sequenzo 0.1.17__cp310-cp310-macosx_10_9_universal2.whl → 0.1.19__cp310-cp310-macosx_10_9_universal2.whl

sequenzo/dissimilarity_measures/src/xsimd/include/xsimd/arch/xsimd_avx512dq.hpp

@@ -9,8 +9,8 @@
  * The full license is in the file LICENSE, distributed with this software. *
  ****************************************************************************/
 
-#ifndef XSIMD_AVX512_DQHPP
-#define XSIMD_AVX512_D_HPP
+#ifndef XSIMD_AVX512DQ_HPP
+#define XSIMD_AVX512DQ_HPP
 
 #include "../types/xsimd_avx512dq_register.hpp"
 
@@ -47,12 +47,12 @@ namespace xsimd
 
         // bitwise_not
         template <class A>
-        XSIMD_INLINE batch<float, A> bitwise_not(batch<float, A> const& self, requires_arch<avx512f>) noexcept
+        XSIMD_INLINE batch<float, A> bitwise_not(batch<float, A> const& self, requires_arch<avx512dq>) noexcept
         {
             return _mm512_xor_ps(self, _mm512_castsi512_ps(_mm512_set1_epi32(-1)));
         }
         template <class A>
-        XSIMD_INLINE batch<double, A> bitwise_not(batch<double, A> const& self, requires_arch<avx512f>) noexcept
+        XSIMD_INLINE batch<double, A> bitwise_not(batch<double, A> const& self, requires_arch<avx512dq>) noexcept
         {
             return _mm512_xor_pd(self, _mm512_castsi512_pd(_mm512_set1_epi32(-1)));
         }
@@ -96,7 +96,7 @@ namespace xsimd
             // tmp1 = [a0..8, b0..8]
             // tmp2 = [a8..f, b8..f]
 #define XSIMD_AVX512_HADDP_STEP1(I, a, b)                                \
-    batch<float, avx512f> res##I;                                        \
+    batch<float, avx512dq> res##I;                                       \
     {                                                                    \
         auto tmp1 = _mm512_shuffle_f32x4(a, b, _MM_SHUFFLE(1, 0, 1, 0)); \
         auto tmp2 = _mm512_shuffle_f32x4(a, b, _MM_SHUFFLE(3, 2, 3, 2)); \
@@ -180,7 +180,7 @@ namespace xsimd
 
         // reduce_add
         template <class A>
-        XSIMD_INLINE float reduce_add(batch<float, A> const& rhs, requires_arch<avx512f>) noexcept
+        XSIMD_INLINE float reduce_add(batch<float, A> const& rhs, requires_arch<avx512dq>) noexcept
         {
             __m256 tmp1 = _mm512_extractf32x8_ps(rhs, 1);
             __m256 tmp2 = _mm512_extractf32x8_ps(rhs, 0);
@@ -188,11 +188,43 @@ namespace xsimd
             return reduce_add(batch<float, avx2>(res1), avx2 {});
         }
 
+        // reduce_mul
+        template <class A>
+        XSIMD_INLINE float reduce_mul(batch<float, A> const& rhs, requires_arch<avx512dq>) noexcept
+        {
+            __m256 tmp1 = _mm512_extractf32x8_ps(rhs, 1);
+            __m256 tmp2 = _mm512_extractf32x8_ps(rhs, 0);
+            __m256 res1 = _mm256_mul_ps(tmp1, tmp2);
+            return reduce_mul(batch<float, avx2>(res1), avx2 {});
+        }
+
+        // swizzle constant mask
+        template <class A, uint32_t V0, uint32_t V1, uint32_t V2, uint32_t V3, uint32_t V4, uint32_t V5, uint32_t V6, uint32_t V7,
+                  uint32_t V8, uint32_t V9, uint32_t V10, uint32_t V11, uint32_t V12, uint32_t V13, uint32_t V14, uint32_t V15>
+        XSIMD_INLINE batch<float, A> swizzle(batch<float, A> const& self,
+                                             batch_constant<uint32_t, A, V0, V1, V2, V3, V4, V5, V6, V7, V8, V9, V10, V11, V12, V13, V14, V15> mask,
+                                             requires_arch<avx512dq>) noexcept
+        {
+            constexpr bool dup_lo = detail::is_dup_lo(mask);
+            constexpr bool dup_hi = detail::is_dup_hi(mask);
+
+            XSIMD_IF_CONSTEXPR(dup_lo || dup_hi)
+            {
+                const batch<float, avx2> half = _mm512_extractf32x8_ps(self, dup_lo ? 0 : 1);
+                constexpr typename std::conditional<dup_lo, batch_constant<uint32_t, avx2, V0 % 8, V1 % 8, V2 % 8, V3 % 8, V4 % 8, V5 % 8, V6 % 8, V7 % 8>,
+                                                    batch_constant<uint32_t, avx2, V8 % 8, V9 % 8, V10 % 8, V11 % 8, V12 % 8, V13 % 8, V14 % 8, V15 % 8>>::type half_mask {};
+                auto permuted = swizzle(half, half_mask, avx2 {});
+                // merge the two slices into an AVX512F register:
+                return _mm512_broadcast_f32x8(permuted); // duplicates the 256-bit perm into both halves
+            }
+            return swizzle(self, mask, avx512f {});
+        }
+
         // convert
         namespace detail
         {
             template <class A>
-            XSIMD_INLINE batch<double, A> fast_cast(batch<int64_t, A> const& x, batch<double, A> const&, requires_arch<avx512dq>) noexcept
+            XSIMD_INLINE batch<double, A> fast_cast(batch<int64_t, A> const& self, batch<double, A> const&, requires_arch<avx512dq>) noexcept
             {
                 return _mm512_cvtepi64_pd(self);
             }

sequenzo/dissimilarity_measures/src/xsimd/include/xsimd/arch/xsimd_avx512f.hpp

@@ -32,6 +32,8 @@ namespace xsimd
         XSIMD_INLINE batch<T, A> incr_if(batch<T, A> const& self, Mask const& mask, requires_arch<common>) noexcept;
         template <class A, class T, size_t I>
         XSIMD_INLINE batch<T, A> insert(batch<T, A> const& self, T val, index<I>, requires_arch<common>) noexcept;
+        template <class A, class T, class ITy, ITy... Is>
+        XSIMD_INLINE batch<T, A> swizzle(batch<T, A> const& self, batch_constant<ITy, A, Is...>, requires_arch<common>) noexcept;
         template <class A>
         XSIMD_INLINE void transpose(batch<uint16_t, A>* matrix_begin, batch<uint16_t, A>* matrix_end, requires_arch<common>) noexcept;
         template <class A>
@@ -562,6 +564,100 @@ namespace xsimd
             }
         }
 
+        // rotl
+        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
+        XSIMD_INLINE batch<T, A> rotl(batch<T, A> const& self, batch<T, A> const& other, requires_arch<avx512f>) noexcept
+        {
+            XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
+            {
+                return _mm512_rolv_epi32(self, other);
+            }
+            XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
+            {
+                return _mm512_rolv_epi64(self, other);
+            }
+            return detail::fwd_to_avx([](__m256i s, __m256i o) noexcept
+                                      { return rotl(batch<T, avx2>(s), batch<T, avx2>(o), avx2 {}); },
+                                      self, other);
+        }
+        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
+        XSIMD_INLINE batch<T, A> rotl(batch<T, A> const& self, int32_t other, requires_arch<avx512f>) noexcept
+        {
+            return rotl(self, batch<T, A>(other), A {});
+        }
+        template <size_t count, class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
+        XSIMD_INLINE batch<T, A> rotl(batch<T, A> const& self, requires_arch<avx512f>) noexcept
+        {
+            constexpr auto bits = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
+            static_assert(count < bits, "Count must be less than the number of bits in T");
+            XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
+            {
+                return _mm512_rol_epi32(self, count);
+            }
+            XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
+            {
+                return _mm512_rol_epi64(self, count);
+            }
+
+            return detail::fwd_to_avx([](__m256i s) noexcept
+                                      { return rotl<count>(batch<T, avx2>(s), avx2 {}); },
+                                      self);
+        }
+
+        // rotr
+        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
+        XSIMD_INLINE batch<T, A> rotr(batch<T, A> const& self, batch<T, A> const& other, requires_arch<avx512f>) noexcept
+        {
+            XSIMD_IF_CONSTEXPR(sizeof(T) < 4)
+            {
+                return detail::fwd_to_avx([](__m256i s, __m256i o) noexcept
+                                          { return rotr(batch<T, avx2>(s), batch<T, avx2>(o), avx2 {}); },
+                                          self, other);
+            }
+            XSIMD_IF_CONSTEXPR(std::is_unsigned<T>::value)
+            {
+                XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
+                {
+                    return _mm512_rorv_epi32(self, other);
+                }
+                else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
+                {
+                    return _mm512_rorv_epi64(self, other);
+                }
+            }
+            return rotr(self, other, common {});
+        }
+        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
+        XSIMD_INLINE batch<T, A> rotr(batch<T, A> const& self, int32_t other, requires_arch<avx512f>) noexcept
+        {
+            return rotr(self, batch<T, A>(other), A {});
+        }
+
+        template <size_t count, class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
+        XSIMD_INLINE batch<T, A> rotr(batch<T, A> const& self, requires_arch<avx512f>) noexcept
+        {
+            constexpr auto bits = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
+            static_assert(count < bits, "Count must be less than the number of bits in T");
+            XSIMD_IF_CONSTEXPR(sizeof(T) < 4)
+            {
+                return detail::fwd_to_avx([](__m256i s) noexcept
+                                          { return rotr<count>(batch<T, avx2>(s), avx2 {}); },
+                                          self);
+            }
+            XSIMD_IF_CONSTEXPR(std::is_unsigned<T>::value)
+            {
+                XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
+                {
+                    return _mm512_ror_epi32(self, count);
+                }
+                else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
+                {
+                    return _mm512_ror_epi64(self, count);
+                }
+            }
+            return rotr<count>(self, common {});
+        }
+
         // bitwise_xor
         template <class A>
         XSIMD_INLINE batch<float, A> bitwise_xor(batch<float, A> const& self, batch<float, A> const& other, requires_arch<avx512f>) noexcept
@@ -578,7 +674,7 @@ namespace xsimd
         XSIMD_INLINE batch_bool<T, A> bitwise_xor(batch_bool<T, A> const& self, batch_bool<T, A> const& other, requires_arch<avx512f>) noexcept
         {
             using register_type = typename batch_bool<T, A>::register_type;
-            return register_type(self.data | other.data);
+            return register_type(self.data ^ other.data);
         }
 
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
@@ -900,6 +996,18 @@ namespace xsimd
         {
             return _mm512_fmsub_pd(x, y, z);
         }
+        // fmas
+        template <class A>
+        XSIMD_INLINE batch<float, A> fmas(batch<float, A> const& x, batch<float, A> const& y, batch<float, A> const& z, requires_arch<avx512f>) noexcept
+        {
+            return _mm512_fmaddsub_ps(x, y, z);
+        }
+
+        template <class A>
+        XSIMD_INLINE batch<double, A> fmas(batch<double, A> const& x, batch<double, A> const& y, batch<double, A> const& z, requires_arch<avx512f>) noexcept
+        {
+            return _mm512_fmaddsub_pd(x, y, z);
+        }
 
         // from bool
         template <class A, class T>
@@ -1312,12 +1420,12 @@ namespace xsimd
         template <class A>
         XSIMD_INLINE batch<float, A> max(batch<float, A> const& self, batch<float, A> const& other, requires_arch<avx512f>) noexcept
         {
-            return _mm512_max_ps(self, other);
+            return _mm512_max_ps(other, self);
         }
         template <class A>
         XSIMD_INLINE batch<double, A> max(batch<double, A> const& self, batch<double, A> const& other, requires_arch<avx512f>) noexcept
         {
-            return _mm512_max_pd(self, other);
+            return _mm512_max_pd(other, self);
         }
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
         XSIMD_INLINE batch<T, A> max(batch<T, A> const& self, batch<T, A> const& other, requires_arch<avx512f>) noexcept
@@ -1362,12 +1470,12 @@ namespace xsimd
         template <class A>
         XSIMD_INLINE batch<float, A> min(batch<float, A> const& self, batch<float, A> const& other, requires_arch<avx512f>) noexcept
         {
-            return _mm512_min_ps(self, other);
+            return _mm512_min_ps(other, self);
         }
         template <class A>
         XSIMD_INLINE batch<double, A> min(batch<double, A> const& self, batch<double, A> const& other, requires_arch<avx512f>) noexcept
         {
-            return _mm512_min_pd(self, other);
+            return _mm512_min_pd(other, self);
         }
         template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
         XSIMD_INLINE batch<T, A> min(batch<T, A> const& self, batch<T, A> const& other, requires_arch<avx512f>) noexcept
@@ -1544,6 +1652,37 @@ namespace xsimd
             return reduce_min(batch<T, avx2>(low));
         }
 
+        // reduce_mul
+        template <class A>
+        XSIMD_INLINE float reduce_mul(batch<float, A> const& rhs, requires_arch<avx512f>) noexcept
+        {
+            return _mm512_reduce_mul_ps(rhs);
+        }
+        template <class A>
+        XSIMD_INLINE double reduce_mul(batch<double, A> const& rhs, requires_arch<avx512f>) noexcept
+        {
+            return _mm512_reduce_mul_pd(rhs);
+        }
+        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
+        XSIMD_INLINE T reduce_mul(batch<T, A> const& self, requires_arch<avx512f>) noexcept
+        {
+            XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
+            {
+                return _mm512_reduce_mul_epi32(self);
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
+            {
+                return _mm512_reduce_mul_epi64(self);
+            }
+            else
+            {
+                __m256i low, high;
+                detail::split_avx512(self, low, high);
+                batch<T, avx2> blow(low), bhigh(high);
+                return reduce_mul(blow, avx2 {}) * reduce_mul(bhigh, avx2 {});
+            }
+        }
+
         // rsqrt
         template <class A>
         XSIMD_INLINE batch<float, A> rsqrt(batch<float, A> const& val, requires_arch<avx512f>) noexcept
@@ -1726,8 +1865,8 @@ namespace xsimd
                 v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31
             };
 #else
-            return _mm512_set_epi16(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15,
-                                    v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31);
+            return _mm512_set_epi16(v31, v30, v29, v28, v27, v26, v25, v24, v23, v22, v21, v20, v19, v18, v17, v16,
+                                    v15, v14, v13, v12, v11, v10, v9, v8, v7, v6, v5, v4, v3, v2, v1, v0);
 #endif
         }
 
@@ -1743,8 +1882,8 @@ namespace xsimd
                 v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31
             };
 #else
-            return _mm512_set_epi16(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15,
-                                    v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31);
+            return _mm512_set_epi16(v31, v30, v29, v28, v27, v26, v25, v24, v23, v22, v21, v20, v19, v18, v17, v16,
+                                    v15, v14, v13, v12, v11, v10, v9, v8, v7, v6, v5, v4, v3, v2, v1, v0);
 #endif
         }
 
@@ -1767,10 +1906,10 @@ namespace xsimd
                 v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63
             };
 #else
-            return _mm512_set_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15,
-                                   v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31,
-                                   v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47,
-                                   v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63);
+            return _mm512_set_epi8(v63, v62, v61, v60, v59, v58, v57, v56, v55, v54, v53, v52, v51, v50, v49, v48,
+                                   v47, v46, v45, v44, v43, v42, v41, v40, v39, v38, v37, v36, v35, v34, v33, v32,
+                                   v31, v30, v29, v28, v27, v26, v25, v24, v23, v22, v21, v20, v19, v18, v17, v16,
+                                   v15, v14, v13, v12, v11, v10, v9, v8, v7, v6, v5, v4, v3, v2, v1, v0);
 #endif
         }
         template <class A, class T, detail::enable_unsigned_integer_t<T> = 0>
@@ -1792,10 +1931,10 @@ namespace xsimd
                 v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63
             };
 #else
-            return _mm512_set_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15,
-                                   v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31,
-                                   v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47,
-                                   v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63);
+            return _mm512_set_epi8(v63, v62, v61, v60, v59, v58, v57, v56, v55, v54, v53, v52, v51, v50, v49, v48,
+                                   v47, v46, v45, v44, v43, v42, v41, v40, v39, v38, v37, v36, v35, v34, v33, v32,
+                                   v31, v30, v29, v28, v27, v26, v25, v24, v23, v22, v21, v20, v19, v18, v17, v16,
+                                   v15, v14, v13, v12, v11, v10, v9, v8, v7, v6, v5, v4, v3, v2, v1, v0);
 #endif
         }
 
@@ -1845,19 +1984,110 @@ namespace xsimd
         }
 
         // slide_left
+        namespace detail
+        {
+            template <size_t N>
+            struct make_slide_left_pattern
+            {
+                static constexpr size_t get(size_t i, size_t)
+                {
+                    return i >= N ? i - N : 0;
+                }
+            };
+
+            template <size_t N, class A, class T>
+            XSIMD_INLINE batch<T, A> slide_left_aligned_u32(batch<T, A> const& x, requires_arch<avx512f>) noexcept
+            {
+                static_assert((N & 3) == 0 || N >= 64, "N must be aligned to 32 bits");
+
+                if (N == 0)
+                {
+                    return x;
+                }
+                if (N >= 64)
+                {
+                    return batch<T, A>(T(0));
+                }
+
+                __mmask16 mask = uint16_t(0xFFFFu << (N / 4));
+
+                if ((N & 15) == 0)
+                {
+                    const uint8_t imm8 = uint8_t(0xe4 << (2 * (N / 16)));
+                    return _mm512_maskz_shuffle_i32x4(mask, x, x, imm8);
+                }
+
+                auto slide_pattern = make_batch_constant<uint32_t, detail::make_slide_left_pattern<N / 4>, A>();
+                return _mm512_maskz_permutexvar_epi32(mask, slide_pattern.as_batch(), x);
+            }
+        }
+
         template <size_t N, class A, class T>
-        XSIMD_INLINE batch<T, A> slide_left(batch<T, A> const&, requires_arch<avx512f>) noexcept
+        XSIMD_INLINE batch<T, A> slide_left(batch<T, A> const& x, requires_arch<avx512f>) noexcept
         {
-            static_assert(N == 0xDEAD, "not implemented yet");
-            return {};
+            constexpr size_t NN = N & ~3;
+            if (N == NN || NN >= 64)
+            {
+                // Call fast path
+                return detail::slide_left_aligned_u32<NN>(x, A {});
+            }
+
+            __m512i xl = detail::slide_left_aligned_u32<NN, A, T>(_mm512_slli_epi32(x, 8 * (N - NN)), A {});
+            __m512i xr = detail::slide_left_aligned_u32<NN + 4, A, T>(_mm512_srli_epi32(x, 32 - 8 * (N - NN)), A {});
+            return _mm512_or_epi32(xl, xr);
         }
 
         // slide_right
+        namespace detail
+        {
+            template <size_t N>
+            struct make_slide_right_pattern
+            {
+                static constexpr size_t get(size_t i, size_t n)
+                {
+                    return i < (n - N) ? i + N : 0;
+                }
+            };
+
+            template <size_t N, class A, class T>
+            XSIMD_INLINE batch<T, A> slide_right_aligned_u32(batch<T, A> const& x, requires_arch<avx512f>) noexcept
+            {
+                static_assert((N & 3) == 0 || N >= 64, "N must be aligned to 32 bits");
+
+                if (N == 0)
+                {
+                    return x;
+                }
+                if (N >= 64)
+                {
+                    return batch<T, A>(T(0));
+                }
+
+                __mmask16 mask = 0xFFFFu >> (N / 4);
+
+                if ((N & 15) == 0)
+                {
+                    const uint8_t imm8 = 0xe4 >> (2 * (N / 16));
+                    return _mm512_maskz_shuffle_i32x4(mask, x, x, imm8);
+                }
+
+                auto slide_pattern = make_batch_constant<uint32_t, detail::make_slide_right_pattern<N / 4>, A>();
+                return _mm512_maskz_permutexvar_epi32(mask, slide_pattern.as_batch(), x);
+            }
+        }
         template <size_t N, class A, class T>
-        XSIMD_INLINE batch<T, A> slide_right(batch<T, A> const&, requires_arch<avx512f>) noexcept
+        XSIMD_INLINE batch<T, A> slide_right(batch<T, A> const& x, requires_arch<avx512f>) noexcept
         {
-            static_assert(N == 0xDEAD, "not implemented yet");
-            return {};
+            constexpr size_t NN = N & ~3;
+            if (N == NN || NN >= 64)
+            {
+                // Call fast path
+                return detail::slide_right_aligned_u32<NN>(x, A {});
+            }
+
+            __m512i xl = detail::slide_right_aligned_u32<NN + 4, A, T>(_mm512_slli_epi32(x, 32 - 8 * (N - NN)), A {});
+            __m512i xr = detail::slide_right_aligned_u32<NN, A, T>(_mm512_srli_epi32(x, 8 * (N - NN)), A {});
+            return _mm512_or_epi32(xl, xr);
         }
 
         // sqrt
@@ -2019,16 +2249,53 @@ namespace xsimd
             return bitwise_cast<int32_t>(swizzle(bitwise_cast<uint32_t>(self), mask, avx512f {}));
         }
 
-        // swizzle (constant version)
-        template <class A, uint32_t... Vs>
-        XSIMD_INLINE batch<float, A> swizzle(batch<float, A> const& self, batch_constant<uint32_t, A, Vs...> mask, requires_arch<avx512f>) noexcept
+        template <class A, uint32_t V0, uint32_t V1, uint32_t V2, uint32_t V3, uint32_t V4, uint32_t V5, uint32_t V6, uint32_t V7,
+                  uint32_t V8, uint32_t V9, uint32_t V10, uint32_t V11, uint32_t V12, uint32_t V13, uint32_t V14, uint32_t V15>
+        XSIMD_INLINE batch<float, A> swizzle(batch<float, A> const& self,
+                                             batch_constant<uint32_t, A, V0, V1, V2, V3, V4, V5, V6, V7, V8, V9, V10, V11, V12, V13, V14, V15> mask,
+                                             requires_arch<avx512f>) noexcept
         {
+            XSIMD_IF_CONSTEXPR(detail::is_identity(mask))
+            {
+                return self;
+            }
+            XSIMD_IF_CONSTEXPR(!detail::is_cross_lane(mask))
+            {
+                constexpr int imm0 = detail::mod_shuffle(V0, V1, V2, V3);
+                constexpr int imm1 = detail::mod_shuffle(V4, V5, V6, V7);
+                constexpr int imm2 = detail::mod_shuffle(V8, V9, V10, V11);
+                constexpr int imm3 = detail::mod_shuffle(V12, V13, V14, V15);
+                XSIMD_IF_CONSTEXPR(imm0 == imm1 && imm0 == imm2 && imm0 == imm3)
+                {
+                    return _mm512_permute_ps(self, imm0);
+                }
+            }
             return swizzle(self, mask.as_batch(), avx512f {});
         }
-
-        template <class A, uint64_t... Vs>
-        XSIMD_INLINE batch<double, A> swizzle(batch<double, A> const& self, batch_constant<uint64_t, A, Vs...> mask, requires_arch<avx512f>) noexcept
+        template <class A, uint64_t V0, uint64_t V1, uint64_t V2, uint64_t V3, uint64_t V4, uint64_t V5, uint64_t V6, uint64_t V7>
+        XSIMD_INLINE batch<double, A> swizzle(batch<double, A> const& self,
+                                              batch_constant<uint64_t, A, V0, V1, V2, V3, V4, V5, V6, V7> mask,
+                                              requires_arch<avx512f>) noexcept
         {
+            XSIMD_IF_CONSTEXPR(detail::is_identity(mask))
+            {
+                return self;
+            }
+            XSIMD_IF_CONSTEXPR(!detail::is_cross_lane(mask))
+            {
+                constexpr auto imm = ((V0 & 1) << 0) | ((V1 & 1) << 1) | ((V2 & 1) << 2) | ((V3 & 1) << 3) | ((V4 & 1) << 4) | ((V5 & 1) << 5) | ((V6 & 1) << 6) | ((V7 & 1) << 7);
+                return _mm512_permute_pd(self, imm);
+            }
+            constexpr bool dup_lo = detail::is_dup_lo(mask);
+            constexpr bool dup_hi = detail::is_dup_hi(mask);
+            XSIMD_IF_CONSTEXPR(dup_lo || dup_hi)
+            {
+                const batch<double, avx2> half = _mm512_extractf64x4_pd(self, dup_lo ? 0 : 1);
+                constexpr typename std::conditional<dup_lo, batch_constant<uint64_t, avx2, V0 % 4, V1 % 4, V2 % 4, V3 % 4>,
+                                                    batch_constant<uint64_t, avx2, V4 % 4, V5 % 4, V6 % 4, V7 % 4>>::type half_mask {};
+                return _mm512_broadcast_f64x4(swizzle(half, half_mask, avx2 {}));
+            }
+            // General case
             return swizzle(self, mask.as_batch(), avx512f {});
         }
 
@@ -2337,8 +2604,47 @@ namespace xsimd
                 2));
         }
 
-    }
+        // first
+        template <class A>
+        XSIMD_INLINE float first(batch<float, A> const& self, requires_arch<avx512f>) noexcept
+        {
+            return _mm512_cvtss_f32(self);
+        }
+
+        template <class A>
+        XSIMD_INLINE double first(batch<double, A> const& self, requires_arch<avx512f>) noexcept
+        {
+            return _mm512_cvtsd_f64(self);
+        }
 
+        template <class A, class T, class = typename std::enable_if<std::is_integral<T>::value, void>::type>
+        XSIMD_INLINE T first(batch<T, A> const& self, requires_arch<avx512f>) noexcept
+        {
+            XSIMD_IF_CONSTEXPR(sizeof(T) == 1)
+            {
+                return static_cast<T>(_mm512_cvtsi512_si32(self) & 0xFF);
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 2)
+            {
+                return static_cast<T>(_mm512_cvtsi512_si32(self) & 0xFFFF);
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 4)
+            {
+                return static_cast<T>(_mm512_cvtsi512_si32(self));
+            }
+            else XSIMD_IF_CONSTEXPR(sizeof(T) == 8)
+            {
+                batch<T, sse4_2> low = _mm512_castsi512_si128(self);
+                return first(low, sse4_2 {});
+            }
+            else
+            {
+                assert(false && "unsupported arch/op combination");
+                return {};
+            }
+        }
+
+    }
 }
 
 #endif

sequenzo/dissimilarity_measures/src/xsimd/include/xsimd/arch/xsimd_avx512vbmi.hpp

@@ -24,54 +24,26 @@ namespace xsimd
     {
         using namespace types;
 
-        namespace detail
-        {
-            template <size_t N, size_t... Is>
-            constexpr std::array<uint8_t, sizeof...(Is)> make_slide_left_bytes_pattern(::xsimd::detail::index_sequence<Is...>)
-            {
-                return { (Is >= N ? Is - N : 0)... };
-            }
-
-            template <size_t N, size_t... Is>
-            constexpr std::array<uint8_t, sizeof...(Is)> make_slide_right_bytes_pattern(::xsimd::detail::index_sequence<Is...>)
-            {
-                return { (Is < (64 - N) ? Is + N : 0)... };
-            }
-        }
-
         // slide_left
-        template <size_t N, class A, class T>
+        template <size_t N, class A, class T, class = typename std::enable_if<(N & 3) != 0 && (N < 64)>::type>
         XSIMD_INLINE batch<T, A> slide_left(batch<T, A> const& x, requires_arch<avx512vbmi>) noexcept
         {
-            if (N == 0)
-            {
-                return x;
-            }
-            if (N >= 64)
-            {
-                return batch<T, A>(T(0));
-            }
+            static_assert((N & 3) != 0 && N < 64, "The AVX512F implementation may have a lower latency.");
 
             __mmask64 mask = 0xFFFFFFFFFFFFFFFFull << (N & 63);
-            alignas(A::alignment()) auto slide_pattern = detail::make_slide_left_bytes_pattern<N>(::xsimd::detail::make_index_sequence<512 / 8>());
-            return _mm512_maskz_permutexvar_epi8(mask, _mm512_load_epi32(slide_pattern.data()), x);
+            auto slide_pattern = make_batch_constant<uint8_t, detail::make_slide_left_pattern<N>, A>();
+            return _mm512_maskz_permutexvar_epi8(mask, slide_pattern.as_batch(), x);
         }
 
         // slide_right
-        template <size_t N, class A, class T>
+        template <size_t N, class A, class T, class = typename std::enable_if<(N & 3) != 0 && (N < 64)>::type>
         XSIMD_INLINE batch<T, A> slide_right(batch<T, A> const& x, requires_arch<avx512vbmi>) noexcept
         {
-            if (N == 0)
-            {
-                return x;
-            }
-            if (N >= 64)
-            {
-                return batch<T, A>(T(0));
-            }
+            static_assert((N & 3) != 0 && N < 64, "The AVX512F implementation may have a lower latency.");
+
             __mmask64 mask = 0xFFFFFFFFFFFFFFFFull >> (N & 63);
-            alignas(A::alignment()) auto slide_pattern = detail::make_slide_right_bytes_pattern<N>(::xsimd::detail::make_index_sequence<512 / 8>());
-            return _mm512_maskz_permutexvar_epi8(mask, _mm512_load_epi32(slide_pattern.data()), x);
+            auto slide_pattern = make_batch_constant<uint8_t, detail::make_slide_right_pattern<N>, A>();
+            return _mm512_maskz_permutexvar_epi8(mask, slide_pattern.as_batch(), x);
         }
 
         // swizzle (dynamic version)