| arm_neon-inl.h |
|
347066 |
| arm_sve-inl.h |
|
224026 |
| emu128-inl.h |
tag |
88528 |
| generic_ops-inl.h |
.h.
#if HWY_IDE && !defined(HWY_HIGHWAY_INCLUDED)
#include "hwy/detect_targets.h"
#include "hwy/ops/emu128-inl.h"
#endif // HWY_IDE
// Relies on the external include guard in highway.h.
HWY_BEFORE_NAMESPACE();
namespace hwy {
namespace HWY_NAMESPACE {
// The lane type of a vector type, e.g. float for Vec<ScalableTag<float>>.
template <class V>
using LaneType = decltype(GetLane(V()));
// Vector type, e.g. Vec128<float> for CappedTag<float, 4>. Useful as the return
// type of functions that do not take a vector argument, or as an argument type
// if the function only has a template argument for D, or for explicit type
// names instead of auto. This may be a built-in type.
template <class D>
using Vec = decltype(Zero(D()));
// Mask type. Useful as the return type of functions that do not take a mask
// argument, or as an argument type if the function only has a template argument
// for D, or for explicit type names instead of auto.
template <class D>
using Mask = decltype(MaskFromVec(Zero(D())));
// Returns the closest value to v within [lo, hi].
template <class V>
HWY_API V Clamp(const V v, const V lo, const V hi) {
return Min(Max(lo, v), hi);
}
// CombineShiftRightBytes (and -Lanes) are not available for the scalar target,
// and RVV has its own implementation of -Lanes.
#if HWY_TARGET != HWY_SCALAR && HWY_TARGET != HWY_RVV
template <size_t kLanes, class D>
HWY_API VFromD<D> CombineShiftRightLanes(D d, VFromD<D> hi, VFromD<D> lo) {
constexpr size_t kBytes = kLanes * sizeof(TFromD<D>);
static_assert(kBytes < 16, "Shift count is per-block");
return CombineShiftRightBytes<kBytes>(d, hi, lo);
}
#endif
// Returns lanes with the most significant bit set and all other bits zero.
template <class D>
HWY_API Vec<D> SignBit(D d) {
const RebindToUnsigned<decltype(d)> du;
return BitCast(d, Set(du, SignMask<TFromD<D>>()));
}
// Returns quiet NaN.
template <class D>
HWY_API Vec<D> NaN(D d) {
const RebindToSigned<D> di;
// LimitsMax sets all exponent and mantissa bits to 1. The exponent plus
// mantissa MSB (to indicate quiet) would be sufficient.
return BitCast(d, Set(di, LimitsMax<TFromD<decltype(di)>>()));
}
// Returns positive infinity.
template <class D>
HWY_API Vec<D> Inf(D d) {
const RebindToUnsigned<D> du;
using T = TFromD<D>;
using TU = TFromD<decltype(du)>;
const TU max_x2 = static_cast<TU>(MaxExponentTimes2<T>());
return BitCast(d, Set(du, max_x2 >> 1));
}
// ------------------------------ ZeroExtendResizeBitCast
// The implementation of detail::ZeroExtendResizeBitCast for the HWY_EMU128
// target is in emu128-inl.h, and the implementation of
// detail::ZeroExtendResizeBitCast for the HWY_SCALAR target is in scalar-inl.h
#if HWY_TARGET != HWY_EMU128 && HWY_TARGET != HWY_SCALAR
namespace detail {
#if HWY_HAVE_SCALABLE
template <size_t kFromVectSize, size_t kToVectSize, class DTo, class DFrom>
HWY_INLINE VFromD<DTo> ZeroExtendResizeBitCast(
hwy::SizeTag<kFromVectSize> /* from_size_tag |
256294 |
| ppc_vsx-inl.h |
|
260547 |
| rvv-inl.h |
tag |
242550 |
| scalar-inl.h |
tag |
64716 |
| set_macros-inl.h |
|
17850 |
| shared-inl.h |
|
27105 |
| tuple-inl.h |
that lack native tuple types, after
// they define VFromD and before they use the tuples e.g. for LoadInterleaved2.
// Assumes we are already in the HWY_NAMESPACE and under an include guard.
// If viewing this header standalone, define VFromD to avoid IDE warnings.
// This is normally set by set_macros-inl.h before this header is included.
#if !defined(HWY_NAMESPACE)
#include "hwy/base.h"
template <class D>
using VFromD = int;
#endif
// On SVE, Vec2..4 are aliases to built-in types.
template <class D>
struct Vec2 {
VFromD<D> v0;
VFromD<D> v1;
};
template <class D>
struct Vec3 {
VFromD<D> v0;
VFromD<D> v1;
VFromD<D> v2;
};
template <class D>
struct Vec4 {
VFromD<D> v0;
VFromD<D> v1;
VFromD<D> v2;
VFromD<D> v3;
};
// D arg is unused but allows deducing D.
template <class D>
HWY_API Vec2<D> Create2(D /* tag |
3397 |
| wasm_128-inl.h |
tag |
230868 |
| wasm_256-inl.h |
|
79946 |
| x86_128-inl.h |
|
457535 |
| x86_256-inl.h |
|
309828 |
| x86_512-inl.h |
|
288498 |