Catcrafts/Crafter.Math
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

more F16 math

Browse source
This commit is contained in:
Jorijn van der Graaf 2026-03-19 05:53:17 +01:00
commit f1fbbe0faf
3 changed files with 82 additions and 42 deletions
Download patch file Download diff file Expand all files Collapse all files

58
interfaces/Crafter.Math-VectorF16.cppm
View file

@ -580,7 +580,7 @@ namespace Crafter {
}
}
constexpr static std::tuple<VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>> Normalize(
constexpr static std::tuple<VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>> Normalize(
VectorF16<Len, Packing, Repeats> A,
VectorF16<Len, Packing, Repeats> C,
VectorF16<Len, Packing, Repeats> E,
@ -705,15 +705,15 @@ namespace Crafter {
__m512h fLenghtG = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecG));
return {
_mm512_mul_ph(A.v, fLenghtA),
_mm512_mul_ph(C.v, fLenghtC),
_mm512_mul_ph(E.v, fLenghtE),
_mm512_mul_ph(G.v, fLenghtG),
VectorF16<Len, Packing, Repeats>(_mm512_mul_ph(A.v, fLenghtA)),
VectorF16<Len, Packing, Repeats>(_mm512_mul_ph(C.v, fLenghtC)),
VectorF16<Len, Packing, Repeats>(_mm512_mul_ph(E.v, fLenghtE)),
VectorF16<Len, Packing, Repeats>(_mm512_mul_ph(G.v, fLenghtG)),
};
}
}
constexpr static std::tuple<VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>> Normalize(
constexpr static std::tuple<VectorF16<Len, Packing, Repeats>, VectorF16<Len, Packing, Repeats>> Normalize(
VectorF16<Len, Packing, Repeats> A,
VectorF16<Len, Packing, Repeats> E
) requires(Packing == 4) {
@ -932,6 +932,36 @@ namespace Crafter {
}
}
constexpr static VectorF16<Len, Packing, Repeats> Length(
VectorF16<Len, Packing, Repeats> A,
VectorF16<Len, Packing, Repeats> C,
VectorF16<Len, Packing, Repeats> E,
VectorF16<Len, Packing, Repeats> G
) requires(Packing == 2) {
VectorF16<Len, Packing, Repeats> lenghtSq = LengthSq(A, C, E, G);
if constexpr(std::is_same_v<VectorType, __m128h>) {
return VectorF16<Len, Packing, Repeats>(_mm_sqrt_ph(lenghtSq.v));
} else if constexpr(std::is_same_v<VectorType, __m256h>) {
return VectorF16<Len, Packing, Repeats>(_mm256_sqrt_ph(lenghtSq.v));
} else {
return VectorF16<Len, Packing, Repeats>(_mm512_sqrt_ph(lenghtSq.v));
}
}
constexpr static VectorF16<Len, Packing, Repeats> Length(
VectorF16<Len, Packing, Repeats> A,
VectorF16<Len, Packing, Repeats> E
) requires(Packing == 2) {
VectorF16<Len, Packing, Repeats> lenghtSq = LengthSq(A, E);
if constexpr(std::is_same_v<VectorType, __m128h>) {
return VectorF16<Len, Packing, Repeats>(_mm_sqrt_ph(lenghtSq.v));
} else if constexpr(std::is_same_v<VectorType, __m256h>) {
return VectorF16<Len, Packing, Repeats>(_mm256_sqrt_ph(lenghtSq.v));
} else {
return VectorF16<Len, Packing, Repeats>(_mm512_sqrt_ph(lenghtSq.v));
}
}
constexpr static VectorF16<Len, Packing, Repeats> LengthSq(
VectorF16<Len, Packing, Repeats> A,
VectorF16<Len, Packing, Repeats> B,
@ -945,6 +975,22 @@ namespace Crafter {
return Dot(A, A, B, B, C, C, D, D, E, E, F, F, G, G, H, H);
}
constexpr static VectorF16<Len, Packing, Repeats> LengthSq(
VectorF16<Len, Packing, Repeats> A,
VectorF16<Len, Packing, Repeats> C,
VectorF16<Len, Packing, Repeats> E,
VectorF16<Len, Packing, Repeats> G
) requires(Packing == 2) {
return Dot(A, A, C, C, E, E, G, G);
}
constexpr static VectorF16<Len, Packing, Repeats> LengthSq(
VectorF16<Len, Packing, Repeats> A,
VectorF16<Len, Packing, Repeats> E
) requires(Packing == 4) {
return Dot(A, A, E, E);
}
constexpr static VectorF16<Len, Packing, Repeats> Dot(
VectorF16<Len, Packing, Repeats> A0, VectorF16<Len, Packing, Repeats> A1,
VectorF16<Len, Packing, Repeats> B0, VectorF16<Len, Packing, Repeats> B1,