gistfile1.txt

#ifdef __arm__
typedef float32x4_t float4;
typedef int32x4_t   int4;
typedef uint32x4_t  uint4;
#else
typedef __m128  float4;
typedef __m128i int4;
#if defined(__AVX__)
typedef __m256  float8;
typedef __m256d double4;
#endif
#endif


#ifdef __arm__

#define vadd_f4(a, b) vaddq_f32((a), (b))
#define vdot_f4(ax, ay, az, bx, by, bz) ( \
     vaddq_f32(vmulq_f32((ax), (bx)), \
           vaddq_f32(vmulq_f32((ay), (by)), \
                      vmulq_f32((az), (bz)))) )

#else // sse

#define vadd_f4(a, b) _mm_add_ps((a), (b))
#define vdot_f4(ax, ay, az, bx, by, bz) ( \
     _mm_add_ps(_mm_mul_ps((ax), (bx)), \
           _mm_add_ps(_mm_mul_ps((ay), (by)), \
                      _mm_mul_ps((az), (bz)))) )

#endif