diff options
author | Thiago Macieira <thiago.macieira@intel.com> | 2022-08-05 15:58:46 -0700 |
---|---|---|
committer | Thiago Macieira <thiago.macieira@intel.com> | 2022-11-28 10:59:21 -0800 |
commit | beab4d30e24442fa7c0c3af5056b0e064b3acc95 (patch) | |
tree | f4903074283d84282d12873db28fb6f65ae2bf2f /src/corelib/global | |
parent | 99c7f0419e66692260be56c0385badeacb3f6760 (diff) |
qfloat16: add a couple faster implementations of qSqrt
Instead of going through float.
Change-Id: Ie1b556d9ebca4ccaadd2fffd170895088a5d2dec
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
Diffstat (limited to 'src/corelib/global')
-rw-r--r-- | src/corelib/global/qfloat16.h | 24 |
1 files changed, 23 insertions, 1 deletions
diff --git a/src/corelib/global/qfloat16.h b/src/corelib/global/qfloat16.h index 10dcfe0f8d..02ad5f303f 100644 --- a/src/corelib/global/qfloat16.h +++ b/src/corelib/global/qfloat16.h @@ -243,13 +243,35 @@ Q_CORE_EXPORT void qFloatFromFloat16(float *, const qfloat16 *, qsizetype length // https://wg21.link/p1467 - disabled until tested using namespace std; return sqrt(f); -#endif +#elif QFLOAT16_IS_NATIVE && defined(__HAVE_FLOAT16) && __HAVE_FLOAT16 + // This C library (glibc) has sqrtf16(). + return sqrtf16(f); +#else + bool mathUpdatesErrno = true; +# if defined(__NO_MATH_ERRNO__) || defined(_M_FP_FAST) + mathUpdatesErrno = false; +# elif defined(math_errhandling) + mathUpdatesErrno = (math_errhandling & MATH_ERRNO); +# endif + + // We don't need to set errno to EDOM if (f >= 0 && f != -0 && !isnan(f)) + // (or if we don't care about errno in the first place). We can merge the + // NaN check with by negating and inverting: !(0 > f), and leaving zero to + // sqrtf(). + if (!mathUpdatesErrno || !(0 > f)) { +# if defined(__AVX512FP16__) + __m128h v = _mm_set_sh(f); + v = _mm_sqrt_sh(v, v); + return _mm_cvtsh_h(v); +# endif + } // WG14's N2601 does not provide a way to tell which types an // implementation supports, so we assume it doesn't and fall back to FP32 float f32 = float(f); f32 = sqrtf(f32); return qfloat16::NearestFloat(f32); +#endif } // The remainder of these utility functions complement qglobal.h |