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

F16 fixes

Browse source
This commit is contained in:
Jorijn van der Graaf 2026-03-24 03:48:56 +01:00
commit 6e6530290b
3 changed files with 189 additions and 506 deletions
Download patch file Download diff file Expand all files Collapse all files

40
interfaces/Crafter.Math-Vector.cppm
View file

@ -458,34 +458,24 @@ namespace Crafter {
}
template <typename T, std::uint32_t Aligment>
struct std::formatter<Crafter::Vector<T, 2, Aligment>> : std::formatter<std::string> {
auto format(const Crafter::Vector<T, 2, Aligment>& obj, format_context& ctx) const {
return std::formatter<std::string>::format(std::format("{{{}, {}}}",
obj.x, obj.y
), ctx);
template <typename T, std::uint32_t Len, std::uint32_t Aligment>
struct std::formatter<Crafter::Vector<T, Len, Aligment>> : std::formatter<std::string> {
auto format(const Crafter::Vector<T, Len, Aligment>& vec, format_context& ctx) const {
std::string out = "{";
for(std::uint32_t i2 = 0; i2 < Len; i2++) {
if constexpr(std::same_as<T, _Float16>) {
out += std::format("{}", static_cast<float>(vec.v[i2]));
} else {
out += std::format("{}", vec.v[i2]);
}
if (i2 + 1 < Len) out += ",";
}
out += "}";
return std::formatter<std::string>::format(out, ctx);
}
};
template <typename T, std::uint32_t Aligment>
struct std::formatter<Crafter::Vector<T, 3, Aligment>> : std::formatter<std::string> {
auto format(const Crafter::Vector<T, 3, Aligment>& obj, format_context& ctx) const {
return std::formatter<std::string>::format(std::format("{{{}, {}, {}}}",
obj.x, obj.y, obj.z
), ctx);
}
};
template <typename T, std::uint32_t Aligment>
struct std::formatter<Crafter::Vector<T, 4, Aligment>> : std::formatter<std::string> {
auto format(const Crafter::Vector<T, 4, Aligment>& obj, format_context& ctx) const {
return std::formatter<std::string>::format(std::format("{{{}, {}, {}, {}}}",
(float)obj.x, (float)obj.y, (float)obj.z, (float)obj.w
), ctx);
}
};
template <typename T, std::uint32_t Len, std::uint32_t Aligment, typename BT>
constexpr Crafter::Vector<T, Len, Aligment> operator*(BT b, const Crafter::Vector<T, Len, Aligment>& v) requires std::is_arithmetic_v<BT> {
return v * b;

239
interfaces/Crafter.Math-VectorF16.cppm
View file

@ -45,6 +45,8 @@ namespace Crafter {
VectorType v;
public:
template <std::uint32_t Len2, std::uint32_t Packing2>
friend class VectorF16;
static constexpr std::uint32_t MaxSize = 32;
static constexpr std::uint8_t Alignment = GetAlingment();
static_assert(Len * Packing <= MaxSize, "Len * Packing exceeds MaxSize");
@ -256,18 +258,18 @@ namespace Crafter {
if constexpr(std::is_same_v<VectorType, __m128h>) {
_Float16 dot = LengthSq();
__m128h vec = _mm_set1_ph(dot);
__m128h sqrt = _mm_rsqrt_ph(vec);
v = _mm_div_ps(v, sqrt);
__m128h sqrt = _mm_sqrt_ph(vec);
v = _mm_div_ph(v, sqrt);
} else if constexpr(std::is_same_v<VectorType, __m256h>) {
_Float16 dot = LengthSq();
__m256h vec = _mm256_set1_ph(dot);
__m256h sqrt = _mm256_rsqrt_ph(vec);
v = _mm256_div_ps(v, sqrt);
__m256h sqrt = _mm256_sqrt_ph(vec);
v = _mm256_div_ph(v, sqrt);
} else {
_Float16 dot = LengthSq();
__m512h vec = _mm512_set1_ph(dot);
__m512h sqrt = _mm512_rsqrt_ph(vec);
v = _mm512_div_ps(v, sqrt);
__m512h sqrt = _mm512_sqrt_ph(vec);
v = _mm512_div_ph(v, sqrt);
}
}
@ -323,42 +325,42 @@ namespace Crafter {
}
}
std::tuple<VectorF16<Len, Packing>, VectorF16<Len, Packing>> SinCos() {
if constexpr (std::is_same_v<VectorType, __m128h>) {
__m256 wide = _mm256_cvtph_ps(_mm_castph_si128(v));
__m256 s, c;
sincos_f32x8(wide, s, c);
return {
VectorF16<Len, Packing>(_mm_castsi128_ph(_mm256_cvtps_ph(s, _MM_FROUND_TO_NEAREST_INT))),
VectorF16<Len, Packing>(_mm_castsi128_ph(_mm256_cvtps_ph(c, _MM_FROUND_TO_NEAREST_INT)))
};
std::tuple<VectorF16<Len, Packing>, VectorF16<Len, Packing>> SinCos() {
if constexpr (std::is_same_v<VectorType, __m128h>) {
__m256 wide = _mm256_cvtph_ps(_mm_castph_si128(v));
__m256 s, c;
sincos_f32x8(wide, s, c);
return {
VectorF16<Len, Packing>(_mm_castsi128_ph(_mm256_cvtps_ph(s, _MM_FROUND_TO_NEAREST_INT))),
VectorF16<Len, Packing>(_mm_castsi128_ph(_mm256_cvtps_ph(c, _MM_FROUND_TO_NEAREST_INT)))
};
} else if constexpr (std::is_same_v<VectorType, __m256h>) {
__m512 wide = _mm512_cvtph_ps(_mm256_castph_si256(v));
__m512 s, c;
sincos_f32x16(wide, s, c);
return {
VectorF16<Len, Packing>(_mm256_castsi256_ph(_mm512_cvtps_ph(s, _MM_FROUND_TO_NEAREST_INT))),
VectorF16<Len, Packing>(_mm256_castsi256_ph(_mm512_cvtps_ph(c, _MM_FROUND_TO_NEAREST_INT)))
};
} else if constexpr (std::is_same_v<VectorType, __m256h>) {
__m512 wide = _mm512_cvtph_ps(_mm256_castph_si256(v));
__m512 s, c;
sincos_f32x16(wide, s, c);
return {
VectorF16<Len, Packing>(_mm256_castsi256_ph(_mm512_cvtps_ph(s, _MM_FROUND_TO_NEAREST_INT))),
VectorF16<Len, Packing>(_mm256_castsi256_ph(_mm512_cvtps_ph(c, _MM_FROUND_TO_NEAREST_INT)))
};
} else {
__m256i lo = _mm512_castsi512_si256(_mm512_castph_si512(v));
__m256i hi = _mm512_extracti64x4_epi64(_mm512_castph_si512(v), 1);
} else {
__m256i lo = _mm512_castsi512_si256(_mm512_castph_si512(v));
__m256i hi = _mm512_extracti64x4_epi64(_mm512_castph_si512(v), 1);
__m512 s_lo, c_lo, s_hi, c_hi;
sincos_f32x16(_mm512_cvtph_ps(lo), s_lo, c_lo);
sincos_f32x16(_mm512_cvtph_ps(hi), s_hi, c_hi);
__m512 s_lo, c_lo, s_hi, c_hi;
sincos_f32x16(_mm512_cvtph_ps(lo), s_lo, c_lo);
sincos_f32x16(_mm512_cvtph_ps(hi), s_hi, c_hi);
auto pack = [](__m256i lo_ph, __m256i hi_ph) {
return _mm512_castsi512_ph(_mm512_inserti64x4(_mm512_castsi256_si512(lo_ph), hi_ph, 1));
};
return {
VectorF16<Len, Packing>(pack(_mm512_cvtps_ph(s_lo, _MM_FROUND_TO_NEAREST_INT), _mm512_cvtps_ph(s_hi, _MM_FROUND_TO_NEAREST_INT))),
VectorF16<Len, Packing>(pack( _mm512_cvtps_ph(c_lo, _MM_FROUND_TO_NEAREST_INT), _mm512_cvtps_ph(c_hi, _MM_FROUND_TO_NEAREST_INT)))
};
auto pack = [](__m256i lo_ph, __m256i hi_ph) {
return _mm512_castsi512_ph(_mm512_inserti64x4(_mm512_castsi256_si512(lo_ph), hi_ph, 1));
};
return {
VectorF16<Len, Packing>(pack(_mm512_cvtps_ph(s_lo, _MM_FROUND_TO_NEAREST_INT), _mm512_cvtps_ph(s_hi, _MM_FROUND_TO_NEAREST_INT))),
VectorF16<Len, Packing>(pack( _mm512_cvtps_ph(c_lo, _MM_FROUND_TO_NEAREST_INT), _mm512_cvtps_ph(c_hi, _MM_FROUND_TO_NEAREST_INT)))
};
}
}
}
template <std::array<bool, Len> values>
constexpr VectorF16<Len, Packing> Negate() {
@ -421,12 +423,12 @@ namespace Crafter {
constexpr static VectorF16<Len, Packing> Cross(VectorF16<Len, Packing> a, VectorF16<Len, Packing> b) requires(Len == 3) {
if constexpr(std::is_same_v<VectorType, __m128h>) {
constexpr std::array<std::uint8_t, 16> shuffleMask1 = GetShuffleMaskEpi8<{1,2,0}>();
constexpr std::array<std::uint8_t, 16> shuffleMask1 = GetShuffleMaskEpi8<{{1,2,0}}>();
__m128i shuffleVec1 = _mm_loadu_epi8(shuffleMask1.data());
__m128h row1 = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(a.v), shuffleVec1));
__m128h row4 = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(b.v), shuffleVec1));
constexpr std::array<std::uint8_t, 16> shuffleMask3 = GetShuffleMaskEpi8<{2,0,1}>();
constexpr std::array<std::uint8_t, 16> shuffleMask3 = GetShuffleMaskEpi8<{{2,0,1}}>();
__m128i shuffleVec3 = _mm_loadu_epi8(shuffleMask3.data());
__m128h row3 = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(a.v), shuffleVec3));
__m128h row2 = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(b.v), shuffleVec3));
@ -434,12 +436,12 @@ namespace Crafter {
__m128h result = _mm_mul_ph(row3, row4);
return _mm_fmsub_ph(row1,row2,result);
} else if constexpr (std::is_same_v<VectorType, __m256h>) {
constexpr std::array<std::uint8_t, 32> shuffleMask1 = GetShuffleMaskEpi8<{1,2,0}>();
constexpr std::array<std::uint8_t, 32> shuffleMask1 = GetShuffleMaskEpi8<{{1,2,0}}>();
__m512i shuffleVec1 = _mm512_castsi256_si512(_mm256_loadu_epi8(shuffleMask1.data()));
__m256h row1 = _mm256_castsi256_ph(_mm512_castsi512_si256(_mm512_shuffle_epi8(_mm512_castsi256_si512(_mm256_castph_si256(a.v)), shuffleVec1)));
__m256h row4 = _mm256_castsi256_ph(_mm512_castsi512_si256(_mm512_shuffle_epi8(_mm512_castsi256_si512(_mm256_castph_si256(b.v)), shuffleVec1)));
constexpr std::array<std::uint8_t, 32> shuffleMask3 = GetShuffleMaskEpi8<{2,0,1}>();
constexpr std::array<std::uint8_t, 32> shuffleMask3 = GetShuffleMaskEpi8<{{2,0,1}}>();
__m512i shuffleVec3 = _mm512_castsi256_si512(_mm256_loadu_epi8(shuffleMask3.data()));
__m256h row3 = _mm256_castsi256_ph(_mm512_castsi512_si256(_mm512_shuffle_epi8(_mm512_castsi256_si512(_mm256_castph_si256(a.v)), shuffleVec3)));
@ -448,13 +450,13 @@ namespace Crafter {
__m256h result = _mm256_mul_ph(row3, row4);
return _mm256_fmsub_ph(row1,row2,result);
} else {
constexpr std::array<std::uint8_t, 64> shuffleMask1 = GetShuffleMaskEpi8<{1,2,0}>();
constexpr std::array<std::uint8_t, 64> shuffleMask1 = GetShuffleMaskEpi8<{{1,2,0}}>();
__m512i shuffleVec1 = _mm512_loadu_epi8(shuffleMask1.data());
__m512h row1 = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(a.v), shuffleVec1));
__m512h row4 = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(b.v), shuffleVec1));
constexpr std::array<std::uint8_t, 64> shuffleMask3 = GetShuffleMaskEpi8<{2,0,1}>();
constexpr std::array<std::uint8_t, 64> shuffleMask3 = GetShuffleMaskEpi8<{{2,0,1}}>();
__m512i shuffleVec3 = _mm512_loadu_epi8(shuffleMask3.data());
__m512h row3 = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(a.v), shuffleVec3));
@ -488,14 +490,14 @@ namespace Crafter {
VectorF16<Len, Packing> G,
VectorF16<Len, Packing> H
) requires(Len == 8) {
constexpr std::uint8_t shuffleMaskA[] = GetShuffleMaskEpi8<{0,0,0,0,0,0,0,0}>();
constexpr std::uint8_t shuffleMaskB[] = GetShuffleMaskEpi8<{1,1,1,1,1,1,1,1}>();
constexpr std::uint8_t shuffleMaskC[] = GetShuffleMaskEpi8<{2,2,2,2,2,2,2,2}>();
constexpr std::uint8_t shuffleMaskD[] = GetShuffleMaskEpi8<{3,3,3,3,3,3,3,3}>();
constexpr std::uint8_t shuffleMaskE[] = GetShuffleMaskEpi8<{4,4,4,4,4,4,4,4}>();
constexpr std::uint8_t shuffleMaskF[] = GetShuffleMaskEpi8<{5,5,5,5,5,5,5,5}>();
constexpr std::uint8_t shuffleMaskG[] = GetShuffleMaskEpi8<{6,6,6,6,6,6,6,6}>();
constexpr std::uint8_t shuffleMaskH[] = GetShuffleMaskEpi8<{7,7,7,7,7,7,7,7}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskA = GetShuffleMaskEpi8<{{0,0,0,0,0,0,0,0}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskB = GetShuffleMaskEpi8<{{1,1,1,1,1,1,1,1}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskC = GetShuffleMaskEpi8<{{2,2,2,2,2,2,2,2}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskD = GetShuffleMaskEpi8<{{3,3,3,3,3,3,3,3}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskE = GetShuffleMaskEpi8<{{4,4,4,4,4,4,4,4}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskF = GetShuffleMaskEpi8<{{5,5,5,5,5,5,5,5}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskG = GetShuffleMaskEpi8<{{6,6,6,6,6,6,6,6}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskH = GetShuffleMaskEpi8<{{7,7,7,7,7,7,7,7}}>();
if constexpr(std::is_same_v<VectorType, __m128h>) {
VectorF16<Len, Packing> lenght = Length(A, B, C, D, E, F, G, H);
@ -503,29 +505,28 @@ namespace Crafter {
__m128h one = _mm_loadu_ph(oneArr);
__m128h fLenght = _mm_div_ph(one, lenght.v);
__m128i shuffleVecA = _mm_loadu_epi8(shuffleMaskA);
__m128i shuffleVecA = _mm_loadu_epi8(shuffleMaskA.data());
__m128h fLenghtA = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecA));
__m128i shuffleVecB = _mm_loadu_epi8(shuffleMaskB);
__m128i shuffleVecB = _mm_loadu_epi8(shuffleMaskB.data());
__m128h fLenghtB = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecB));
__m128i shuffleVecC = _mm_loadu_epi8(shuffleMaskC);
__m128i shuffleVecC = _mm_loadu_epi8(shuffleMaskC.data());
__m128h fLenghtC = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecC));
__m128i shuffleVecD = _mm_loadu_epi8(shuffleMaskD);
__m128i shuffleVecD = _mm_loadu_epi8(shuffleMaskD.data());
__m128h fLenghtD = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecD));
__m128i shuffleVecE = _mm_loadu_epi8(shuffleMaskE);
__m128i shuffleVecE = _mm_loadu_epi8(shuffleMaskE.data());
__m128h fLenghtE = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecE));
__m128i shuffleVecF = _mm_loadu_epi8(shuffleMaskF);
__m128i shuffleVecF = _mm_loadu_epi8(shuffleMaskF.data());
__m128h fLenghtF = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecF));
__m128i shuffleVecG = _mm_loadu_epi8(shuffleMaskG);
__m128i shuffleVecG = _mm_loadu_epi8(shuffleMaskG.data());
__m128h fLenghtG = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecG));
__m128i shuffleVecH = _mm_loadu_epi8(shuffleMaskH);
__m128i shuffleVecH = _mm_loadu_epi8(shuffleMaskH.data());
__m128h fLenghtH = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecH));
return {
@ -544,28 +545,28 @@ namespace Crafter {
__m256h one = _mm256_loadu_ph(oneArr);
__m256h fLenght = _mm256_div_ph(one, lenght.v);
__m256i shuffleVecA = _mm256_loadu_epi8(shuffleMaskA);
__m256i shuffleVecA = _mm256_loadu_epi8(shuffleMaskA.data());
__m256h fLenghtA = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecA));
__m256i shuffleVecB = _mm256_loadu_epi8(shuffleMaskB);
__m256i shuffleVecB = _mm256_loadu_epi8(shuffleMaskB.data());
__m256h fLenghtB = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecB));
__m256i shuffleVecC = _mm256_loadu_epi8(shuffleMaskC);
__m256i shuffleVecC = _mm256_loadu_epi8(shuffleMaskC.data());
__m256h fLenghtC = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecC));
__m256i shuffleVecD = _mm256_loadu_epi8(shuffleMaskD);
__m256i shuffleVecD = _mm256_loadu_epi8(shuffleMaskD.data());
__m256h fLenghtD = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecD));
__m256i shuffleVecE = _mm256_loadu_epi8(shuffleMaskE);
__m256i shuffleVecE = _mm256_loadu_epi8(shuffleMaskE.data());
__m256h fLenghtE = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecE));
__m256i shuffleVecF = _mm256_loadu_epi8(shuffleMaskF);
__m256i shuffleVecF = _mm256_loadu_epi8(shuffleMaskF.data());
__m256h fLenghtF = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecF));
__m256i shuffleVecG = _mm256_loadu_epi8(shuffleMaskG);
__m256i shuffleVecG = _mm256_loadu_epi8(shuffleMaskG.data());
__m256h fLenghtG = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecG));
__m256i shuffleVecH = _mm256_loadu_epi8(shuffleMaskH);
__m256i shuffleVecH = _mm256_loadu_epi8(shuffleMaskH.data());
__m256h fLenghtH = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecH));
return {
@ -584,28 +585,28 @@ namespace Crafter {
__m512h one = _mm512_loadu_ph(oneArr);
__m512h fLenght = _mm512_div_ph(one, lenght.v);
__m512i shuffleVecA = _mm512_loadu_epi8(shuffleMaskA);
__m512i shuffleVecA = _mm512_loadu_epi8(shuffleMaskA.data());
__m512h fLenghtA = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecA));
__m512i shuffleVecB = _mm512_loadu_epi8(shuffleMaskB);
__m512i shuffleVecB = _mm512_loadu_epi8(shuffleMaskB.data());
__m512h fLenghtB = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecB));
__m512i shuffleVecC = _mm512_loadu_epi8(shuffleMaskC);
__m512i shuffleVecC = _mm512_loadu_epi8(shuffleMaskC.data());
__m512h fLenghtC = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecC));
__m512i shuffleVecD = _mm512_loadu_epi8(shuffleMaskD);
__m512i shuffleVecD = _mm512_loadu_epi8(shuffleMaskD.data());
__m512h fLenghtD = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecD));
__m512i shuffleVecE = _mm512_loadu_epi8(shuffleMaskE);
__m512i shuffleVecE = _mm512_loadu_epi8(shuffleMaskE.data());
__m512h fLenghtE = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecE));
__m512i shuffleVecF = _mm512_loadu_epi8(shuffleMaskF);
__m512i shuffleVecF = _mm512_loadu_epi8(shuffleMaskF.data());
__m512h fLenghtF = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecF));
__m512i shuffleVecG = _mm512_loadu_epi8(shuffleMaskG);
__m512i shuffleVecG = _mm512_loadu_epi8(shuffleMaskG.data());
__m512h fLenghtG = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecG));
__m512i shuffleVecH = _mm512_loadu_epi8(shuffleMaskH);
__m512i shuffleVecH = _mm512_loadu_epi8(shuffleMaskH.data());
__m512h fLenghtH = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecH));
return {
@ -627,10 +628,10 @@ namespace Crafter {
VectorF16<Len, Packing> E,
VectorF16<Len, Packing> G
) requires(Len == 4) {
constexpr std::uint8_t shuffleMaskA[] = GetShuffleMaskEpi8<{0,0,0,0}>();
constexpr std::uint8_t shuffleMaskC[] = GetShuffleMaskEpi8<{1,1,1,1}>();
constexpr std::uint8_t shuffleMaskE[] = GetShuffleMaskEpi8<{2,2,2,2}>();
constexpr std::uint8_t shuffleMaskG[] = GetShuffleMaskEpi8<{3,3,3,3}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskA = GetShuffleMaskEpi8<{{0,0,0,0}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskC = GetShuffleMaskEpi8<{{1,1,1,1}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskE = GetShuffleMaskEpi8<{{2,2,2,2}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskG = GetShuffleMaskEpi8<{{3,3,3,3}}>();
if constexpr(std::is_same_v<VectorType, __m128h>) {
VectorF16<Len, Packing> lenght = Length(A, C, E, G);
@ -638,16 +639,16 @@ namespace Crafter {
__m128h one = _mm_loadu_ph(oneArr);
__m128h fLenght = _mm_div_ph(one, lenght.v);
__m128i shuffleVecA = _mm_loadu_epi8(shuffleMaskA);
__m128i shuffleVecA = _mm_loadu_epi8(shuffleMaskA.data());
__m128h fLenghtA = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecA));
__m128i shuffleVecC = _mm_loadu_epi8(shuffleMaskC);
__m128i shuffleVecC = _mm_loadu_epi8(shuffleMaskC.data());
__m128h fLenghtC = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecC));
__m128i shuffleVecE = _mm_loadu_epi8(shuffleMaskE);
__m128i shuffleVecE = _mm_loadu_epi8(shuffleMaskE.data());
__m128h fLenghtE = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecE));
__m128i shuffleVecG = _mm_loadu_epi8(shuffleMaskG);
__m128i shuffleVecG = _mm_loadu_epi8(shuffleMaskG.data());
__m128h fLenghtG = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecG));
return {
@ -662,16 +663,16 @@ namespace Crafter {
__m256h one = _mm256_loadu_ph(oneArr);
__m256h fLenght = _mm256_div_ph(one, lenght.v);
__m256i shuffleVecA = _mm256_loadu_epi8(shuffleMaskA);
__m256i shuffleVecA = _mm256_loadu_epi8(shuffleMaskA.data());
__m256h fLenghtA = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecA));
__m256i shuffleVecC = _mm256_loadu_epi8(shuffleMaskC);
__m256i shuffleVecC = _mm256_loadu_epi8(shuffleMaskC.data());
__m256h fLenghtC = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecC));
__m256i shuffleVecE = _mm256_loadu_epi8(shuffleMaskE);
__m256i shuffleVecE = _mm256_loadu_epi8(shuffleMaskE.data());
__m256h fLenghtE = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecE));
__m256i shuffleVecG = _mm256_loadu_epi8(shuffleMaskG);
__m256i shuffleVecG = _mm256_loadu_epi8(shuffleMaskG.data());
__m256h fLenghtG = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecG));
return {
@ -686,16 +687,16 @@ namespace Crafter {
__m512h one = _mm512_loadu_ph(oneArr);
__m512h fLenght = _mm512_div_ph(one, lenght.v);
__m512i shuffleVecA = _mm512_loadu_epi8(shuffleMaskA);
__m512i shuffleVecA = _mm512_loadu_epi8(shuffleMaskA.data());
__m512h fLenghtA = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecA));
__m512i shuffleVecC = _mm512_loadu_epi8(shuffleMaskC);
__m512i shuffleVecC = _mm512_loadu_epi8(shuffleMaskC.data());
__m512h fLenghtC = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecC));
__m512i shuffleVecE = _mm512_loadu_epi8(shuffleMaskE);
__m512i shuffleVecE = _mm512_loadu_epi8(shuffleMaskE.data());
__m512h fLenghtE = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecE));
__m512i shuffleVecG = _mm512_loadu_epi8(shuffleMaskG);
__m512i shuffleVecG = _mm512_loadu_epi8(shuffleMaskG.data());
__m512h fLenghtG = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecG));
return {
@ -711,8 +712,8 @@ namespace Crafter {
VectorF16<Len, Packing> A,
VectorF16<Len, Packing> E
) requires(Len == 2) {
constexpr std::uint8_t shuffleMaskA[] = GetShuffleMaskEpi8<{0,0}>();
constexpr std::uint8_t shuffleMaskE[] = GetShuffleMaskEpi8<{1,1}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskA = GetShuffleMaskEpi8<{{0,0}}>();
constexpr std::array<std::uint8_t, Alignment*2> shuffleMaskE = GetShuffleMaskEpi8<{{1,1}}>();
if constexpr(std::is_same_v<VectorType, __m128h>) {
VectorF16<Len, Packing> lenght = Length(A, E);
@ -720,10 +721,10 @@ namespace Crafter {
__m128h one = _mm_loadu_ph(oneArr);
__m128h fLenght = _mm_div_ph(one, lenght.v);
__m128i shuffleVecA = _mm_loadu_epi8(shuffleMaskA);
__m128i shuffleVecA = _mm_loadu_epi8(shuffleMaskA.data());
__m128h fLenghtA = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecA));
__m128i shuffleVecE = _mm_loadu_epi8(shuffleMaskE);
__m128i shuffleVecE = _mm_loadu_epi8(shuffleMaskE.data());
__m128h fLenghtE = _mm_castsi128_ph(_mm_shuffle_epi8(_mm_castph_si128(fLenght), shuffleVecE));
return {
@ -736,10 +737,10 @@ namespace Crafter {
__m256h one = _mm256_loadu_ph(oneArr);
__m256h fLenght = _mm256_div_ph(one, lenght.v);
__m256i shuffleVecA = _mm256_loadu_epi8(shuffleMaskA);
__m256i shuffleVecA = _mm256_loadu_epi8(shuffleMaskA.data());
__m256h fLenghtA = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecA));
__m256i shuffleVecE = _mm256_loadu_epi8(shuffleMaskE);
__m256i shuffleVecE = _mm256_loadu_epi8(shuffleMaskE.data());
__m256h fLenghtE = _mm256_castsi256_ph(_mm256_shuffle_epi8(_mm256_castph_si256(fLenght), shuffleVecE));
return {
@ -752,10 +753,10 @@ namespace Crafter {
__m512h one = _mm512_loadu_ph(oneArr);
__m512h fLenght = _mm512_div_ph(one, lenght.v);
__m512i shuffleVecA = _mm512_loadu_epi8(shuffleMaskA);
__m512i shuffleVecA = _mm512_loadu_epi8(shuffleMaskA.data());
__m512h fLenghtA = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecA));
__m512i shuffleVecE = _mm512_loadu_epi8(shuffleMaskE);
__m512i shuffleVecE = _mm512_loadu_epi8(shuffleMaskE.data());
__m512h fLenghtE = _mm512_castsi512_ph(_mm512_shuffle_epi8(_mm512_castph_si512(fLenght), shuffleVecE));
return {
@ -1158,19 +1159,19 @@ namespace Crafter {
}
}
template <std::array<std::uint8_t, Len> ShuffleValues>
template <std::array<bool, Len> ShuffleValues>
constexpr static VectorF16<Len, Packing> Blend(VectorF16<Len, Packing> a, VectorF16<Len, Packing> b) {
if constexpr(std::is_same_v<VectorType, __m128h>) {
return _mm_castsi128_ph(_mm_blend_epi16(GetBlendMaskEpi16<ShuffleValues>(), _mm_castph_si128(a.v), _mm_castph_si128(b)));
return _mm_castsi128_ph(_mm_blend_epi16(_mm_castph_si128(a.v), _mm_castph_si128(b.v), GetBlendMaskEpi16<ShuffleValues>()));
} else if constexpr(std::is_same_v<VectorType, __m256h>) {
#ifndef __AVX512BW__
#ifndef __AVX512VL__
static_assert(false, "No __AVX512BW__ and __AVX512VL__ support");
#endif
#endif
return _mm256_castsi256_ph(_mm256_mask_blend_epi16(_mm256_castph_si256(a.v), _mm256_castph_si256(b), GetBlendMaskEpi16<ShuffleValues>()));
return _mm256_castsi256_ph(_mm256_mask_blend_epi16(_mm256_castph_si256(a.v), _mm256_castph_si256(b.v), GetBlendMaskEpi16<ShuffleValues>()));
} else {
return _mm512_castsi512_ph(_mm512_blend_epi16(GetBlendMaskEpi16<ShuffleValues>(), _mm512_castph_si512(a.v), _mm512_castph_si512(b)));
return _mm512_castsi512_ph(_mm512_blend_epi16(GetBlendMaskEpi16<ShuffleValues>(), _mm512_castph_si512(a.v), _mm512_castph_si512(b.v)));
}
}
@ -1189,34 +1190,36 @@ namespace Crafter {
}
constexpr static VectorF16<4, Packing> QuanternionFromEuler(VectorF16<3, Packing> EulerHalf) requires(Len == 4) {
VectorF16<4, Packing> sin = EulerHalf.Sin();
VectorF16<4, Packing> cos = EulerHalf.Cos();
std::tuple<VectorF16<3, Packing>, VectorF16<3, Packing>> sinCos = EulerHalf.SinCos();
VectorF16<4, Packing> sin = std::get<0>(sinCos);
VectorF16<4, Packing> cos = std::get<1>(sinCos);
VectorF16<4, Packing> row1 = cos.template Shuffle<{{0,0,0,0}}>();
VectorF16<4, Packing> row1 = cos.template Shuffle<{{0,0,0,0}}>();
row1 = Blend<{{0,1,1,1}}>(sin, row1);
VectorF16<4, Packing> row2 = cos.template Shuffle<{{1,1,1,1}}>();
row2 = Blend<{{1,0,1,1}}>(sin, row2);
row1 = row2;
row1 *= row2;
VectorF16<4, Packing> row3 = cos.template Shuffle<{{2,2,2,2}}>();
row3 = Blend<{{1,1,0,1}}>(sin, row3);
row1 *= row3;
VectorF16<4, Packing> row4 = sin.template Shuffle<{{0,0,0,0}}>();
row4 = Blend<{{1,0,0,0}}>(sin, row4);
row1 *= row4;
row1 = row1.template Negate<{{true,false,true}}>();
VectorF16<4, Packing> row4 = sin.template Shuffle<{{0,0,0,0}}>();
row4 = Blend<{{0,1,1,1}}>(cos, row4);
VectorF16<4, Packing> row5 = sin.template Shuffle<{{1,1,1,1}}>();
row5 = Blend<{{0,1,0,0}}>(sin, row5);
row5 = Blend<{{1,0,1,1}}>(cos, row5);
row4 *= row5;
VectorF16<4, Packing> row6 = sin.template Shuffle<{{2,2,2,2}}>();
row6 = Blend<{{0,0,1,0}}>(sin, row6);
row6 = Blend<{{1,1,0,1}}>(cos, row6);
row6 = row6.template Negate<{{true,false,true,false}}>();
row1 = MulitplyAdd(row5, row6, row1);
row1 = MulitplyAdd(row4, row6, row1);
return row1;
}
@ -1323,7 +1326,7 @@ namespace Crafter {
return mask;
}
template <std::array<std::uint8_t, Len> ShuffleValues>
template <std::array<bool, Len> ShuffleValues>
static consteval std::uint16_t GetBlendMaskEpi16() requires (std::is_same_v<VectorType, __m256h>){
std::uint16_t mask = 0;
for (std::uint8_t i2 = 0; i2 < Packing; i2++) {
@ -1336,7 +1339,7 @@ namespace Crafter {
return mask;
}
template <std::array<std::uint8_t, Len> ShuffleValues>
template <std::array<bool, Len> ShuffleValues>
static consteval std::uint32_t GetBlendMaskEpi16() requires (std::is_same_v<VectorType, __m512h>){
std::uint32_t mask = 0;
for (std::uint8_t i2 = 0; i2 < Packing; i2++) {