diff options
| author | Thiago Macieira <thiago.macieira@intel.com> | 2021-09-28 17:26:09 -0700 |
|---|---|---|
| committer | Thiago Macieira <thiago.macieira@intel.com> | 2021-10-13 12:49:49 -0700 |
| commit | 0df7831d01b03910ae74a432c2a07a29d99fc315 (patch) | |
| tree | 70348515fd1812c3712b9815f661a42069e805c1 /src/corelib/global/qfloat16.cpp | |
| parent | cb0ea3989598a83b2326f63e7da1669018a965e5 (diff) | |
qfloat16: re-merge the F16C content into qfloat16.cpp
The bugs we had noticed previously are believed to be fixed. MSVC will
do per-function updating of AVX content where necessary and the ICC
issue is no longer relevant.
Change-Id: I2bbf422288924c198645fffd16a9235f2d73cc19
Reviewed-by: Alexandru Croitor <alexandru.croitor@qt.io>
Diffstat (limited to 'src/corelib/global/qfloat16.cpp')
| -rw-r--r-- | src/corelib/global/qfloat16.cpp | 42 |
1 files changed, 29 insertions, 13 deletions
diff --git a/src/corelib/global/qfloat16.cpp b/src/corelib/global/qfloat16.cpp index 9ef197b3a1..834d82e0ea 100644 --- a/src/corelib/global/qfloat16.cpp +++ b/src/corelib/global/qfloat16.cpp @@ -197,17 +197,36 @@ static inline bool hasFastF16() return qCpuHasFeature(F16C) && qCpuHasFeature(AVX); } -extern "C" { -#ifdef QFLOAT16_INCLUDE_FAST -# define f16cextern static -#else -# define f16cextern extern -#endif - -f16cextern void qFloatToFloat16_fast(quint16 *out, const float *in, qsizetype len) noexcept; -f16cextern void qFloatFromFloat16_fast(float *out, const quint16 *in, qsizetype len) noexcept; +QT_FUNCTION_TARGET(F16C) +void qFloatToFloat16_fast(quint16 *out, const float *in, qsizetype len) Q_DECL_NOEXCEPT +{ + qsizetype i = 0; + int epilog_i; + for (; i < len - 7; i += 8) + _mm_storeu_si128((__m128i *)(out + i), _mm256_cvtps_ph(_mm256_loadu_ps(in + i), 0)); + if (i < len - 3) { + _mm_storel_epi64((__m128i *)(out + i), _mm_cvtps_ph(_mm_loadu_ps(in + i), 0)); + i += 4; + } + // Inlining "qfloat16::qfloat16(float f)": + for (epilog_i = 0; i < len && epilog_i < 3; ++i, ++epilog_i) + out[i] = _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(in[i]), 0), 0); +} -#undef f16cextern +QT_FUNCTION_TARGET(F16C) +void qFloatFromFloat16_fast(float *out, const quint16 *in, qsizetype len) Q_DECL_NOEXCEPT +{ + qsizetype i = 0; + int epilog_i; + for (; i < len - 7; i += 8) + _mm256_storeu_ps(out + i, _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(in + i)))); + if (i < len - 3) { + _mm_storeu_ps(out + i, _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)(in + i)))); + i += 4; + } + // Inlining "qfloat16::operator float()": + for (epilog_i = 0; i < len && epilog_i < 3; ++i, ++epilog_i) + out[i] = _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(in[i]))); } #elif defined(__ARM_FP16_FORMAT_IEEE) && defined(__ARM_NEON__) && (__ARM_FP & 2) @@ -292,6 +311,3 @@ Q_CORE_EXPORT void qFloatFromFloat16(float *out, const qfloat16 *in, qsizetype l QT_END_NAMESPACE #include "qfloat16tables.cpp" -#ifdef QFLOAT16_INCLUDE_FAST -# include "qfloat16_f16c.c" -#endif |