Specialize avx512bw bool loader for aligned access

Author: serge-sans-paille

include/xsimd/arch/generic/xsimd_generic_memory.hpp

@@ -262,7 +262,7 @@ namespace xsimd
 
         // load
         template <class A, class T>
-        XSIMD_INLINE batch_bool<T, A> load(bool const* mem, batch_bool<T, A>, requires_arch<generic>) noexcept
+        XSIMD_INLINE batch_bool<T, A> load_unaligned(bool const* mem, batch_bool<T, A>, requires_arch<generic>) noexcept
         {
             using batch_type = batch<T, A>;
             batch_type ref(0);
@@ -273,6 +273,12 @@ namespace xsimd
             return ref != batch_type::load_aligned(&buffer[0]);
         }
 
+        template <class A, class T>
+        XSIMD_INLINE batch_bool<T, A> load_aligned(bool const* mem, batch_bool<T, A> b, requires_arch<generic>) noexcept
+        {
+            return load_unaligned(mem, b, A {});
+        }
+
         // load_aligned
         namespace detail
         {

include/xsimd/arch/xsimd_avx512bw.hpp

@@ -316,19 +316,18 @@ namespace xsimd
         }
 
         // load
-        template <class A, class T>
-        XSIMD_INLINE batch_bool<T, A> load(bool const* mem, batch_bool<T, A>, requires_arch<avx512bw>) noexcept
+        template <class A, class T, class = typename std::enable_if<batch_bool<T, A>::size == 64, void>::type>
+        XSIMD_INLINE batch_bool<T, A> load_unaligned(bool const* mem, batch_bool<T, A>, requires_arch<avx512bw>) noexcept
         {
-            using register_type = typename batch_bool<T, A>::register_type;
-            XSIMD_IF_CONSTEXPR(batch_bool<T, A>::size == 64)
-            {
-                __m512i bool_val = _mm512_loadu_si512((__m512i const*)mem);
-                return (register_type)_mm512_cmpgt_epu8_mask(bool_val, _mm512_setzero_si512());
-            }
-            else
-            {
-                return load(mem, batch_bool<T, A>(), avx512dq {});
-            }
+            __m512i bool_val = _mm512_loadu_si512((__m512i const*)mem);
+            return _mm512_cmpgt_epu8_mask(bool_val, _mm512_setzero_si512());
+        }
+
+        template <class A, class T, class = typename std::enable_if<batch_bool<T, A>::size == 64, void>::type>
+        XSIMD_INLINE batch_bool<T, A> load_aligned(bool const* mem, batch_bool<T, A>, requires_arch<avx512bw>) noexcept
+        {
+            __m512i bool_val = _mm512_load_si512((__m512i const*)mem);
+            return _mm512_cmpgt_epu8_mask(bool_val, _mm512_setzero_si512());
         }
 
         // max

include/xsimd/arch/xsimd_avx512f.hpp

@@ -1209,14 +1209,14 @@ namespace xsimd
             }
         }
 
-        // load
+        // load mask
         template <class A, class T>
-        XSIMD_INLINE batch_bool<T, A> load(bool const* mem, batch_bool<T, A>, requires_arch<avx512f>) noexcept
+        XSIMD_INLINE batch_bool<T, A> load_unaligned(bool const* mem, batch_bool<T, A>, requires_arch<avx512f>) noexcept
         {
             using register_type = typename batch_bool<T, A>::register_type;
             constexpr auto size = batch_bool<T, A>::size;
             constexpr auto iter = size / 8;
-            static_assert(size % 8 == 0, "incorrect size of bool batch");
+            static_assert((size % 8) == 0, "incorrect size of bool batch");
             register_type mask = 0;
             for (std::size_t i = 0; i < iter; ++i)
             {

include/xsimd/types/xsimd_batch.hpp

@@ -968,13 +968,13 @@ namespace xsimd
     template <class T, class A>
     XSIMD_INLINE batch_bool<T, A> batch_bool<T, A>::load_aligned(bool const* mem) noexcept
     {
-        return kernel::load<A>(mem, batch_bool<T, A>(), A {});
+        return kernel::load_aligned<A>(mem, batch_bool<T, A>(), A {});
     }
 
     template <class T, class A>
     XSIMD_INLINE batch_bool<T, A> batch_bool<T, A>::load_unaligned(bool const* mem) noexcept
     {
-        return load_aligned(mem);
+        return kernel::load_unaligned<A>(mem, batch_bool<T, A>(), A {});
     }
 
     /**