ARM NEON Helpers

BCNeonImageUtils.h

//
//  BCNeonImageUtils.h
//  IRStudio
//
//  Created by Brian Cook on 5/30/12.
//  Copyright (c) 2012 All rights reserved.
//

#ifndef IRStudio_BCNeonImageUtils_h
#define IRStudio_BCNeonImageUtils_h

static void neon_BGRA_to_ARGB(unsigned char *src, int width, int height) {
    int numPixels = width * height;
    int iterations = numPixels / 8;
    
    __asm__ __volatile__("1:                        \n"
                         "vld4.8    {d0-d3}, [%0]   \n"
                         "vswp.8    d0, d3          \n" //swap Blue and Alpha
                         "vswp.8    d1, d2          \n" //swap Red and Green
                         "vst4.8    {d0-d3}, [%0]!  \n"
                         "subs      %1, %1, #1      \n"
                         "bne       1b              \n"
                         :
                         : "r"(src), "r"(iterations)
                         : "memory"
                         );
}

static void neon_BGRA_to_RGBA(unsigned char *src, int width, int height) {
    int numPixels = width * height;
    int iterations = numPixels / 8;
    
    __asm__ __volatile__("1:                        \n"
                         "vld4.8    {d0-d3}, [%0]   \n"
                         "vswp.8    d0, d2          \n" //swap Red and Blue
                         "vst4.8    {d0-d3}, [%0]!  \n"
                         "subs      %1, %1, #1      \n"
                         "bne       1b              \n"
                         :
                         : "r"(src), "r"(iterations)
                         : "memory"
                         );
}

static void neon_asm_switch_red_and_blue(unsigned char *src, int numPixels) {
	__asm__ __volatile__("1:			        \n"
			     "vld4.8	{d0-d3}, [%0]	        \n"
			     "vswp.8	d1, d3			\n"
			     "vst4.8	{d0-d3}, [%0]!	        \n"
			     "subs	%1, %1, #1		\n"
			     "bne	1b			\n"
			     : 
			     : "r"(src), "r"(numPixels)
			     : "memory"
			     );
}

#endif