S.O.N.G/TheGaborator
S.O.N.G/TheGaborator
3
0
Fork 0
Code Issues Pull requests Releases Activity

Use standard intrinsic types, adjust __builtin with intrinsics

Browse source
This commit is contained in:
DataHoarder 2022-07-15 18:19:31 +02:00
parent cfe987a51d
commit f86f6d213a
Signed by: DataHoarder
SSH key fingerprint: SHA256:OLTRf6Fl87G52SiR7sWLGNzlJt4WOX+tfI2yxo0z7xk
1 changed files with 38 additions and 38 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

76
gaborator/vector_math.h
View file

@ -47,12 +47,12 @@ static inline
std::complex<double> complex_mul(std::complex<double> a_,
std::complex<double> b_)
{
__v2df a = _mm_setr_pd(a_.real(), a_.imag());
__v2df b = _mm_setr_pd(b_.real(), b_.imag());
__v2df as = (__v2df) _mm_shuffle_pd(a, a, 0x1);
__v2df t0 = _mm_mul_pd(a, _mm_shuffle_pd(b, b, 0x0));
__v2df t1 = _mm_mul_pd(as, _mm_shuffle_pd(b, b, 0x3));
__v2df c = __builtin_ia32_addsubpd(t0, t1); // SSE3
__m128d a = _mm_setr_pd(a_.real(), a_.imag());
__m128d b = _mm_setr_pd(b_.real(), b_.imag());
__m128d as = (__m128d) _mm_shuffle_pd(a, a, 0x1);
__m128d t0 = _mm_mul_pd(a, _mm_shuffle_pd(b, b, 0x0));
__m128d t1 = _mm_mul_pd(as, _mm_shuffle_pd(b, b, 0x3));
__m128d c = _mm_addsub_pd(t0, t1); // SSE3
return std::complex<double>(c[0], c[1]);
}
#else
@ -113,11 +113,11 @@ complex_magnitude_naive(I *GABORATOR_RESTRICT inv,
// Perform two complex float multiplies in parallel
static inline
__v4sf complex_mul_vec2(__v4sf aa, __v4sf bb) {
__v4sf aas =_mm_shuffle_ps(aa, aa, 0xb1);
__v4sf t0 = _mm_mul_ps(aa, _mm_moveldup_ps(bb));
__v4sf t1 = _mm_mul_ps(aas, _mm_movehdup_ps(bb));
return __builtin_ia32_addsubps(t0, t1); // SSE3
__m128 complex_mul_vec2(__m128 aa, __m128 bb) {
__m128 aas =_mm_shuffle_ps(aa, aa, 0xb1);
__m128 t0 = _mm_mul_ps(aa, _mm_moveldup_ps(bb));
__m128 t1 = _mm_mul_ps(aas, _mm_movehdup_ps(bb));
return _mm_addsub_ps(t0, t1); // SSE3
}
// Calculate the elementwise product of a complex float vector
@ -131,12 +131,12 @@ elementwise_product(std::complex<float> *cv,
{
assert((n & 1) == 0);
n >>= 1;
__v4sf *c = (__v4sf *) cv;
const __v4sf *a = (const __v4sf *) av;
const __v4sf *b = (const __v4sf *) bv;
__m128 *c = (__m128 *) cv;
const __m128 *a = (const __m128 *) av;
const __m128 *b = (const __m128 *) bv;
while (n--) {
__v4sf aa = *a++;
__v4sf bb = *b++;
__m128 aa = *a++;
__m128 bb = *b++;
*c++ = complex_mul_vec2(aa, bb);
}
}
@ -154,13 +154,13 @@ elementwise_product(std::complex<float> *cv,
{
assert((n & 3) == 0);
n >>= 2;
__v4sf *c = (__v4sf *) cv;
const __v4sf *a = (const __v4sf *) av;
const __v4sf *b = (const __v4sf *) bv;
__m128 *c = (__m128 *) cv;
const __m128 *a = (const __m128 *) av;
const __m128 *b = (const __m128 *) bv;
while (n--) {
__v4sf a0 = (__v4sf) _mm_loadu_si128((const __m128i *) a++);
__v4sf a1 = (__v4sf) _mm_loadu_si128((const __m128i *) a++);
__v4sf bb = *b++;
__m128 a0 = (__m128) _mm_loadu_si128((const __m128i *) a++);
__m128 a1 = (__m128) _mm_loadu_si128((const __m128i *) a++);
__m128 bb = *b++;
*c++ = _mm_mul_ps(a0, _mm_unpacklo_ps(bb, bb));
*c++ = _mm_mul_ps(a1, _mm_unpackhi_ps(bb, bb));
}
@ -175,13 +175,13 @@ elementwise_product_times_scalar(std::complex<float> *cv,
{
assert((n & 1) == 0);
n >>= 1;
const __v4sf *a = (const __v4sf *) av;
const __v4sf *b = (const __v4sf *) bv;
const __v4sf dd = (__v4sf) { d.real(), d.imag(), d.real(), d.imag() };
__v4sf *c = (__v4sf *) cv;
const __m128 *a = (const __m128 *) av;
const __m128 *b = (const __m128 *) bv;
const __m128 dd = (__m128) { d.real(), d.imag(), d.real(), d.imag() };
__m128 *c = (__m128 *) cv;
while (n--) {
__v4sf aa = *a++;
__v4sf bb = *b++;
__m128 aa = *a++;
__m128 bb = *b++;
*c++ = complex_mul_vec2(complex_mul_vec2(aa, bb), dd);
}
}
@ -200,19 +200,19 @@ complex_magnitude(std::complex<float> *inv,
*outv++ = std::sqrt(v.real() * v.real() + v.imag() * v.imag());
n--;
}
const __v4sf *in = (const __v4sf *) inv;
__v4sf *out = (__v4sf *) outv;
const __m128 *in = (const __m128 *) inv;
__m128 *out = (__m128 *) outv;
while (n >= 4) {
__v4sf aa = *in++; // c0re c0im c1re c1im
__v4sf aa2 = _mm_mul_ps(aa, aa); // c0re^2 c0im^2 c1re^2 c1im^2
__v4sf bb = *in++; // c2re c2im c3re c3im
__v4sf bb2 = _mm_mul_ps(bb, bb); // etc
__m128 aa = *in++; // c0re c0im c1re c1im
__m128 aa2 = _mm_mul_ps(aa, aa); // c0re^2 c0im^2 c1re^2 c1im^2
__m128 bb = *in++; // c2re c2im c3re c3im
__m128 bb2 = _mm_mul_ps(bb, bb); // etc
// Gather the real parts: x0 x2 y0 y2
// 10 00 10 00 = 0x88
__v4sf re2 =_mm_shuffle_ps(aa2, bb2, 0x88);
__v4sf im2 =_mm_shuffle_ps(aa2, bb2, 0xdd);
__v4sf mag2 = _mm_add_ps(re2, im2);
__v4sf mag = __builtin_ia32_sqrtps(mag2);
__m128 re2 =_mm_shuffle_ps(aa2, bb2, 0x88);
__m128 im2 =_mm_shuffle_ps(aa2, bb2, 0xdd);
__m128 mag2 = _mm_add_ps(re2, im2);
__m128 mag = _mm_sqrt_ps(mag2);
// Unaligned store
_mm_storeu_si128((__m128i *)out, (__m128i)mag);
out++;