ARMアセンブリ言語の実装色々とNEON命令のサンプル ref: https://qiita.com/take-iwiw/items/cea0a2cb4d2709cb7ee5

call.c

#include <stdio.h>
int main()
{
    int a;
    int ret;
    extern int func1(int *a);
    ret = func1(&a);
    printf("ret = %d, a = %d\n", ret, a);
}

inline.c

#include <stdio.h>

int func1(int *a)
{
    int ret = 0;
    __asm__ __volatile__ (
        "mov r1, #99;"
        "str r1, [%1];"
        "mov %0, #88;"
        : "=r"(ret)     // Output operands
        : "r"(a)        // Input operands
        :               // Overwritten registers
    );
    return ret;
}

int add(int a, int b)
{
    int ret = 0;
    __asm__ __volatile__ (
        "add %0, %1, %2;"
        : "=r"(ret)         // Output operands
        : "r"(a), "r"(b)    // Input operands
        :                   // Overwritten registers
    );
    return ret;
}

int add_label(int a, int b)
{
    int ret = 0;
    __asm__ __volatile__ (
        "add %[Rret], %[Ra], %[Rb];"
        : [Rret]"=r"(ret)           // Output operands
        : [Ra]"r"(a), [Rb]"r"(b)    // Input operands
        :                           // Overwritten registers
    );
    return ret;
}

void copy(int *src, int *dst)
{
    __asm__ __volatile__ (
        "ldr r2, [%0];" // r2 <- *src
        "str r2, [%1];" // r2 -> *dst
        :                       // Output operands
        : "r"(src), "r"(dst)    // Input operands
        : "r2", "memory"        // Overwritten registers
    );
    return;
}

int main()
{
    int a, b, ret;

    ret = func1(&a);
    printf("ret = %d, a = %d\n", ret, a);

    ret = add(1, 2);
    printf("ret = %d\n", ret);

    ret = add_label(1, 2);
    printf("ret = %d\n", ret);

    a = 10; b = 0;
    copy(&a, &b);
    printf("a = %d b = %d\n", a, b);
}

simd.c

#include <stdio.h>
#include <stdint.h>
#include <string.h>
#include <arm_neon.h>

void add_emb()
{
    uint8_t a[8];   // d0
    uint8_t b[8];   // d1
    uint8_t c[8];   // result
    memset(a, 0x12, sizeof(a));
    a[2] = 0xEE;    // 飽和確認用
    memset(b, 0x34, sizeof(b));
    memset(c, 0x00, sizeof(c));

    uint8x8_t va, vb, vc;       // 8-bit(uint8_t) x 8レーン
    va = vld1_u8(a);
    vb = vld1_u8(b);
    vc = vqadd_u8(va, vb);
    vst1_u8(c, vc);

    for (int i = 0; i < 8; i++)
        printf("%02X + %02X = %02X\n", a[i], b[i], c[i]);

    return;
}

int main()
{
    add_emb();
}

simd_auto.c

#include <stdio.h>
#include <stdint.h>
#include <string.h>

#define NUM 128

void add(uint8_t * a, uint8_t * b, uint8_t * c)
{
    for(int i = 0; i < NUM; i++) {
        c[i] = a[i] + b[i];
    }
}

int main()
{
    uint8_t a[NUM], b[NUM], c[NUM];
    memset(a, 0x12, sizeof(a));
    memset(b, 0x34, sizeof(a));
    add(a, b, c);
    printf("%08X\n", *(uint32_t*)&c[0]);
}

sub.s

    .file   "sub.s"
    .text
    .global func1
    .type   func1, %function
func1:
    mov r1, #99
    str r1, [r0]    @ r0 is the first arg
    mov r0, #88
    bx  lr
    .end

ビルドコマンドと実行結果

gcc simd.c -mfpu=neon
./a.out
12 + 34 = 46
12 + 34 = 46
EE + 34 = FF
12 + 34 = 46
12 + 34 = 46
12 + 34 = 46
12 + 34 = 46
12 + 34 = 46

ビルドコマンド確認用

gcc -mfpu=neon -O3 -S simd_auto.c
more simd_auto.s