#ifdef _MSC_VER
#pragma once
#endif

#ifndef SSC_INTRIN
#define SSC_INTRIN

#include "StgSharpC.h"
#include <immintrin.h>

typedef union column_2 {
        __m256 stream[1];
        __m128 column[2];
        float m[4][2];
} __2_columnset;

typedef struct column_3 {
        __m256 part;
        __m128 column[3];
        float m[4][3];
} __3_columnset;

typedef union column_4 {
        __m512 stream;
        __m256 half[2];
        __m128 column[4];
        float m[4][4];
} __4_columnset;

typedef void (SSCDECL *VECTORNORMALIZEPROC)(__m128 *source, __m128 *target);
typedef void (SSCDECL *MATRIXTRANSPOSEPROC)(void *source, void *target);
typedef void (SSCDECL *MATRIXDETPROC)(void *mat, void *transpose);
typedef void (SSCDECL *MATARITHMETIC)(void* left, void* right, void* ans);
typedef void (SSCDECL *SCALERMATARITHMETIC)(void* left, void* right, void* ans);

#pragma region matix function

INTERNAL void normalize(__m128 *source, __m128 *target);

INTERNAL void SSCDECL addmatrix2_sse(__2_columnset* left, __2_columnset* right, __2_columnset* ans);
INTERNAL void SSCDECL addmatrix2_avx(__2_columnset* left, __2_columnset* right, __2_columnset* ans);
INTERNAL void SSCDECL transpose23(__2_columnset *source, __3_columnset *target);
INTERNAL void SSCDECL transpose24(__2_columnset *source, __4_columnset *target);


INTERNAL void SSCDECL addmatrix3_sse(__3_columnset* left, __3_columnset* right, __3_columnset* ans);
INTERNAL void SSCDECL addmatrix3_avx(__3_columnset* left, __3_columnset* right, __3_columnset* ans);
INTERNAL void SSCDECL addmatrix3_512(__3_columnset* left, __3_columnset* right, __3_columnset* ans);
INTERNAL void SSCDECL transpose32(__3_columnset *source, __2_columnset *target);
INTERNAL void SSCDECL transpose33(__3_columnset *source, __3_columnset *target);
INTERNAL void SSCDECL transpose34(__3_columnset *source, __4_columnset *target);
INTERNAL float SSCDECL det_mat3(__3_columnset *matrix, __3_columnset *tranpose);

INTERNAL void SSCDECL transpose42(__4_columnset *source, __2_columnset *target);
INTERNAL void SSCDECL transpose43(__4_columnset *source, __3_columnset *target);
INTERNAL void SSCDECL transpose44(__4_columnset *matrix, __4_columnset *target);
INTERNAL float SSCDECL det_mat4(__4_columnset *matrix, __4_columnset *transpose);

#pragma endregion

#pragma region quick_hash

INTERNAL int quick_string_hash(byte *str, int length);

#pragma endregion

typedef struct ssc_intrinsic {
        MATARITHMETIC add_m2;
        MATARITHMETIC add_m3;
        MATARITHMETIC add_m4;
        MATRIXDETPROC det_matrix33;
        MATRIXDETPROC det_matrix44;
        VECTORNORMALIZEPROC normalize_3;
        int (*string_quick_hash)(char *str, int length);
        MATRIXTRANSPOSEPROC transpose_23;
        MATRIXTRANSPOSEPROC transpose_24;
        MATRIXTRANSPOSEPROC transpose_32;
        MATRIXTRANSPOSEPROC transpose_33;
        MATRIXTRANSPOSEPROC transpose_34;
        MATRIXTRANSPOSEPROC transpose_42;
        MATRIXTRANSPOSEPROC transpose_43;
        MATRIXTRANSPOSEPROC transpose_44;
} ssc_intrinsic;

enum most_advance_instruction
{
        SSE,
        AVX,
        AVX512,
        NEON
};

#endif // !SSC_INTRIN