void test_vrsqrteQf32 (void)
{
float32x4_t out_float32x4_t;
float32x4_t arg0_float32x4_t;
out_float32x4_t = vrsqrteq_f32 (arg0_float32x4_t);
}
inline float32x4_t cv_vrsqrtq_f32(float32x4_t val)
{
float32x4_t e = vrsqrteq_f32(val);
e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(e, e), val), e);
e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(e, e), val), e);
return e;
}
static float32x4_t vsqrtq_f32(float32x4_t s) {
int i;
float32x4_t x = vrsqrteq_f32(s);
// Code to handle sqrt(0).
// If the input to sqrtf() is zero, a zero will be returned.
// If the input to vrsqrteq_f32() is zero, positive infinity is returned.
const uint32x4_t vec_p_inf = vdupq_n_u32(0x7F800000);
// check for divide by zero
const uint32x4_t div_by_zero = vceqq_u32(vec_p_inf, vreinterpretq_u32_f32(x));
// zero out the positive infinity results
x = vreinterpretq_f32_u32(vandq_u32(vmvnq_u32(div_by_zero),
vreinterpretq_u32_f32(x)));
// from arm documentation
// The Newton-Raphson iteration:
// x[n+1] = x[n] * (3 - d * (x[n] * x[n])) / 2)
// converges to (1/√d) if x0 is the result of VRSQRTE applied to d.
//
// Note: The precision did not improve after 2 iterations.
for (i = 0; i < 2; i++) {
x = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, x), s), x);
}
// sqrt(s) = s * 1/sqrt(s)
return vmulq_f32(s, x);;
}
inline v_float32x4 v_invsqrt(const v_float32x4& x)
{
float32x4_t e = vrsqrteq_f32(x.val);
e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x.val, e), e), e);
e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x.val, e), e), e);
return v_float32x4(e);
}
inline v_float32x4 v_sqrt(const v_float32x4& x)
{
float32x4_t x1 = vmaxq_f32(x.val, vdupq_n_f32(FLT_MIN));
float32x4_t e = vrsqrteq_f32(x1);
e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x1, e), e), e);
e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x1, e), e), e);
return v_float32x4(vmulq_f32(x.val, e));
}
float32x4_t test_vrsqrteq_f32(float32x4_t in) {
// CHECK-LABEL: @test_vrsqrteq_f32
// CHECK: call <4 x float> @llvm.arm64.neon.frsqrte.v4f32(<4 x float> %in)
return vrsqrteq_f32(in);
}