Feature/view to simd

include/seqan3/core/simd/all.hpp

@@ -36,6 +36,7 @@
 #include <seqan3/core/simd/simd_algorithm.hpp>
 #include <seqan3/core/simd/simd_traits.hpp>
 #include <seqan3/core/simd/simd.hpp>
+#include <seqan3/core/simd/view_to_simd.hpp>
 
 /*!\namespace seqan3::simd
  * \brief The SeqAn3 namespace for simd data types, algorithms and meta functions.

include/seqan3/core/simd/concept.hpp

@@ -31,34 +31,34 @@ namespace seqan3::detail
 template <typename simd_t>
 SEQAN3_CONCEPT Simd = requires (simd_t a, simd_t b)
 {
-    typename simd_traits<simd_t>::scalar_type;
-    typename simd_traits<simd_t>::mask_type;
-    typename simd_traits<simd_t>::swizzle_type;
+    typename simd_traits<std::remove_reference_t<simd_t>>::scalar_type;
+    typename simd_traits<std::remove_reference_t<simd_t>>::mask_type;
+    typename simd_traits<std::remove_reference_t<simd_t>>::swizzle_type;
 
     // require that static member variables are defined
-    requires std::Integral<decltype(simd_traits<simd_t>::length)>;
-    requires std::Integral<decltype(simd_traits<simd_t>::max_length)>;
+    requires std::Integral<decltype(simd_traits<std::remove_reference_t<simd_t>>::length)>;
+    requires std::Integral<decltype(simd_traits<std::remove_reference_t<simd_t>>::max_length)>;
 
     // assume array access that returns a scalar_type type
-    { a[0] } -> typename simd_traits<simd_t>::scalar_type;
+    { a[0] } -> typename simd_traits<std::remove_reference_t<simd_t>>::scalar_type;
 
     // require comparison operators
-    requires std::Same<decltype(a == b), typename simd_traits<simd_t>::mask_type>;
-    requires std::Same<decltype(a != b), typename simd_traits<simd_t>::mask_type>;
-    requires std::Same<decltype(a <  b), typename simd_traits<simd_t>::mask_type>;
-    requires std::Same<decltype(a >  b), typename simd_traits<simd_t>::mask_type>;
-    requires std::Same<decltype(a <= b), typename simd_traits<simd_t>::mask_type>;
-    requires std::Same<decltype(a >= b), typename simd_traits<simd_t>::mask_type>;
+    requires std::Same<decltype(a == b), typename simd_traits<std::remove_reference_t<simd_t>>::mask_type>;
+    requires std::Same<decltype(a != b), typename simd_traits<std::remove_reference_t<simd_t>>::mask_type>;
+    requires std::Same<decltype(a <  b), typename simd_traits<std::remove_reference_t<simd_t>>::mask_type>;
+    requires std::Same<decltype(a >  b), typename simd_traits<std::remove_reference_t<simd_t>>::mask_type>;
+    requires std::Same<decltype(a <= b), typename simd_traits<std::remove_reference_t<simd_t>>::mask_type>;
+    requires std::Same<decltype(a >= b), typename simd_traits<std::remove_reference_t<simd_t>>::mask_type>;
 
     // require arithmetic operators
-    requires std::Same<decltype(a + b), simd_t>;
-    requires std::Same<decltype(a - b), simd_t>;
-    requires std::Same<decltype(a * b), simd_t>;
-    requires std::Same<decltype(a / b), simd_t>;
-    requires std::Same<decltype(a += b), simd_t &>;
-    requires std::Same<decltype(a -= b), simd_t &>;
-    requires std::Same<decltype(a *= b), simd_t &>;
-    requires std::Same<decltype(a /= b), simd_t &>;
+    requires std::Same<decltype(a + b), std::remove_reference_t<simd_t>>;
+    requires std::Same<decltype(a - b), std::remove_reference_t<simd_t>>;
+    requires std::Same<decltype(a * b), std::remove_reference_t<simd_t>>;
+    requires std::Same<decltype(a / b), std::remove_reference_t<simd_t>>;
+    requires std::Same<decltype(a += b), std::remove_reference_t<simd_t> &>;
+    requires std::Same<decltype(a -= b), std::remove_reference_t<simd_t> &>;
+    requires std::Same<decltype(a *= b), std::remove_reference_t<simd_t> &>;
+    requires std::Same<decltype(a /= b), std::remove_reference_t<simd_t> &>;
 };
 //!\endcond
 

include/seqan3/core/simd/detail/builtin_simd.hpp

@@ -101,14 +101,15 @@ struct builtin_simd_traits_helper<builtin_simd_t>
 
     //!\brief Whether builtin_simd_t is a builtin type or not?
     //!\hideinitializer
-    static constexpr bool value = is_power_of_two(length) && std::is_same_v<builtin_simd_t, transformation_trait_or_t<builtin_simd<scalar_type, length>, void>>;
+    static constexpr bool value = is_power_of_two(length) &&
+                                  std::is_same_v<builtin_simd_t,
+                                                 transformation_trait_or_t<builtin_simd<scalar_type, length>, void>>;
 };
 
 /*!\brief This class inherits from std::true_type, **iff**
  * seqan3::detail::builtin_simd<scalar_t, length>::type is a builtin simd type.
  * \ingroup simd
- * \tparam scalar_type The underlying type of a simd vector
- * \tparam length_v The number of packed values in a simd vector
+ * \tparam builtin_simd_t The type to check.
  *
  * \include test/snippet/core/simd/detail/is_builtin_simd.cpp
  * \sa https://gcc.gnu.org/onlinedocs/gcc/Vector-Extensions.html
@@ -117,6 +118,14 @@ template <typename builtin_simd_t>
 struct is_builtin_simd : std::bool_constant<builtin_simd_traits_helper<builtin_simd_t>::value>
 {};
 
+/*!\brief Helper variable to test whether a type is a simd builtin type.
+ * \ingroup simd
+ * \tparam builtin_simd_t The type to check.
+ * \see seqan3::detail::is_builtin_simd
+ */
+template <typename builtin_simd_t>
+constexpr bool is_builtin_simd_v = is_builtin_simd<builtin_simd_t>::value;
+
 /*!\brief This function specializes seqan3::detail::default_simd_max_length for
  * seqan3::detail::builtin_simd.
  * \ingroup simd
@@ -139,6 +148,36 @@ constexpr auto default_simd_max_length<builtin_simd> = []()
 #endif
 }();
 
+/*!\brief This class inherits from std::true_type, **iff** the builtin simd type is supported by the current
+ *        architecture.
+ * \ingroup simd
+ * \tparam builtin_simd_t The type to check.
+ *
+ * \details
+ *
+ * A builtin simd type is native if the following conditions are true:
+ * * the default simd max length is not equal to `0`.
+ * * the max length of the simd type is at least 16 (SSE4)
+ * * the max length of the simd type is at most 64 (AVX512)
+ */
+template <typename builtin_simd_t>
+struct is_native_builtin_simd :
+    std::bool_constant<(default_simd_max_length<builtin_simd> != 0) &&
+                       ((builtin_simd_traits_helper<builtin_simd_t>::length *
+                            sizeof(typename builtin_simd_traits_helper<builtin_simd_t>::scalar_type)) >= 16) &&
+                       ((builtin_simd_traits_helper<builtin_simd_t>::length *
+                            sizeof(typename builtin_simd_traits_helper<builtin_simd_t>::scalar_type)) <= 64)>
+{};
+
+
+/*!\brief Helper variable to test whether a type is a native simd builtin type.
+ * \ingroup simd
+ * \tparam builtin_simd_t The type to check.
+ * \see seqan3::detail::is_native_builtin_simd_v
+ */
+template <typename builtin_simd_t>
+constexpr bool is_native_builtin_simd_v = is_native_builtin_simd<builtin_simd_t>::value;
+
 } // namespace seqan3::detail
 
 namespace seqan3

include/seqan3/core/simd/detail/simd_algorithm_sse4.hpp

@@ -0,0 +1,102 @@
+// -----------------------------------------------------------------------------------------------------
+// Copyright (c) 2006-2019, Knut Reinert & Freie Universität Berlin
+// Copyright (c) 2016-2019, Knut Reinert & MPI für molekulare Genetik
+// This file may be used, modified and/or redistributed under the terms of the 3-clause BSD-License
+// shipped with this file and also available at: https://github.com/seqan/seqan3/blob/master/LICENSE.md
+// -----------------------------------------------------------------------------------------------------
+
+/*!\file
+ * \brief Provides specific algorithm implementations for SSE4 instruction set.
+ * \author Rene Rahn <rene.rahn AT fu-berlin.de>
+ */
+
+#pragma once
+
+#include <array>
+#include <immintrin.h>
+
+#include <seqan3/core/simd/concept.hpp>
+#include <seqan3/core/simd/detail/builtin_simd.hpp>
+#include <seqan3/core/simd/simd_traits.hpp>
+
+namespace seqan3::detail
+{
+
+/*!\brief Transposes the given simd vector matrix using SSE4 intrinsics.
+ * \ingroup simd
+ * \tparam simd_t The simd vector type; must model seqan3::detail::Simd and must be a simd built-in and native type.
+ * \param[in,out] matrix The matrix that is transposed in place.
+ *
+ * \details
+ *
+ * Does inplace transformation using SSE4 intrinsics.
+ */
+template <Simd simd_t>
+inline void transpose_matrix_sse4(std::array<simd_t, simd_traits<simd_t>::length> & matrix)
+{
+    static_assert(simd_traits<simd_t>::length == simd_traits<simd_t>::max_length, "Expects byte scalar type.");
+    static_assert(is_native_builtin_simd_v<simd_t>, "The passed simd vector is not a native SSE4 simd vector type.");
+    static_assert(is_builtin_simd_v<simd_t>, "The passed simd vector is not a builtin vector type.");
+
+    // we need a look-up table to reverse the lowest 4 bits
+    // in order to place the permute the transposed rows
+    constexpr std::array<char, 16> bit_reverse{0, 8, 4, 12, 2, 10, 6, 14, 1, 9, 5, 13, 3, 11, 7, 15};
+
+    // transpose a 16x16 byte matrix
+    //
+    // matrix =
+    // A0 A1 A2 ... Ae Af
+    // B0 B1 B2 ... Be Bf
+    // ...
+    // P0 P1 P2 ... Pe Pf
+    __m128i tmp1[16];
+    for (int i = 0; i < 8; ++i)
+    {
+        tmp1[i]   = _mm_unpacklo_epi8(reinterpret_cast<__m128i &>(matrix[2*i]),
+                                      reinterpret_cast<__m128i &>(matrix[2*i+1]));
+        tmp1[i+8] = _mm_unpackhi_epi8(reinterpret_cast<__m128i &>(matrix[2*i]),
+                                      reinterpret_cast<__m128i &>(matrix[2*i+1]));
+    }
+    // tmp1[0]  = A0 B0 A1 B1 ... A7 B7
+    // tmp1[1]  = C0 D0 C1 D1 ... C7 D7
+    // ...
+    // tmp1[7]  = O0 P0 O1 P1 ... O7 P7
+    // tmp1[8]  = A8 B8 A9 B9 ... Af Bf
+    // ...
+    // tmp1[15] = O8 P8 O9 P9 ... Of Pf
+    __m128i tmp2[16];
+    for (int i = 0; i < 8; ++i)
+    {
+        tmp2[i]   = _mm_unpacklo_epi16(tmp1[2*i], tmp1[2*i+1]);
+        tmp2[i+8] = _mm_unpackhi_epi16(tmp1[2*i], tmp1[2*i+1]);
+    }
+    // tmp2[0]  = A0 B0 C0 D0 ... A3 B3 C3 D3
+    // tmp2[1]  = E0 F0 G0 H0 ... E3 F3 G3 H3
+    // ...
+    // tmp2[3]  = M0 N0 O0 P0 ... M3 N3 O3 P3
+    // tmp2[4]  = A8 B8 C8 D8 ... Ab Bb Cb Db
+    // ...
+    // tmp2[7]  = M8 N8 O8 P8 ... Mb Nb Ob Pb
+    // tmp2[8]  = A4 B4 C4 D4 ... A7 B7 C7 D7
+    // ..
+    // tmp2[12] = Ac Bc Cc Dc ... Af Bf Cf Df
+    // ...
+    // tmp2[15] = Mc Nc Oc Pc ... Mf Nf Of Pf
+    for (int i = 0; i < 8; ++i)
+    {
+        tmp1[i]   = _mm_unpacklo_epi32(tmp2[2*i], tmp2[2*i+1]);
+        tmp1[i+8] = _mm_unpackhi_epi32(tmp2[2*i], tmp2[2*i+1]);
+    }
+    // tmp1[0]  = A0 B0 .... H0 A1 B1 .... H1
+    // tmp1[1]  = I0 J0 .... P0 I1 J1 .... P1
+    // ...
+    // tmp1[4]  = A0 B0 .... H0 A1 B1 .... H1
+    // tmp1[1]  = I0 J0 .... P0 I1 J1 .... P1
+    for (int i = 0; i < 8; ++i)
+    {
+        matrix[bit_reverse[i]]   = reinterpret_cast<simd_t>(_mm_unpacklo_epi64(tmp1[2*i], tmp1[2*i+1]));
+        matrix[bit_reverse[i+8]] = reinterpret_cast<simd_t>(_mm_unpackhi_epi64(tmp1[2*i], tmp1[2*i+1]));
+    }
+}
+
+} // namespace seqan3::detail