#pragma once
#include <immintrin.h>
#include <iostream>
#include <opencv2/opencv.hpp>

/* helper tables */
extern const unsigned char icvSaturate8u_cv[];
#define CV_FAST_CAST_8U(t)  ( (-256 <= (t) && (t) <= 512) ? icvSaturate8u_cv[(t)+256] : 0 )
#define CV_CALC_MIN_8U(a,b) (a) -= CV_FAST_CAST_8U((a) - (b))
#define CV_CALC_MAX_8U(a,b) (a) += CV_FAST_CAST_8U((b) - (a))


// Any 1 to 1.

// memset for vin is meant to clear the source buffer so that
// SIMD that reads full multiple of 16 bytes will not trigger msan errors.
// memset is not needed for production, as the garbage values are processed but
// not used, although there may be edge cases for subsampling.
// The size of the buffer is based on the largest read, which can be inferred
// by the source type (e.g. ARGB) and the mask (last parameter), or by examining
// the source code for how much the source pointers are advanced.

//数据4对齐
#define ALIGN_4(x) (((x) + 3) & ~3)
//数据2对齐
#define ALIGN_2(x) (((x) + 1) & ~1)

// Subsampled source needs to be increase by 1 of not even.
#define SS(width, shift) (((width) + (1 << (shift)) - 1) >> (shift))
#define SIMD_ALIGNED(var) __declspec(align(16)) var
//#define ANY11(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK)             
void FastCopy_Any(const unsigned char* src_ptr, unsigned char* dst_ptr, int width);

void FastCopy(unsigned char* pSrc, unsigned char* pDst, int nWidth);

void distanceATS_L1_8u(unsigned char* const src,int nWidth,int nHeight,int nPitch, unsigned char* dst);

void OptimizeCalMatAdd_8U(unsigned char* pRelated, int nRelatedWidth, int nRelatedHeight, int nRelatedPitch,
	unsigned char* pTarget, int nTargetWidth, int nTargetHeight, int nTargetPitch,
	int nLeft, int nTop,
	int nChannel
);

//ANY11(FastCopy_Any, FastCopy, 0, 1, 1, 63);