| 21 | /// This type is ensured to be 16-bit size. |
| 22 | #if 1 // ndef __ARM_FP16_FORMAT_IEEE |
| 23 | class Half final { |
| 24 | public: |
| 25 | /// Whether `Half` has a full native HW support. |
| 26 | static constexpr bool kNative = false; |
| 27 | /// Whether `Half` has a native HW support for single/double conversion. |
| 28 | template<typename T> |
| 29 | static constexpr bool kNativeConversion = ( |
| 30 | ( |
| 31 | std::is_same_v<T, float> && |
| 32 | #if defined(NPY_HAVE_FP16) || defined(NPY_HAVE_VSX3) |
| 33 | true |
| 34 | #else |
| 35 | false |
| 36 | #endif |
| 37 | ) || ( |
| 38 | std::is_same_v<T, double> && |
| 39 | #if defined(NPY_HAVE_AVX512FP16) || (defined(NPY_HAVE_VSX3) && defined(NPY_HAVE_VSX3_HALF_DOUBLE)) |
| 40 | true |
| 41 | #else |
| 42 | false |
| 43 | #endif |
| 44 | ) |
| 45 | ); |
| 46 | |
| 47 | /// Default constructor. initialize nothing. |
| 48 | Half() = default; |
| 49 | |
| 50 | /// Construct from float |
| 51 | /// If there are no hardware optimization available, rounding will always |
| 52 | /// be set to ties to even. |
| 53 | explicit Half(float f) |
| 54 | { |
| 55 | #if defined(NPY_HAVE_FP16) |
| 56 | __m128 mf = _mm_load_ss(&f); |
| 57 | bits_ = static_cast<uint16_t>(_mm_cvtsi128_si32(_mm_cvtps_ph(mf, _MM_FROUND_TO_NEAREST_INT))); |
| 58 | #elif defined(NPY_HAVE_VSX3) && defined(NPY_HAVE_VSX_ASM) |
| 59 | __vector float vf32 = vec_splats(f); |
| 60 | __vector unsigned short vf16; |
| 61 | __asm__ __volatile__ ("xvcvsphp %x0,%x1" : "=wa" (vf16) : "wa" (vf32)); |
| 62 | #ifdef __BIG_ENDIAN__ |
| 63 | bits_ = vec_extract(vf16, 1); |
| 64 | #else |
| 65 | bits_ = vec_extract(vf16, 0); |
| 66 | #endif |
| 67 | #else |
| 68 | bits_ = half_private::FromFloatBits(BitCast<uint32_t>(f)); |
| 69 | #endif |
| 70 | } |
| 71 | |
| 72 | /// Construct from double. |
| 73 | /// If there are no hardware optimization available, rounding will always |
| 74 | /// be set to ties to even. |
| 75 | explicit Half(double f) |
| 76 | { |
| 77 | #if defined(NPY_HAVE_AVX512FP16) |
| 78 | __m128d md = _mm_load_sd(&f); |
| 79 | bits_ = static_cast<uint16_t>(_mm_cvtsi128_si32(_mm_castph_si128(_mm_cvtpd_ph(md)))); |
| 80 | #elif defined(NPY_HAVE_VSX3) && defined(NPY_HAVE_VSX3_HALF_DOUBLE) |
no test coverage detected