WIP

xtensor-stack · Sep 17, 2024 · dd15471 · dd15471
1 parent fa526c1
commit dd15471
Show file tree

Hide file tree

Showing 2 changed files with 10 additions and 2 deletions.
diff --git a/include/xsimd/arch/generic/xsimd_generic_logical.hpp b/include/xsimd/arch/generic/xsimd_generic_logical.hpp
@@ -26,9 +26,9 @@ namespace xsimd
 
         // count
         template <class A, class T>
-        XSIMD_INLINE size_t count(batch_bool<T, A> const& x, requires_arch<generic>) noexcept
+        XSIMD_INLINE size_t count(batch_bool<T, A> const& self, requires_arch<generic>) noexcept
         {
-            uint64_t m = x.mask();
+            uint64_t m = self.mask();
             XSIMD_IF_CONSTEXPR(batch_bool<T, A>::size < 14)
             {
                 // https://graphics.stanford.edu/~seander/bithacks.html#CountBitsSet64
@@ -41,9 +41,14 @@ namespace xsimd
 #define builtin_popcount(v) __builtin_popcount(v)
 #endif
 #endif
+
 #ifdef builtin_popcount
                 return builtin_popcount(m);
 #else
+                alignas(A::alignment()) bool buffer[batch_bool<T, A>::size];
+                self.store_aligned(buffer);
+                return std::count(std::begin(buffer), std::end(buffer), true);
+#if 0
                 // FIXME: we could do better by dispatching to the appropriate
                 // popcount instruction depending on the arch...
                 XSIMD_IF_CONSTEXPR(batch_bool<T, A>::size < 32)
@@ -61,6 +66,7 @@ namespace xsimd
                     m = (m + (m >> 4)) & (uint64_t) ~(uint64_t)0 / 255 * 15; // temp
                     return (m * ((uint64_t) ~(uint64_t)0 / 255)) >> (sizeof(uint64_t) - 1) * CHAR_BIT; // count
                 }
+#endif
 #endif
             }
         }

diff --git a/include/xsimd/arch/xsimd_emulated.hpp b/include/xsimd/arch/xsimd_emulated.hpp
@@ -230,6 +230,7 @@ namespace xsimd
             return r;
         }
 
+#if 0
         // count
         template <class A, class T, size_t N = 8 * sizeof(T) * batch<T, A>::size>
         XSIMD_INLINE size_t count(batch_bool<T, A> const& x, requires_arch<emulated<N>>) noexcept
@@ -241,6 +242,7 @@ namespace xsimd
             m = (m + (m >> 4)) & (uint64_t) ~(uint64_t)0 / 255 * 15; // temp
             return (m * ((uint64_t) ~(uint64_t)0 / 255)) >> (sizeof(uint64_t) - 1) * CHAR_BIT; // count
         }
+#endif
 
         // store_complex
         namespace detail