#ifndef __SCALER_NEON_H__
#define __SCALER_NEON_H__
#include <stdint.h>

//
//	arm NEON / C integer scalers for miyoomini
//	args/	src :	src offset		address of top left corner
//		dst :	dst offset		address	of top left corner
//		sw  :	src width		pixels
//		sh  :	src height		pixels
//		sp  :	src pitch (stride)	bytes	if 0, (src width * [2|4]) is used
//		dp  :	dst pitch (stride)	bytes	if 0, (src width * [2|4] * multiplier) is used
//
//	** NOTE **
//	since 32bit aligned addresses need to be processed for NEON scalers,
//	x-offset and stride pixels must be even# in the case of 16bpp,
//	if odd#, then handled by the C scaler
//

typedef void (*scale_neon_t)(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);


//	NEON scalers
void scale1x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale1x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale2x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale2x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale3x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale3x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale4x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale4x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale5x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale5x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale6x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale6x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);

//	C scalers
void scale1x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale1x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale2x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale2x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale3x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale3x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale4x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale4x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
//	c16b/c32b: faster when -Ofast/-O3 and aligned width, however dp must be 4xN
void scale4x_c16b(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale4x_c32b(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale5x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale5x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale6x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
void scale6x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);

#endif