rate control
    int8_t      mb_qp_delta; 

	// for CABAC, the following field should be saved in the context
	uint8_t     is_copy;
	uint16_t    cbp_y;
	uint16_t    cbp_c;
	uint16_t	cbp;
    // just pad to 16 bytes aligned
    uint8_t     pad[CACHE_SIZE - 7];
} T264_mb_context_t;

typedef struct
{
    uint8_t*  Y[4], *U, *V;
    int32_t   poc;
    T264_mb_context_t* mb;
    int32_t   qp;
} T264_frame_t;

typedef struct T264_search_context_t
{
    // all candidate
    T264_vector_t* vec;
    int32_t vec_num;
    T264_vector_t vec_best;
    int32_t height;
    int32_t width;
    int8_t limit_x;
    int8_t limit_y;
    int32_t offset;
    int32_t mb_part;
    int32_t list_index;
} T264_search_context_t;

//this is for CABAC

typedef struct
{
	/* model */
	struct
	{
		int i_model;
		int i_cost;
	} slice[3];

	/* context */
	struct
	{
		int i_state;
		int i_mps;
		int i_count;
	} ctxstate[399];

	/* state */
	int i_low;
	int i_range;

	int i_sym_cnt;

	/* bit stream */
	int b_first_bit;
	int i_bits_outstanding;
	void *s;

} T264_cabac_t;

#ifndef _PREDICT_H_1
#define _PREDICT_H_1
typedef void (*T264_predict_16x16_mode_t)(uint8_t* dst, int32_t dst_stride, uint8_t* top, uint8_t* left);
typedef void (*T264_predict_4x4_mode_t)  (uint8_t* dst, int32_t dst_stride, uint8_t* top, uint8_t* left);
typedef void (*T264_predict_8x8_mode_t)  (uint8_t* dst, int32_t dst_stride, uint8_t* top, uint8_t* left);
#endif

typedef void (*expand8to16_t)(uint8_t* src, int32_t src_stride, int32_t quarter_width, int32_t quarter_height, int16_t* dst);
typedef void (*expand8to16sub_t)(uint8_t* pred, int32_t quarter_width, int32_t quarter_height, int16_t* dst, uint8_t* src, int32_t src_stride);
typedef void (*contract16to8_t)(int16_t* src, int32_t quarter_width, int32_t quarter_height, uint8_t* dst, int32_t dst_stride);
typedef void (*contract16to8add_t)(int16_t* src, int32_t quarter_width, int32_t quarter_height, uint8_t* org, uint8_t* dst, int32_t dst_stride);
typedef void (*memcpy_stride_u_t)(void* src, int32_t width, int32_t height, int32_t src_stride, void* dst, int32_t dst_stride);
typedef uint32_t (*T264_cmp_t)(uint8_t* src, int32_t src_stride, uint8_t* data, int32_t dst_stride);
typedef void (*T264_pia_t)(uint8_t* p1, uint8_t* p2, int32_t p1_stride, int32_t p2_stride, uint8_t* dst, int32_t dst_stride);
typedef uint32_t (*T264_satd_i16x16_u_t)(uint8_t* src, int32_t src_stride, uint8_t* data, int32_t dst_stride);

typedef void (*T264_dct_t)(int16_t* data);
typedef void (*T264_quant4x4_t)(int16_t* data, const int32_t Qp, int32_t is_intra);
typedef void (*T264_quant4x4dc_t)(int16_t* data, const int32_t Qp);
typedef void (*T264_quant2x2dc_t)(int16_t* data, const int32_t Qp, int32_t is_intra);
typedef void (*T264_iquant_t)(int16_t* data, const int32_t Qp);

typedef void (*T264_eighth_pixel_mc_u_t)(uint8_t* src, int32_t src_stride, uint8_t* dst, int16_t mvx, int16_t mvy, int32_t width, int32_t height);
typedef void (*T264_interpolate_halfpel_t)(uint8_t* src, int32_t src_stride, uint8_t* dst, int32_t dst_stride, int32_t width, int32_t height);
typedef void (*T264_pixel_avg_t)(uint8_t* p1, uint8_t* p2, int32_t p1_stride, int32_t p2_stride, uint8_t* dst, int32_t dst_stride, int32_t w, int32_t h);
typedef uint32_t (*T264_search_t)(T264_t* t, T264_search_context_t* context);
typedef void (*T264_emms_t)();
typedef decoder_state_t (*action_t)(T264_t* t);

typedef struct
{
    // COPY BEGIN
    // NOTE: copied from T264_mb_cache_t except padding words!!!
    // intra 4x4 mode
    uint8_t     mode_i4x4[16];
    // inter sub partition size(if no sub partition -1)
    uint8_t     submb_part[16];
    // inter 4x4 block mv
    T264_vector_t vec[2][4 * 4];
	// for CABAC, mv delta
	int16_t		mvd[2][4*4][2];
    // non zero count
    uint8_t     nnz[16 + 4 + 4];
    // sad
    uint32_t sad;

    // I16x16, I4x4, P_MODE, ....
    uint8_t     mb_mode;
    uint8_t     mb_mode_uv;
    // intra 16x16 mode
    uint8_t     mode_i16x16;
    // inter
    uint8_t     mb_part;
    // rate control
    int8_t      mb_qp_delta; 
	// for CABAC, the following 4 fields should be saved in contexts
	uint8_t     is_copy;
	uint16_t    cbp_y;
	uint16_t    cbp_c;
	uint16_t	cbp;
    // COPY END

    T264_mb_context_t* context;
    uint8_t     mb_part2[2];    // b slice use

    int32_t     mb_neighbour;
    int16_t     mb_x;
    int16_t     mb_y;
    int16_t		mb_xy;
    int32_t     lambda;

    uint8_t*    src_y;
    uint8_t*    dst_y;	
    uint8_t*    src_u;
    uint8_t*    dst_u;
    uint8_t*    src_v;
    uint8_t*    dst_v;

    // save the predict value for intra encode
    DECLARE_ALIGNED_MATRIX_H(pred_i16x16, 16, 16, uint8_t, CACHE_SIZE);
    DECLARE_ALIGNED_MATRIX_H(pred_i8x8u, 8, 8, uint8_t, CACHE_SIZE);
    DECLARE_ALIGNED_MATRIX_H(pred_i8x8v, 8, 8, uint8_t, CACHE_SIZE);
    DECLARE_ALIGNED_MATRIX_H(pred_p16x16, 16, 16, uint8_t, CACHE_SIZE);
 
    int16_t     dct_y_z[16][4*4];               // 块进行Z扫描后的系数
    int16_t     dct_uv_z[2][4][4*4];
    int16_t     dc4x4_z[16];                    // Z扫描后的16个DC系数
    int16_t     dc2x2_z[2][4];                  // Z扫描后的4个DC系数
    // for CABAC, the following 4 fileds have been moved to the beginning
    uint8_t     sub_is_copy[4];

    int8_t i4x4_pred_mode_ref[5 * 8];   // see load_context for detail layout
    uint8_t nnz_ref[6 * 8];
	uint32_t sad_ref[3]; //left, top, top-right
    struct
    {
        T264_vector_t vec[2];
        uint8_t part;
        uint8_t subpart;
    } vec_ref[5 * 8];
	
	// for CABAC, mv delta
	int16_t mvd_ref[2][5 * 8][2];
} T264_mb_t;

struct T264_t
{
    T264_frame_t refn[MAX_REFFRAMES];
    T264_frame_t* ref[2][MAX_REFFRAMES];
    int32_t refl0_num;
    int32_t refl1_num;
    T264_frame_t cur;
    T264_frame_t* rec;
    int32_t     width;
    int32_t     height;
    int32_t     stride;
    int32_t     stride_uv;
    int32_t     edged_stride;
    int32_t     edged_stride_uv;
    int32_t     edged_width;
    int32_t     edged_height;
    int32_t     qp_y, qp_uv;
    void*        bs;
    uint8_t*    bs_buf;
    uint32_t    flags;
    int32_t     mb_width, mb_height;
	int32_t		mb_stride;
    uint32_t    idr_pic_id;
    /* the frame unique id in the whole encoding session, for statistic(rc uses it too) */
    uint32_t    frame_id;
    /* the frame_num in the bitstream semantic, for generating bitstream only */
    uint32_t    frame_num;
    /* the frame unique id in the current encoding gop, for deciding the slice type usage */
    uint32_t    frame_no;
    /* the frame unique id that the last key frame id */
    uint32_t    last_i_frame_id;
    uint32_t    poc;
    uint32_t    slice_type;
    int32_t     skip;
    uint32_t    sad_all;
    T264_frame_t pending_bframes[MAX_BREFNUMS];
    int32_t     pending_bframes_num;
    int32_t     header_bits;
    int32_t     frame_bits;

    /* ++ decoder section ++ */
    /* source section */
    uint8_t* src_buf;   /* source buffer start ptr */ 
    uint8_t* src_end;
    uint8_t* nal_buf;   /* buffer one whole nal unit */
    int32_t nal_len;    /* one nal unit length */

    uint32_t shift;     /* nal decode use */
    uint32_t shift1;

    action_t action;

    int32_t need_deblock;
    T264_frame_t* cur_frame;
    /* frame rate info */
    int32_t aspect_ratio;
    int32_t video_format;
    T264_frame_t output;
    /* -- decoder section -- */

    T264_param_t   param;
    T264_nal_t     nal;
    T264_seq_set_t ss;
    T264_pic_set_t ps;
    T264_slice_t   slice;
    T264_stat_t    stat;

    T264_mb_t      mb;
    int16_t        subpel_pts;
    T264_plugin_t  plugins[MAX_PLUGINS];

	//for CABAC
	T264_cabac_t cabac;

    int32_t        plug_num;
    T264_predict_16x16_mode_t pred16x16[4 + 3];
	T264_predict_8x8_mode_t   pred8x8[4 + 3];
	T264_predict_4x4_mode_t   pred4x4[9 + 3];
	T264_cmp_t cmp[8];
    T264_cmp_t sad[8];
    T264_pia_t pia[9];  //for pixel avearage func

    T264_dct_t fdct4x4;
    T264_dct_t fdct4x4dc;
    T264_dct_t fdct2x2dc;
    T264_dct_t idct4x4;
    T264_dct_t idct4x4dc;
    T264_dct_t idct2x2dc;

    T264_quant4x4_t   quant4x4;
    T264_quant4x4dc_t quant4x4dc;
    T264_quant2x2dc_t quant2x2dc;
    T264_iquant_t     iquant4x4;
    T264_iquant_t     iquant4x4dc;
    T264_iquant_t     iquant2x2dc;

    expand8to16_t     expand8to16;
    contract16to8_t   contract16to8;
    expand8to16sub_t  expand8to16sub;
    contract16to8add_t contract16to8add;
    memcpy_stride_u_t   memcpy_stride_u;
    T264_eighth_pixel_mc_u_t eighth_pixel_mc_u;
    T264_interpolate_halfpel_t interpolate_halfpel_h;
    T264_interpolate_halfpel_t interpolate_halfpel_v;
    T264_interpolate_halfpel_t interpolate_halfpel_hv;
    T264_pixel_avg_t pixel_avg;
    T264_satd_i16x16_u_t T264_satd_16x16_u;
    T264_search_t   search;
    T264_emms_t emms;
};

/* private func(for encoder & decoder share) */
void T264_init_cpu(T264_t* t);
void T264_mb_load_context(T264_t* t, int32_t mb_y, int32_t mb_x);
void T264_extend_border(T264_t* t, T264_frame_t* f);
void T264_interpolate_halfpel(T264_t* t, T264_frame_t* f);

/* extern api */
T264_t* T264_open(T264_param_t* para);
void T264_close(T264_t* t);
int32_t T264_encode(T264_t* t, uint8_t* src, uint8_t* dst, int32_t dst_size);

T264_t* T264dec_open();
void T264dec_close(T264_t* t);
void T264dec_buffer(T264_t* t, uint8_t* buf, int32_t len);
decoder_state_t T264dec_parse(T264_t* t);
T264_frame_t* T264dec_flush_frame(T264_t* t);

#ifdef __cplusplus
};
#endif

#endif