inline int32x2_t cv_vrnd_s32_f32(float32x2_t v)
{
static int32x2_t v_sign = vdup_n_s32(1 << 31),
v_05 = vreinterpret_s32_f32(vdup_n_f32(0.5f));
int32x2_t v_addition = vorr_s32(v_05, vand_s32(v_sign, vreinterpret_s32_f32(v)));
return vcvt_s32_f32(vadd_f32(v, vreinterpret_f32_s32(v_addition)));
}
void test_vaddf32 (void)
{
float32x2_t out_float32x2_t;
float32x2_t arg0_float32x2_t;
float32x2_t arg1_float32x2_t;
out_float32x2_t = vadd_f32 (arg0_float32x2_t, arg1_float32x2_t);
}
static inline void mix_block_into(cbox_sample_t **outputs, int oofs, float *src_leftright)
{
float *dst_left = outputs[oofs];
float *dst_right = outputs[oofs + 1];
for (size_t i = 0; i < CBOX_BLOCK_SIZE; i += 2)
{
float32x2_t lr1 = vld1_f32(&src_leftright[2 * i]);
float32x2_t lr2 = vld1_f32(&src_leftright[2 * i + 2]);
float32x2x2_t lr12 = vtrn_f32(lr1, lr2);
float32x2_t dl1 = vld1_f32(&dst_left[i]);
float32x2_t dr1 = vld1_f32(&dst_right[i]);
float32x2_t l1 = vadd_f32(dl1, lr12.val[0]);
vst1_f32(&dst_left[i], l1);
float32x2_t r1 = vadd_f32(dr1, lr12.val[1]);
vst1_f32(&dst_right[i], r1);
}
}
inline uint32x2_t cv_vrnd_u32_f32(float32x2_t v)
{
static float32x2_t v_05 = vdup_n_f32(0.5f);
return vcvt_u32_f32(vadd_f32(v, v_05));
}
inline float32x2_t vadd(const float32x2_t & v0, const float32x2_t & v1) { return vadd_f32(v0, v1); }
