173 lines
17 KiB
C
173 lines
17 KiB
C
#ifndef __SCALER_NEON_H__
|
|
#define __SCALER_NEON_H__
|
|
#include <stdint.h>
|
|
|
|
//
|
|
// arm NEON / C integer scalers for rg35xx
|
|
// 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);
|
|
|
|
// Functions for generic call
|
|
// n/c = neon or c
|
|
// 16/32 = bpp
|
|
// xmul = 1,2,3,4,5,6
|
|
// ymul = 1,2,3,4(xmul < 5) / 1,2,3,4,5(xmul == 5) / 1,2,3,4,5,6(xmul == 6)
|
|
void scaler_n16(uint32_t xmul, uint32_t ymul, void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scaler_n32(uint32_t xmul, uint32_t ymul, void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scaler_c16(uint32_t xmul, uint32_t ymul, void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scaler_c32(uint32_t xmul, uint32_t ymul, void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
|
|
// NEON memcpy
|
|
void memcpy_neon(void* dst, void* src, uint32_t size);
|
|
|
|
// NEON scalers
|
|
void scale1x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale1x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale2x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale2x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale3x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale3x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale4x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale4x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale5x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale5x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale6x_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale6x_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
|
|
void scale1x1_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x1_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x2_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x2_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x3_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x3_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x4_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x4_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x1_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x1_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x2_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x2_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x3_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x3_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x4_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x4_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x1_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x1_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x2_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x2_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x3_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x3_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x4_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x4_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x1_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x1_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x2_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x2_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x3_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x3_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x4_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x4_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x1_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x1_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x2_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x2_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x3_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x3_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x4_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x4_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x5_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x5_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x1_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x1_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x2_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x2_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x3_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x3_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x4_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x4_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x5_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x5_n32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x6_n16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x6_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, uint32_t ymul);
|
|
void scale1x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale2x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale2x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale3x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale3x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale4x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale4x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale5x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale5x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale6x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
void scale6x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp, uint32_t ymul);
|
|
|
|
void scale1x1_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x1_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x2_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x2_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x3_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x3_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x4_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale1x4_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x1_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x1_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x2_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x2_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x3_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x3_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x4_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale2x4_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x1_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x1_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x2_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x2_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x3_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x3_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x4_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale3x4_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x1_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x1_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x2_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x2_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x3_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x3_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x4_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale4x4_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x1_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x1_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x2_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x2_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x3_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x3_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x4_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x4_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x5_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale5x5_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x1_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x1_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x2_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x2_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x3_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x3_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x4_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x4_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x5_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x5_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x6_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
void scale6x6_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
|
|
|
|
#endif
|