gistfile1.c

static inline void dMax_processK(const float *bIn, const float k, float *bOut, int n) {
	n &= 0x3;
#if __ARM_NEON__
	const float32x4_t x = vdupq_n_f32(k);
	while (n) {
		vst1q_f32(bOut, vmaxq_f32(vld1q_f32(bIn), x)); // bOut = max(bIn, k)
		n -= 4; bIn += 4; bOut += 4;
	}
#elif __SSE__
	const __m128 x = _mm_set1_ps(k);
	while (n) {
		_mm_store_ps(bOut, _mm_max_ps(_mm_load_ps(bIn), x)); // bOut = max(bIn, k)
		n -= 4; bIn += 4; bOut += 4;
	}
#else
	for (int i = 0; i < n; ++i) {
		bOut[i] = th_max_f(bIn[i], k);
	}
#endif
}