diff options
| author | Thiago Macieira <thiago.macieira@intel.com> | 2017-04-03 15:41:44 -0700 |
|---|---|---|
| committer | Thiago Macieira <thiago.macieira@intel.com> | 2017-04-06 04:15:04 +0000 |
| commit | c817b33b45f2886c89d8a768c61ffffaa46f67a4 (patch) | |
| tree | 98e39e893502c4c03b54ba8d00276aed701910e1 | |
| parent | 264d814773a15806df497e872e4b19c613c94725 (diff) | |
f16c: Use the packed intrinsics instead of scalar ones
MSVC, Apple's Clang and Clang prior to 3.9 do not recognize _cvtss_sh
and _cvtsh_ss. So expand the operation to use directly the packed
intrinsics.
Change-Id: I27b55fdf514247549455fffd14b2046fd638593d
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
| -rw-r--r-- | config.tests/common/f16c/f16c.cpp | 11 | ||||
| -rw-r--r-- | src/corelib/global/qfloat16.h | 8 |
2 files changed, 10 insertions, 9 deletions
diff --git a/config.tests/common/f16c/f16c.cpp b/config.tests/common/f16c/f16c.cpp index fc73e4fc55..aeeb35eac3 100644 --- a/config.tests/common/f16c/f16c.cpp +++ b/config.tests/common/f16c/f16c.cpp @@ -41,14 +41,11 @@ int main(int, char**) { - float f = 1.f; - unsigned short s = _cvtss_sh(f, 0); - float g = _cvtsh_ss(s); - bool result = f == g; - (void)result; __m128i a = _mm_setzero_si128(); - __m256 b = _mm256_cvtph_ps(a); - __m128i c = _mm256_cvtps_ph(b, 0); + __m128 b = _mm_cvtph_ps(a); + __m256 b256 = _mm256_cvtph_ps(a); + __m128i c = _mm_cvtps_ph(b, 0); + c = _mm256_cvtps_ph(b256, 0); (void)c; return 0; } diff --git a/src/corelib/global/qfloat16.h b/src/corelib/global/qfloat16.h index 0fa83db93f..654183acd6 100644 --- a/src/corelib/global/qfloat16.h +++ b/src/corelib/global/qfloat16.h @@ -118,7 +118,9 @@ QT_WARNING_DISABLE_CLANG("-Wc99-extensions") inline qfloat16::qfloat16(float f) Q_DECL_NOTHROW { #if defined(QT_COMPILER_SUPPORTS_F16C) && defined(__F16C__) - b16 = _cvtss_sh(f, 0); + __m128 packsingle = _mm_set_ss(f); + __m128i packhalf = _mm_cvtps_ph(packsingle, 0); + b16 = _mm_extract_epi16(packhalf, 0); #elif defined (__ARM_FP16_FORMAT_IEEE) __fp16 f16 = f; memcpy(&b16, &f16, sizeof(quint16)); @@ -134,7 +136,9 @@ QT_WARNING_POP inline qfloat16::operator float() const Q_DECL_NOTHROW { #if defined(QT_COMPILER_SUPPORTS_F16C) && defined(__F16C__) - return _cvtsh_ss(b16); + __m128i packhalf = _mm_cvtsi32_si128(b16); + __m128 packsingle = _mm_cvtph_ps(packhalf); + return _mm_cvtss_f32(packsingle); #elif defined (__ARM_FP16_FORMAT_IEEE) __fp16 f16; memcpy(&f16, &b16, sizeof(quint16)); |