/********************************************************************
	INCLUDE FILES
********************************************************************/

#define DBG_YOLOV5S_DETECTOUT   0
#define PREDICT_NUMBERS         20000//20000
#define FINAL_PRED_NUMBERS      10000//10000
#define YOLOV5S_AXIS_NUM       5
#define NN_AXIS_NUM			4

#define NUM_CLASS   22

#define INPUT_W     640
#define INPUT_H     640

#define scale_out_w	640 
#define scale_out_h	640

typedef struct _YOLOV5S_Bbox {
	float xmin;
	float ymin;
	float xmax;
	float ymax;
	float score;
	int cid;
} YOLOV5S_Bbox;

typedef struct _YOLOV5S_Anchor {
	FLOAT width;
	FLOAT height;
} YOLOV5S_Anchor;

typedef struct _TRANSPOSE_PARM {
	UINT32 in_addr;
	UINT32 out_addr;                       
	UINT32 tmp_addr;                       
	UINT32 width;
	UINT32 height;
	UINT32 channels;
	UINT32 batch_num;
	UINT32 in_line_ofs;                       
	UINT32 in_channel_ofs;                     
	UINT32 in_batch_ofs;						   							
	UINT32 out_line_ofs;                        
	UINT32 out_channel_ofs;                     
	UINT32 out_batch_ofs;						                          
	INT8 in_frac_bits;
	INT8 in_int_bits;
	INT8 in_sign_bits;
	INT8 in_reserved;
	INT8 out_frac_bits;
	INT8 out_int_bits;
	INT8 out_sign_bits;
	INT8 out_reserved;
	INT32 in_mul;                              
	INT8 in_shf;
	INT32 out_mul;                              
	INT8 out_shf;
	INT8 order[NN_AXIS_NUM];                
}TRANSPOSE_PARM;

VOID print_detection_out_info(INT32 num, YOLOV5S_Bbox *yolov5s_final_bbox_buf);
void transform(INT32 num_kept, INT32 ih, INT32 iw, INT32 input_h, INT32 input_w, YOLOV5S_Bbox *yolov5s_final_bbox, BOOL is_padding);
INT32 yolov5s_nms(YOLOV5S_Bbox *yolov5s_final_bbox, YOLOV5S_Bbox *yolov5s_post_bbox, FLOAT nms_thre, INT32 *bbox_num);
void transpose_data(TRANSPOSE_PARM *t_parm);
void post_process_single(YOLOV5S_Bbox *yolov5s_post_bbox,INT32 input_w, INT32 input_h, INT32 height, INT32 width, INT32 idx, FLOAT conf_thre, INT16 *layer_float, INT32 *strides, YOLOV5S_Anchor *anchor, INT32 *bbox_num, INT8 frac_bits, float scale_ratio);