protot/3rdparty/bx/include/bx/simd_ni.inl

559 lines
13 KiB
C++

/*
* Copyright 2010-2016 Branimir Karadzic. All rights reserved.
* License: https://github.com/bkaradzic/bx#license-bsd-2-clause
*/
#ifndef BX_SIMD_NI_H_HEADER_GUARD
#define BX_SIMD_NI_H_HEADER_GUARD
namespace bx
{
template<typename Ty>
BX_SIMD_INLINE Ty simd_shuf_xAzC_ni(Ty _a, Ty _b)
{
const Ty xAyB = simd_shuf_xAyB(_a, _b);
const Ty zCwD = simd_shuf_zCwD(_a, _b);
const Ty result = simd_shuf_xyAB(xAyB, zCwD);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_shuf_yBwD_ni(Ty _a, Ty _b)
{
const Ty xAyB = simd_shuf_xAyB(_a, _b);
const Ty zCwD = simd_shuf_zCwD(_a, _b);
const Ty result = simd_shuf_zwCD(xAyB, zCwD);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_madd_ni(Ty _a, Ty _b, Ty _c)
{
const Ty mul = simd_mul(_a, _b);
const Ty result = simd_add(mul, _c);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_nmsub_ni(Ty _a, Ty _b, Ty _c)
{
const Ty mul = simd_mul(_a, _b);
const Ty result = simd_sub(_c, mul);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_div_nr_ni(Ty _a, Ty _b)
{
const Ty oneish = simd_isplat<Ty>(0x3f800001);
const Ty est = simd_rcp_est(_b);
const Ty iter0 = simd_mul(_a, est);
const Ty tmp1 = simd_nmsub(_b, est, oneish);
const Ty result = simd_madd(tmp1, iter0, iter0);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_rcp_ni(Ty _a)
{
const Ty one = simd_splat<Ty>(1.0f);
const Ty result = simd_div(one, _a);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_orx_ni(Ty _a)
{
const Ty zwxy = simd_swiz_zwxy(_a);
const Ty tmp0 = simd_or(_a, zwxy);
const Ty tmp1 = simd_swiz_yyyy(_a);
const Ty tmp2 = simd_or(tmp0, tmp1);
const Ty mf000 = simd_ild<Ty>(UINT32_MAX, 0, 0, 0);
const Ty result = simd_and(tmp2, mf000);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_orc_ni(Ty _a, Ty _b)
{
const Ty aorb = simd_or(_a, _b);
const Ty mffff = simd_isplat<Ty>(UINT32_MAX);
const Ty result = simd_xor(aorb, mffff);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_neg_ni(Ty _a)
{
const Ty zero = simd_zero<Ty>();
const Ty result = simd_sub(zero, _a);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_selb_ni(Ty _mask, Ty _a, Ty _b)
{
const Ty sel_a = simd_and(_a, _mask);
const Ty sel_b = simd_andc(_b, _mask);
const Ty result = simd_or(sel_a, sel_b);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_sels_ni(Ty _test, Ty _a, Ty _b)
{
const Ty mask = simd_sra(_test, 31);
const Ty result = simd_selb(mask, _a, _b);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_not_ni(Ty _a)
{
const Ty mffff = simd_isplat<Ty>(UINT32_MAX);
const Ty result = simd_xor(_a, mffff);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_min_ni(Ty _a, Ty _b)
{
const Ty mask = simd_cmplt(_a, _b);
const Ty result = simd_selb(mask, _a, _b);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_max_ni(Ty _a, Ty _b)
{
const Ty mask = simd_cmpgt(_a, _b);
const Ty result = simd_selb(mask, _a, _b);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_abs_ni(Ty _a)
{
const Ty a_neg = simd_neg(_a);
const Ty result = simd_max(a_neg, _a);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_imin_ni(Ty _a, Ty _b)
{
const Ty mask = simd_icmplt(_a, _b);
const Ty result = simd_selb(mask, _a, _b);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_imax_ni(Ty _a, Ty _b)
{
const Ty mask = simd_icmpgt(_a, _b);
const Ty result = simd_selb(mask, _a, _b);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_clamp_ni(Ty _a, Ty _min, Ty _max)
{
const Ty tmp = simd_min(_a, _max);
const Ty result = simd_max(tmp, _min);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_lerp_ni(Ty _a, Ty _b, Ty _s)
{
const Ty ba = simd_sub(_b, _a);
const Ty result = simd_madd(_s, ba, _a);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_sqrt_nr_ni(Ty _a)
{
const Ty half = simd_splat<Ty>(0.5f);
const Ty one = simd_splat<Ty>(1.0f);
const Ty tmp0 = simd_rsqrt_est(_a);
const Ty tmp1 = simd_mul(tmp0, _a);
const Ty tmp2 = simd_mul(tmp1, half);
const Ty tmp3 = simd_nmsub(tmp0, tmp1, one);
const Ty result = simd_madd(tmp3, tmp2, tmp1);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_sqrt_nr1_ni(Ty _a)
{
const Ty half = simd_splat<Ty>(0.5f);
Ty result = _a;
for (uint32_t ii = 0; ii < 11; ++ii)
{
const Ty tmp1 = simd_div(_a, result);
const Ty tmp2 = simd_add(tmp1, result);
result = simd_mul(tmp2, half);
}
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_rsqrt_ni(Ty _a)
{
const Ty one = simd_splat<Ty>(1.0f);
const Ty sqrt = simd_sqrt(_a);
const Ty result = simd_div(one, sqrt);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_rsqrt_nr_ni(Ty _a)
{
const Ty rsqrt = simd_rsqrt_est(_a);
const Ty iter0 = simd_mul(_a, rsqrt);
const Ty iter1 = simd_mul(iter0, rsqrt);
const Ty half = simd_splat<Ty>(0.5f);
const Ty half_rsqrt = simd_mul(half, rsqrt);
const Ty three = simd_splat<Ty>(3.0f);
const Ty three_sub_iter1 = simd_sub(three, iter1);
const Ty result = simd_mul(half_rsqrt, three_sub_iter1);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_rsqrt_carmack_ni(Ty _a)
{
const Ty half = simd_splat<Ty>(0.5f);
const Ty ah = simd_mul(half, _a);
const Ty ashift = simd_sra(_a, 1);
const Ty magic = simd_isplat<Ty>(0x5f3759df);
const Ty msuba = simd_isub(magic, ashift);
const Ty msubasq = simd_mul(msuba, msuba);
const Ty tmp0 = simd_splat<Ty>(1.5f);
const Ty tmp1 = simd_mul(ah, msubasq);
const Ty tmp2 = simd_sub(tmp0, tmp1);
const Ty result = simd_mul(msuba, tmp2);
return result;
}
namespace simd_logexp_detail
{
template<typename Ty>
BX_SIMD_INLINE Ty simd_poly1(Ty _a, float _b, float _c)
{
const Ty bbbb = simd_splat<Ty>(_b);
const Ty cccc = simd_splat<Ty>(_c);
const Ty result = simd_madd(cccc, _a, bbbb);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_poly2(Ty _a, float _b, float _c, float _d)
{
const Ty bbbb = simd_splat<Ty>(_b);
const Ty poly = simd_poly1(_a, _c, _d);
const Ty result = simd_madd(poly, _a, bbbb);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_poly3(Ty _a, float _b, float _c, float _d, float _e)
{
const Ty bbbb = simd_splat<Ty>(_b);
const Ty poly = simd_poly2(_a, _c, _d, _e);
const Ty result = simd_madd(poly, _a, bbbb);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_poly4(Ty _a, float _b, float _c, float _d, float _e, float _f)
{
const Ty bbbb = simd_splat<Ty>(_b);
const Ty poly = simd_poly3(_a, _c, _d, _e, _f);
const Ty result = simd_madd(poly, _a, bbbb);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_poly5(Ty _a, float _b, float _c, float _d, float _e, float _f, float _g)
{
const Ty bbbb = simd_splat<Ty>(_b);
const Ty poly = simd_poly4(_a, _c, _d, _e, _f, _g);
const Ty result = simd_madd(poly, _a, bbbb);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_logpoly(Ty _a)
{
#if 1
const Ty result = simd_poly5(_a
, 3.11578814719469302614f, -3.32419399085241980044f
, 2.59883907202499966007f, -1.23152682416275988241f
, 0.318212422185251071475f, -0.0344359067839062357313f
);
#elif 0
const Ty result = simd_poly4(_a
, 2.8882704548164776201f, -2.52074962577807006663f
, 1.48116647521213171641f, -0.465725644288844778798f
, 0.0596515482674574969533f
);
#elif 0
const Ty result = simd_poly3(_a
, 2.61761038894603480148f, -1.75647175389045657003f
, 0.688243882994381274313f, -0.107254423828329604454f
);
#else
const Ty result = simd_poly2(_a
, 2.28330284476918490682f, -1.04913055217340124191f
, 0.204446009836232697516f
);
#endif
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_exppoly(Ty _a)
{
#if 1
const Ty result = simd_poly5(_a
, 9.9999994e-1f, 6.9315308e-1f
, 2.4015361e-1f, 5.5826318e-2f
, 8.9893397e-3f, 1.8775767e-3f
);
#elif 0
const Ty result = simd_poly4(_a
, 1.0000026f, 6.9300383e-1f
, 2.4144275e-1f, 5.2011464e-2f
, 1.3534167e-2f
);
#elif 0
const Ty result = simd_poly3(_a
, 9.9992520e-1f, 6.9583356e-1f
, 2.2606716e-1f, 7.8024521e-2f
);
#else
const Ty result = simd_poly2(_a
, 1.0017247f, 6.5763628e-1f
, 3.3718944e-1f
);
#endif // 0
return result;
}
} // namespace simd_internal
template<typename Ty>
BX_SIMD_INLINE Ty simd_log2_ni(Ty _a)
{
const Ty expmask = simd_isplat<Ty>(0x7f800000);
const Ty mantmask = simd_isplat<Ty>(0x007fffff);
const Ty one = simd_splat<Ty>(1.0f);
const Ty c127 = simd_isplat<Ty>(127);
const Ty aexp = simd_and(_a, expmask);
const Ty aexpsr = simd_srl(aexp, 23);
const Ty tmp0 = simd_isub(aexpsr, c127);
const Ty exp = simd_itof(tmp0);
const Ty amask = simd_and(_a, mantmask);
const Ty mant = simd_or(amask, one);
const Ty poly = simd_logexp_detail::simd_logpoly(mant);
const Ty mandiff = simd_sub(mant, one);
const Ty result = simd_madd(poly, mandiff, exp);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_exp2_ni(Ty _a)
{
const Ty min = simd_splat<Ty>( 129.0f);
const Ty max = simd_splat<Ty>(-126.99999f);
const Ty tmp0 = simd_min(_a, min);
const Ty aaaa = simd_max(tmp0, max);
const Ty half = simd_splat<Ty>(0.5f);
const Ty tmp2 = simd_sub(aaaa, half);
const Ty ipart = simd_ftoi(tmp2);
const Ty iround = simd_itof(ipart);
const Ty fpart = simd_sub(aaaa, iround);
const Ty c127 = simd_isplat<Ty>(127);
const Ty tmp5 = simd_iadd(ipart, c127);
const Ty expipart = simd_sll(tmp5, 23);
const Ty expfpart = simd_logexp_detail::simd_exppoly(fpart);
const Ty result = simd_mul(expipart, expfpart);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_pow_ni(Ty _a, Ty _b)
{
const Ty alog2 = simd_log2(_a);
const Ty alog2b = simd_mul(alog2, _b);
const Ty result = simd_exp2(alog2b);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_dot3_ni(Ty _a, Ty _b)
{
const Ty xyzw = simd_mul(_a, _b);
const Ty xxxx = simd_swiz_xxxx(xyzw);
const Ty yyyy = simd_swiz_yyyy(xyzw);
const Ty zzzz = simd_swiz_zzzz(xyzw);
const Ty tmp1 = simd_add(xxxx, yyyy);
const Ty result = simd_add(zzzz, tmp1);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_cross3_ni(Ty _a, Ty _b)
{
// a.yzx * b.zxy - a.zxy * b.yzx == (a * b.yzx - a.yzx * b).yzx
#if 0
const Ty a_yzxw = simd_swiz_yzxw(_a);
const Ty a_zxyw = simd_swiz_zxyw(_a);
const Ty b_zxyw = simd_swiz_zxyw(_b);
const Ty b_yzxw = simd_swiz_yzxw(_b);
const Ty tmp = simd_mul(a_yzxw, b_zxyw);
const Ty result = simd_nmsub(a_zxyw, b_yzxw, tmp);
#else
const Ty a_yzxw = simd_swiz_yzxw(_a);
const Ty b_yzxw = simd_swiz_yzxw(_b);
const Ty tmp0 = simd_mul(_a, b_yzxw);
const Ty tmp1 = simd_nmsub(a_yzxw, _b, tmp0);
const Ty result = simd_swiz_yzxw(tmp1);
#endif
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_normalize3_ni(Ty _a)
{
const Ty dot3 = simd_dot3(_a, _a);
const Ty invSqrt = simd_rsqrt(dot3);
const Ty result = simd_mul(_a, invSqrt);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_dot_ni(Ty _a, Ty _b)
{
const Ty xyzw = simd_mul(_a, _b);
const Ty yzwx = simd_swiz_yzwx(xyzw);
const Ty tmp0 = simd_add(xyzw, yzwx);
const Ty zwxy = simd_swiz_zwxy(tmp0);
const Ty result = simd_add(tmp0, zwxy);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_ceil_ni(Ty _a)
{
const Ty tmp0 = simd_ftoi(_a);
const Ty tmp1 = simd_itof(tmp0);
const Ty mask = simd_cmplt(tmp1, _a);
const Ty one = simd_splat<Ty>(1.0f);
const Ty tmp2 = simd_and(one, mask);
const Ty result = simd_add(tmp1, tmp2);
return result;
}
template<typename Ty>
BX_SIMD_INLINE Ty simd_floor_ni(Ty _a)
{
const Ty tmp0 = simd_ftoi(_a);
const Ty tmp1 = simd_itof(tmp0);
const Ty mask = simd_cmpgt(tmp1, _a);
const Ty one = simd_splat<Ty>(1.0f);
const Ty tmp2 = simd_and(one, mask);
const Ty result = simd_sub(tmp1, tmp2);
return result;
}
template<typename Ty>
BX_SIMD_FORCE_INLINE Ty simd_round_ni(Ty _a)
{
const Ty tmp = simd_ftoi(_a);
const Ty result = simd_itof(tmp);
return result;
}
template<typename Ty>
BX_SIMD_INLINE bool simd_test_any_ni(Ty _a)
{
const Ty mask = simd_sra(_a, 31);
const Ty zwxy = simd_swiz_zwxy(mask);
const Ty tmp0 = simd_or(mask, zwxy);
const Ty tmp1 = simd_swiz_yyyy(tmp0);
const Ty tmp2 = simd_or(tmp0, tmp1);
int res;
simd_stx(&res, tmp2);
return 0 != res;
}
template<typename Ty>
BX_SIMD_INLINE bool simd_test_all_ni(Ty _a)
{
const Ty bits = simd_sra(_a, 31);
const Ty m1248 = simd_ild<Ty>(1, 2, 4, 8);
const Ty mask = simd_and(bits, m1248);
const Ty zwxy = simd_swiz_zwxy(mask);
const Ty tmp0 = simd_or(mask, zwxy);
const Ty tmp1 = simd_swiz_yyyy(tmp0);
const Ty tmp2 = simd_or(tmp0, tmp1);
int res;
simd_stx(&res, tmp2);
return 0xf == res;
}
} // namespace bx
#endif // BX_SIMD_NI_H_HEADER_GUARD