react-native-executorch 0.10.6 → 0.10.7

package/third-party/common/phonemis/src/third-party/xsimd/arch/common/xsimd_common_details.hpp

@@ -0,0 +1,386 @@
+/***************************************************************************
+ * Copyright (c) Johan Mabille, Sylvain Corlay, Wolf Vollprecht and         *
+ * Martin Renou                                                             *
+ * Copyright (c) QuantStack                                                 *
+ * Copyright (c) Serge Guelton                                              *
+ *                                                                          *
+ * Distributed under the terms of the BSD 3-Clause License.                 *
+ *                                                                          *
+ * The full license is in the file LICENSE, distributed with this software. *
+ ****************************************************************************/
+
+#ifndef XSIMD_COMMON_DETAILS_HPP
+#define XSIMD_COMMON_DETAILS_HPP
+
+#include <complex>
+
+#include "../../math/xsimd_rem_pio2.hpp"
+#include "../../types/xsimd_common_arch.hpp"
+#include "../../types/xsimd_utils.hpp"
+#include "../xsimd_constants.hpp"
+
+namespace xsimd
+{
+    // Forward declaration. Should we put them in a separate file?
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> abs(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> abs(batch<std::complex<T>, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE bool any(batch_bool<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> atan2(batch<T, A> const& self, batch<T, A> const& other) noexcept;
+    template <class A, class T_out, class T_in>
+    XSIMD_INLINE batch<T_out, A> batch_cast(batch<T_in, A> const&, batch<T_out, A> const& out) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> bitofsign(batch<T, A> const& self) noexcept;
+    template <class T_out, class T_in, class A>
+    XSIMD_INLINE batch<T_out, A> bitwise_cast(batch<T_in, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> cos(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> cosh(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> exp(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> fma(batch<T, A> const& x, batch<T, A> const& y, batch<T, A> const& z) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> fms(batch<T, A> const& x, batch<T, A> const& y, batch<T, A> const& z) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> fmas(batch<T, A> const& x, batch<T, A> const& y, batch<T, A> const& z) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> frexp(const batch<T, A>& x, const batch<as_integer_t<T>, A>& e) noexcept;
+    template <class T, class A, uint64_t... Coefs>
+    XSIMD_INLINE batch<T, A> horner(const batch<T, A>& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> hypot(const batch<T, A>& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch_bool<T, A> is_even(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch_bool<T, A> is_flint(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch_bool<T, A> is_odd(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE typename batch<T, A>::batch_bool_type isinf(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE typename batch<T, A>::batch_bool_type isfinite(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE typename batch<T, A>::batch_bool_type isnan(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> ldexp(const batch<T, A>& x, const batch<as_integer_t<T>, A>& e) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> log(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> nearbyint(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<as_integer_t<T>, A> nearbyint_as_int(const batch<T, A>& x) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE T reduce_add(batch<T, A> const&) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE T reduce_mul(batch<T, A> const&) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> select(batch_bool<T, A> const&, batch<T, A> const&, batch<T, A> const&) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<std::complex<T>, A> select(batch_bool<T, A> const&, batch<std::complex<T>, A> const&, batch<std::complex<T>, A> const&) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> sign(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> signnz(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> sin(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> sinh(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE std::pair<batch<T, A>, batch<T, A>> sincos(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> sqrt(batch<T, A> const& self) noexcept;
+    template <class T, class A, class Vt, Vt... Values>
+    XSIMD_INLINE std::enable_if_t<std::is_arithmetic<T>::value, batch<T, A>>
+    swizzle(batch<T, A> const& x, batch_constant<Vt, A, Values...> mask) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> tan(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<as_float_t<T>, A> to_float(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<as_integer_t<T>, A> to_int(batch<T, A> const& self) noexcept;
+    template <class T, class A>
+    XSIMD_INLINE batch<T, A> trunc(batch<T, A> const& self) noexcept;
+
+    namespace kernel
+    {
+
+        namespace detail
+        {
+            // Prevent -ffast-math from reassociating floating-point
+            // arithmetic across this point.  The reason string
+            // documents *why* at each call site; unused at runtime.
+            //
+            // Zero-cost register constraints per target:
+            //   x86  "+x"  (XMM/YMM/ZMM, also scalar float/double)
+            //   ARM  "+w"  (V-reg / SVE Z-reg, also scalar float/double)
+            //   PPC  "+wa" (VS register, also scalar float/double)
+            //   RISC-V "+f" (F/D register, scalar float/double)
+            //   RISC-V RVV "+vr" (V register; GCC 15+ / Clang 20+)
+            //
+            // On unknown targets the "+m" fallback spills; it is
+            // only emitted when the compiler can actually reassociate.
+            template <class T>
+            XSIMD_INLINE void reassociation_barrier(T& x, const char*) noexcept
+            {
+#if XSIMD_REASSOCIATIVE_MATH && XSIMD_WITH_INLINE_ASM && !defined(__EMSCRIPTEN__)
+#if XSIMD_WITH_SSE2
+                __asm__ volatile("" : "+x"(x));
+#elif XSIMD_WITH_NEON || XSIMD_WITH_SVE
+                __asm__ volatile("" : "+w"(x));
+#elif XSIMD_WITH_VSX
+                __asm__ volatile("" : "+wa"(x));
+#else
+                __asm__ volatile("" : "+m"(x));
+#endif
+#else
+                (void)x;
+#endif
+            }
+
+            // RISC-V scalar float/double: use F/D registers instead of
+            // spilling through "+m".  These overloads also serve
+            // emulated batches on RISC-V via the std::array overload.
+#if XSIMD_REASSOCIATIVE_MATH && XSIMD_WITH_INLINE_ASM && defined(__riscv)
+            XSIMD_INLINE void reassociation_barrier(float& x, const char*) noexcept
+            {
+                __asm__ volatile("" : "+f"(x));
+            }
+            XSIMD_INLINE void reassociation_barrier(double& x, const char*) noexcept
+            {
+                __asm__ volatile("" : "+f"(x));
+            }
+#endif
+
+            template <class T, size_t N>
+            XSIMD_INLINE void reassociation_barrier(std::array<T, N>& arr, const char* reason) noexcept
+            {
+                for (auto& v : arr)
+                    reassociation_barrier(v, reason);
+            }
+
+            template <class T, class A>
+            XSIMD_INLINE void reassociation_barrier(batch<T, A>& b, const char* reason) noexcept
+            {
+#if XSIMD_REASSOCIATIVE_MATH && XSIMD_WITH_RVV && XSIMD_WITH_INLINE_ASM && ((__GNUC__ >= 15) || (__clang_major__ >= 20))
+                __asm__ volatile("" : "+vr"(b.data.value.value));
+                (void)reason;
+#else
+                reassociation_barrier(b.data, reason);
+#endif
+            }
+
+            template <class F, class A, class T, class... Batches>
+            XSIMD_INLINE batch<T, A> apply(F&& func, batch<T, A> const& self, batch<T, A> const& other) noexcept
+            {
+                constexpr std::size_t size = batch<T, A>::size;
+                alignas(A::alignment()) T self_buffer[size];
+                alignas(A::alignment()) T other_buffer[size];
+                self.store_aligned(&self_buffer[0]);
+                other.store_aligned(&other_buffer[0]);
+                for (std::size_t i = 0; i < size; ++i)
+                {
+                    self_buffer[i] = func(self_buffer[i], other_buffer[i]);
+                }
+                return batch<T, A>::load_aligned(self_buffer);
+            }
+
+            template <class U, class F, class A, class T>
+            XSIMD_INLINE batch<U, A> apply_transform(F&& func, batch<T, A> const& self) noexcept
+            {
+                static_assert(batch<T, A>::size == batch<U, A>::size,
+                              "Source and destination sizes must match");
+                constexpr std::size_t src_size = batch<T, A>::size;
+                constexpr std::size_t dest_size = batch<U, A>::size;
+                alignas(A::alignment()) T self_buffer[src_size];
+                alignas(A::alignment()) U other_buffer[dest_size];
+                self.store_aligned(&self_buffer[0]);
+                for (std::size_t i = 0; i < src_size; ++i)
+                {
+                    other_buffer[i] = func(self_buffer[i]);
+                }
+                return batch<U, A>::load_aligned(other_buffer);
+            }
+        }
+
+        // some common fast_cast conversion
+        namespace detail
+        {
+            template <class A>
+            XSIMD_INLINE batch<uint8_t, A> fast_cast(batch<int8_t, A> const& self, batch<uint8_t, A> const&, requires_arch<common>) noexcept
+            {
+                return bitwise_cast<uint8_t>(self);
+            }
+            template <class A>
+            XSIMD_INLINE batch<uint16_t, A> fast_cast(batch<int16_t, A> const& self, batch<uint16_t, A> const&, requires_arch<common>) noexcept
+            {
+                return bitwise_cast<uint16_t>(self);
+            }
+            template <class A>
+            XSIMD_INLINE batch<uint32_t, A> fast_cast(batch<int32_t, A> const& self, batch<uint32_t, A> const&, requires_arch<common>) noexcept
+            {
+                return bitwise_cast<uint32_t>(self);
+            }
+            template <class A>
+            XSIMD_INLINE batch<uint64_t, A> fast_cast(batch<int64_t, A> const& self, batch<uint64_t, A> const&, requires_arch<common>) noexcept
+            {
+                return bitwise_cast<uint64_t>(self);
+            }
+            template <class A>
+            XSIMD_INLINE batch<int8_t, A> fast_cast(batch<uint8_t, A> const& self, batch<int8_t, A> const&, requires_arch<common>) noexcept
+            {
+                return bitwise_cast<int8_t>(self);
+            }
+            template <class A>
+            XSIMD_INLINE batch<int16_t, A> fast_cast(batch<uint16_t, A> const& self, batch<int16_t, A> const&, requires_arch<common>) noexcept
+            {
+                return bitwise_cast<int16_t>(self);
+            }
+            template <class A>
+            XSIMD_INLINE batch<int32_t, A> fast_cast(batch<uint32_t, A> const& self, batch<int32_t, A> const&, requires_arch<common>) noexcept
+            {
+                return bitwise_cast<int32_t>(self);
+            }
+            template <class A>
+            XSIMD_INLINE batch<int64_t, A> fast_cast(batch<uint64_t, A> const& self, batch<int64_t, A> const&, requires_arch<common>) noexcept
+            {
+                return bitwise_cast<int64_t>(self);
+            }
+
+            // Provide a common uint32_t -> float cast only if we have a
+            // non-common int32_t -> float fast_cast
+            template <class A, class = decltype(fast_cast(std::declval<batch<int32_t, A> const&>(), std::declval<batch<float, A> const&>(), A {}))>
+            XSIMD_INLINE batch<float, A> fast_cast(batch<uint32_t, A> const& v, batch<float, A> const&, requires_arch<common>) noexcept
+            {
+                // see https://stackoverflow.com/questions/34066228/how-to-perform-uint32-float-conversion-with-sse
+                batch<uint32_t, A> msk_lo(0xFFFF);
+                batch<float, A> cnst65536f(65536.0f);
+
+                auto v_lo = batch_cast<int32_t>(v & msk_lo); /* extract the 16 lowest significant bits of self                             */
+                auto v_hi = batch_cast<int32_t>(v >> 16); /* 16 most significant bits of v                                                 */
+                auto v_lo_flt = batch_cast<float>(v_lo); /* No rounding                                                                */
+                auto v_hi_flt = batch_cast<float>(v_hi); /* No rounding                                                                */
+                v_hi_flt = cnst65536f * v_hi_flt; /* No rounding                                                            */
+                return v_hi_flt + v_lo_flt; /* Rounding may occur here, mul and add may fuse to fma for haswell and newer   */
+            }
+
+            // Provide a common float -> uint32_t cast only if we have a
+            // non-common float -> int32_t fast_cast
+            template <class A, class = decltype(fast_cast(std::declval<batch<float, A> const&>(), std::declval<batch<int32_t, A> const&>(), A {}))>
+            XSIMD_INLINE batch<uint32_t, A> fast_cast(batch<float, A> const& v, batch<uint32_t, A> const&, requires_arch<common>) noexcept
+            {
+                auto is_large = v >= batch<float, A>(1u << 31);
+                auto small_v = bitwise_cast<float>(batch_cast<int32_t>(v));
+                auto large_v = bitwise_cast<float>(
+                    batch_cast<int32_t>(v - batch<float, A>(1u << 31))
+                    ^ batch<int32_t, A>(1u << 31));
+                return bitwise_cast<uint32_t>(select(is_large, large_v, small_v));
+            }
+        }
+
+        namespace detail
+        {
+            // Generic conversion handling machinery. Each architecture must define
+            // conversion function when such conversions exits in the form of
+            // intrinsic. Then we use that information to automatically decide whether
+            // to use scalar or vector conversion when doing load / store / batch_cast
+            struct with_fast_conversion
+            {
+            };
+            struct with_slow_conversion
+            {
+            };
+
+            template <class A, class From, class To, class = void>
+            struct conversion_type_impl
+            {
+                using type = with_slow_conversion;
+            };
+
+            using xsimd::detail::void_t;
+
+            template <class A, class From, class To>
+            struct conversion_type_impl<A, From, To,
+                                        void_t<decltype(fast_cast(std::declval<const batch<From, A>&>(),
+                                                                  std::declval<const batch<To, A>&>(),
+                                                                  std::declval<const A&>()))>>
+            {
+                using type = with_fast_conversion;
+            };
+
+            template <class A, class From, class To>
+            using conversion_type = typename conversion_type_impl<A, From, To>::type;
+        }
+
+        namespace detail
+        {
+            /* origin: boost/simdfunction/horn.hpp*/
+            /*
+             * ====================================================
+             * copyright 2016 NumScale SAS
+             *
+             * Distributed under the Boost Software License, Version 1.0.
+             * (See copy at http://boost.org/LICENSE_1_0.txt)
+             * ====================================================
+             */
+            template <class B, uint64_t c>
+            XSIMD_INLINE B coef() noexcept
+            {
+                using value_type = typename B::value_type;
+                return B(bit_cast<value_type>(as_unsigned_integer_t<value_type>(c)));
+            }
+            template <class B>
+            XSIMD_INLINE B horner(const B&) noexcept
+            {
+                return B(typename B::value_type(0.));
+            }
+
+            template <class B, uint64_t c0>
+            XSIMD_INLINE B horner(const B&) noexcept
+            {
+                return coef<B, c0>();
+            }
+
+            template <class B, uint64_t c0, uint64_t c1, uint64_t... args>
+            XSIMD_INLINE B horner(const B& self) noexcept
+            {
+                return fma(self, horner<B, c1, args...>(self), coef<B, c0>());
+            }
+
+            /* origin: boost/simdfunction/horn1.hpp*/
+            /*
+             * ====================================================
+             * copyright 2016 NumScale SAS
+             *
+             * Distributed under the Boost Software License, Version 1.0.
+             * (See copy at http://boost.org/LICENSE_1_0.txt)
+             * ====================================================
+             */
+            template <class B>
+            XSIMD_INLINE B horner1(const B&) noexcept
+            {
+                return B(1.);
+            }
+
+            template <class B, uint64_t c0>
+            XSIMD_INLINE B horner1(const B& x) noexcept
+            {
+                return x + detail::coef<B, c0>();
+            }
+
+            template <class B, uint64_t c0, uint64_t c1, uint64_t... args>
+            XSIMD_INLINE B horner1(const B& x) noexcept
+            {
+                return fma(x, horner1<B, c1, args...>(x), detail::coef<B, c0>());
+            }
+        }
+
+    }
+
+}
+
+#endif

package/third-party/common/phonemis/src/third-party/xsimd/arch/common/xsimd_common_logical.hpp

@@ -0,0 +1,240 @@
+/***************************************************************************
+ * Copyright (c) Johan Mabille, Sylvain Corlay, Wolf Vollprecht and         *
+ * Martin Renou                                                             *
+ * Copyright (c) QuantStack                                                 *
+ * Copyright (c) Serge Guelton                                              *
+ *                                                                          *
+ * Distributed under the terms of the BSD 3-Clause License.                 *
+ *                                                                          *
+ * The full license is in the file LICENSE, distributed with this software. *
+ ****************************************************************************/
+
+#ifndef XSIMD_COMMON_LOGICAL_HPP
+#define XSIMD_COMMON_LOGICAL_HPP
+
+#include "./xsimd_common_bit.hpp"
+#include "./xsimd_common_details.hpp"
+
+#include <climits>
+
+namespace xsimd
+{
+
+    namespace kernel
+    {
+
+        using namespace types;
+
+        // count
+        template <class A, class T>
+        XSIMD_INLINE size_t count(batch_bool<T, A> const& self, requires_arch<common>) noexcept
+        {
+            return xsimd::detail::popcount(self.mask());
+        }
+
+        template <class A, class T>
+        XSIMD_INLINE size_t countl_zero(batch_bool<T, A> const& self, requires_arch<common>) noexcept
+        {
+            constexpr size_t unused_bits = 64 - batch_bool<T, A>::size;
+            constexpr uint64_t lower_mask = batch_bool<T, A>::size < 64 ? ((uint64_t)1 << (batch_bool<T, A>::size % 64)) - 1 : (uint64_t)-1;
+            return xsimd::detail::countl_zero(self.mask() & lower_mask) - unused_bits;
+        }
+
+        template <class A, class T>
+        XSIMD_INLINE size_t countl_one(batch_bool<T, A> const& self, requires_arch<common>) noexcept
+        {
+            constexpr size_t unused_bits = 64 - batch_bool<T, A>::size;
+            constexpr uint64_t upper_mask = batch_bool<T, A>::size < 64 ? ~(((uint64_t)1 << (batch_bool<T, A>::size % 64)) - 1) : (uint64_t)0;
+            return xsimd::detail::countl_one(self.mask() | upper_mask) - unused_bits;
+        }
+
+        template <class A, class T>
+        XSIMD_INLINE size_t countr_zero(batch_bool<T, A> const& self, requires_arch<common>) noexcept
+        {
+            constexpr uint64_t stop = batch_bool<T, A>::size < 64 ? (uint64_t)1 << (batch_bool<T, A>::size % 64) : 0;
+            return xsimd::detail::countr_zero(self.mask() | stop);
+        }
+
+        template <class A, class T>
+        XSIMD_INLINE size_t countr_one(batch_bool<T, A> const& self, requires_arch<common>) noexcept
+        {
+            constexpr uint64_t stop = batch_bool<T, A>::size < 64 ? ~((uint64_t)1 << (batch_bool<T, A>::size % 64)) : (uint64_t)-1;
+            return xsimd::detail::countr_one(self.mask() & stop);
+        }
+
+        // from  mask
+        template <class A, class T>
+        XSIMD_INLINE batch_bool<T, A> from_mask(batch_bool<T, A> const&, uint64_t mask, requires_arch<common>) noexcept
+        {
+            alignas(A::alignment()) bool buffer[batch_bool<T, A>::size];
+            // This is inefficient and should never be called.
+            for (size_t i = 0; i < batch_bool<T, A>::size; ++i)
+                buffer[i] = mask & (1ull << i);
+            return batch_bool<T, A>::load_aligned(buffer);
+        }
+
+        // ge
+        template <class A, class T>
+        XSIMD_INLINE batch_bool<T, A> ge(batch<T, A> const& self, batch<T, A> const& other, requires_arch<common>) noexcept
+        {
+            return other <= self;
+        }
+
+        // gt
+        template <class A, class T>
+        XSIMD_INLINE batch_bool<T, A> gt(batch<T, A> const& self, batch<T, A> const& other, requires_arch<common>) noexcept
+        {
+            return other < self;
+        }
+
+        // is_even
+        template <class A, class T>
+        XSIMD_INLINE batch_bool<T, A> is_even(batch<T, A> const& self, requires_arch<common>) noexcept
+        {
+            return is_flint(self * T(0.5));
+        }
+
+        // is_flint
+        template <class A, class T>
+        XSIMD_INLINE batch_bool<T, A> is_flint(batch<T, A> const& self, requires_arch<common>) noexcept
+        {
+            auto frac = select(isnan(self - self), constants::nan<batch<T, A>>(), self - trunc(self));
+            return frac == T(0.);
+        }
+
+        // is_odd
+        template <class A, class T>
+        XSIMD_INLINE batch_bool<T, A> is_odd(batch<T, A> const& self, requires_arch<common>) noexcept
+        {
+            return is_even(self - T(1.));
+        }
+
+        // isinf
+        template <class A, class T, class = std::enable_if_t<std::is_integral<T>::value>>
+        XSIMD_INLINE batch_bool<T, A> isinf(batch<T, A> const&, requires_arch<common>) noexcept
+        {
+            return batch_bool<T, A>(false);
+        }
+        template <class A>
+        XSIMD_INLINE batch_bool<float, A> isinf(batch<float, A> const& self, requires_arch<common>) noexcept
+        {
+#ifdef __FAST_MATH__
+            (void)self;
+            return { false };
+#else
+            return abs(self) == std::numeric_limits<float>::infinity();
+#endif
+        }
+        template <class A>
+        XSIMD_INLINE batch_bool<double, A> isinf(batch<double, A> const& self, requires_arch<common>) noexcept
+        {
+#ifdef __FAST_MATH__
+            (void)self;
+            return { false };
+#else
+            return abs(self) == std::numeric_limits<double>::infinity();
+#endif
+        }
+
+        // isfinite
+        template <class A, class T, class = std::enable_if_t<std::is_integral<T>::value>>
+        XSIMD_INLINE batch_bool<T, A> isfinite(batch<T, A> const&, requires_arch<common>) noexcept
+        {
+            return batch_bool<T, A>(true);
+        }
+        template <class A>
+        XSIMD_INLINE batch_bool<float, A> isfinite(batch<float, A> const& self, requires_arch<common>) noexcept
+        {
+            return (self - self) == 0.f;
+        }
+        template <class A>
+        XSIMD_INLINE batch_bool<double, A> isfinite(batch<double, A> const& self, requires_arch<common>) noexcept
+        {
+            return (self - self) == 0.;
+        }
+
+        // isnan
+        template <class A, class T, class = std::enable_if_t<std::is_integral<T>::value>>
+        XSIMD_INLINE batch_bool<T, A> isnan(batch<T, A> const&, requires_arch<common>) noexcept
+        {
+            return batch_bool<T, A>(false);
+        }
+
+        // le
+        template <class A, class T, class = std::enable_if_t<std::is_integral<T>::value>>
+        XSIMD_INLINE batch_bool<T, A> le(batch<T, A> const& self, batch<T, A> const& other, requires_arch<common>) noexcept
+        {
+            return (self < other) || (self == other);
+        }
+
+        // neq
+        template <class A, class T>
+        XSIMD_INLINE batch_bool<T, A> neq(batch<T, A> const& self, batch<T, A> const& other, requires_arch<common>) noexcept
+        {
+            return !(other == self);
+        }
+
+        // logical_and
+        template <class A, class T>
+        XSIMD_INLINE batch<T, A> logical_and(batch<T, A> const& self, batch<T, A> const& other, requires_arch<common>) noexcept
+        {
+            return detail::apply([](T x, T y) noexcept
+                                 { return x && y; },
+                                 self, other);
+        }
+
+        // logical_or
+        template <class A, class T>
+        XSIMD_INLINE batch<T, A> logical_or(batch<T, A> const& self, batch<T, A> const& other, requires_arch<common>) noexcept
+        {
+            return detail::apply([](T x, T y) noexcept
+                                 { return x || y; },
+                                 self, other);
+        }
+
+        // mask
+        template <class A, class T>
+        XSIMD_INLINE uint64_t mask(batch_bool<T, A> const& self, requires_arch<common>) noexcept
+        {
+            alignas(A::alignment()) bool buffer[batch_bool<T, A>::size];
+            self.store_aligned(buffer);
+            // This is inefficient and should never be called.
+            uint64_t res = 0;
+            for (size_t i = 0; i < batch_bool<T, A>::size; ++i)
+                if (buffer[i])
+                    res |= 1ul << i;
+            return res;
+        }
+
+        // select
+        namespace detail
+        {
+            template <typename T, typename A>
+            using is_batch_bool_register_same = std::is_same<typename batch_bool<T, A>::register_type, typename batch<T, A>::register_type>;
+        }
+
+        template <class A, class T, std::enable_if_t<detail::is_batch_bool_register_same<T, A>::value, int> = 3>
+        XSIMD_INLINE batch_bool<T, A> select(batch_bool<T, A> const& cond, batch_bool<T, A> const& true_br, batch_bool<T, A> const& false_br, requires_arch<common>)
+        {
+            using register_type = typename batch_bool<T, A>::register_type;
+            // Do not cast, but rather reinterpret the masks as batches.
+            const auto true_v = batch<T, A> { static_cast<register_type>(true_br) };
+            const auto false_v = batch<T, A> { static_cast<register_type>(false_br) };
+            return batch_bool<T, A> { select(cond, true_v, false_v) };
+        }
+
+        template <class A, class T, std::enable_if_t<!detail::is_batch_bool_register_same<T, A>::value, int> = 3>
+        XSIMD_INLINE batch_bool<T, A> select(batch_bool<T, A> const& cond, batch_bool<T, A> const& true_br, batch_bool<T, A> const& false_br, requires_arch<common>)
+        {
+            return (true_br & cond) | (bitwise_andnot(false_br, cond));
+        }
+
+        template <class A, class T, bool... Values>
+        XSIMD_INLINE batch_bool<T, A> select(batch_bool_constant<T, A, Values...> const& cond, batch_bool<T, A> const& true_br, batch_bool<T, A> const& false_br, requires_arch<common>)
+        {
+            return (true_br & cond) | (false_br & ~cond);
+        }
+    }
+}
+
+#endif