WIP

Author: serge-sans-paille

include/xsimd/arch/xsimd_neon.hpp

@@ -991,10 +991,19 @@ namespace xsimd
             return dispatcher.apply(register_type(lhs), register_type(rhs));
         }
 
-        template <class A, class T, detail::enable_sized_integral_t<T, 8> = 0>
+        template <class A, class T, detail::enable_sized_signed_t<T, 8> = 0>
+        XSIMD_INLINE batch_bool<T, A> lt(batch<T, A> const& lhs, batch<T, A> const& rhs, requires_arch<neon>) noexcept
+        {
+            using register_type = typename batch<T, A>::register_type;
+            return batch_bool<T, A>(vshrq_n_s64(vqsubq_s64(register_type(rhs), register_type(lhs)), 63));
+        }
+
+        template <class A, class T, detail::enable_sized_unsigned_t<T, 8> = 0>
         XSIMD_INLINE batch_bool<T, A> lt(batch<T, A> const& lhs, batch<T, A> const& rhs, requires_arch<neon>) noexcept
         {
-            return batch_bool<T, A>({ lhs.get(0) < rhs.get(0), lhs.get(1) < rhs.get(1) });
+            using register_type = typename batch<T, A>::register_type;
+            register_type acc = { 0x8FFFFFFFFFFFFFFFull, 0x8FFFFFFFFFFFFFFFull };
+            return batch_bool<T, A>(vreinterpretq_u64_s64(detail::bitwise_not_s64(vshrq_n_s64(vreinterpretq_s64_u64(vqaddq_u64(vqsubq_u64(register_type(lhs), register_type(rhs)), acc)), 63))));
         }
 
         /******
@@ -1035,6 +1044,23 @@ namespace xsimd
             {
                 return vreinterpretq_u64_u32(vmvnq_u32(vreinterpretq_u32_u64(arg)));
             }
+
+            XSIMD_INLINE float32x4_t bitwise_not_f32(float32x4_t arg) noexcept
+            {
+                return vreinterpretq_f32_u32(vmvnq_u32(vreinterpretq_u32_f32(arg)));
+            }
+
+            template <class V>
+            XSIMD_INLINE V bitwise_not_neon(V const& arg) noexcept
+            {
+                const neon_dispatcher::unary dispatcher = {
+                    std::make_tuple(wrap::vmvnq_u8, wrap::vmvnq_s8, wrap::vmvnq_u16, wrap::vmvnq_s16,
+                                    wrap::vmvnq_u32, wrap::vmvnq_s32,
+                                    bitwise_not_u64, bitwise_not_s64,
+                                    bitwise_not_f32)
+                };
+                return dispatcher.apply(arg);
+            }
         }
 
         WRAP_BINARY_INT_EXCLUDING_64(vcgtq, detail::comp_return_type)
@@ -1055,14 +1081,15 @@ namespace xsimd
         XSIMD_INLINE batch_bool<T, A> gt(batch<T, A> const& lhs, batch<T, A> const& rhs, requires_arch<neon>) noexcept
         {
             using register_type = typename batch<T, A>::register_type;
-            return vshrq_n_s64(vqsubq_s64(register_type(lhs), register_type(rhs)), 63);
+            return batch_bool<T, A>(vshrq_n_s64(vqsubq_s64(register_type(lhs), register_type(rhs)), 63));
         }
 
         template <class A, class T, detail::enable_sized_unsigned_t<T, 8> = 0>
         XSIMD_INLINE batch_bool<T, A> gt(batch<T, A> const& lhs, batch<T, A> const& rhs, requires_arch<neon>) noexcept
         {
             using register_type = typename batch<T, A>::register_type;
-            return detail::bitwise_not_s64(vshrq_n_s64(vreinterpretq_s64_u64(vqsubq_u64(register_type(rhs), register_type(lhs))), 63));
+            register_type acc = { 0x8FFFFFFFFFFFFFFFull, 0x8FFFFFFFFFFFFFFFull };
+            return batch_bool<T, A>(vreinterpretq_u64_s64(detail::bitwise_not_s64(vshrq_n_s64(vreinterpretq_s64_u64(vqaddq_u64(vqsubq_u64(register_type(rhs), register_type(lhs)), acc)), 63))));
         }
 
         /******
@@ -1086,7 +1113,7 @@ namespace xsimd
         template <class A, class T, detail::enable_sized_integral_t<T, 8> = 0>
         XSIMD_INLINE batch_bool<T, A> ge(batch<T, A> const& lhs, batch<T, A> const& rhs, requires_arch<neon>) noexcept
         {
-            return batch_bool<T, A>({ lhs.get(0) >= rhs.get(0), lhs.get(1) >= rhs.get(1) });
+            return detail::bitwise_not_neon(lt(rhs, lt, A {}));
         }
 
         /*******************
@@ -1236,26 +1263,6 @@ namespace xsimd
 
         WRAP_UNARY_INT_EXCLUDING_64(vmvnq)
 
-        namespace detail
-        {
-            XSIMD_INLINE float32x4_t bitwise_not_f32(float32x4_t arg) noexcept
-            {
-                return vreinterpretq_f32_u32(vmvnq_u32(vreinterpretq_u32_f32(arg)));
-            }
-
-            template <class V>
-            XSIMD_INLINE V bitwise_not_neon(V const& arg) noexcept
-            {
-                const neon_dispatcher::unary dispatcher = {
-                    std::make_tuple(wrap::vmvnq_u8, wrap::vmvnq_s8, wrap::vmvnq_u16, wrap::vmvnq_s16,
-                                    wrap::vmvnq_u32, wrap::vmvnq_s32,
-                                    bitwise_not_u64, bitwise_not_s64,
-                                    bitwise_not_f32)
-                };
-                return dispatcher.apply(arg);
-            }
-        }
-
         template <class A, class T, detail::enable_neon_type_t<T> = 0>
         XSIMD_INLINE batch<T, A> bitwise_not(batch<T, A> const& arg, requires_arch<neon>) noexcept
         {