Factorize sum style cleanup into remove_garbage

jfalcou · Jun 30, 2022 · c50e4f5 · c50e4f5
1 parent 49b9aa2
commit c50e4f5
Show file tree

Hide file tree

Showing 4 changed files with 26 additions and 40 deletions.
diff --git a/include/eve/detail/function/simd/arm/neon/sum.hpp b/include/eve/detail/function/simd/arm/neon/sum.hpp
@@ -9,27 +9,10 @@
 
 #include <eve/detail/abi.hpp>
 #include <eve/detail/category.hpp>
-#include <eve/module/core/regular/slide_right.hpp>
+#include <eve/detail/remove_garbage.hpp>
 
 namespace eve::detail
 {
-  template<typename T, typename N>
-  EVE_FORCEINLINE wide<T,N> arm_cleanup(wide<T,N> v) noexcept
-  {
-    // Clean up potential garbage
-    using ec_t = expected_cardinal_t<T,arm_64_>;
-    if constexpr(N::value < ec_t::value)
-    {
-      v = bit_cast( slide_right ( bit_cast(v,as<wide<T,ec_t>>())
-                                , index<ec_t::value - N::value>
-                                )
-                  , as(v)
-                  );
-    }
-
-    return v;
-  }
-
   template<typename T, typename N>
   EVE_FORCEINLINE wide<T,N> arm_sum_impl(wide<T,N> v) noexcept
   {
@@ -57,7 +40,7 @@ namespace eve::detail
     {
       if constexpr( std::same_as<abi_t<T,N>, arm_64_> )
       {
-        v = arm_cleanup(v);
+        v = slide_garbage(v);
         if constexpr(sizeof(T) <= 4)  v = arm_sum_impl(v);
         if constexpr(sizeof(T) <= 2)  v = arm_sum_impl(v);
         if constexpr(sizeof(T) <= 1)  v = arm_sum_impl(v);
@@ -85,7 +68,7 @@ namespace eve::detail
     {
       if constexpr(current_api >= asimd)
       {
-        v = arm_cleanup(v);
+        v = slide_garbage(v);
         constexpr auto c = categorize<wide<T, N>>();
 
               if constexpr( c== category::float64x2 ) return vaddvq_f64(v);

diff --git a/include/eve/detail/function/simd/x86/sum.hpp b/include/eve/detail/function/simd/x86/sum.hpp
@@ -9,7 +9,7 @@
 
 #include <eve/detail/abi.hpp>
 #include <eve/detail/category.hpp>
-#include <eve/module/core/regular/slide_right.hpp>
+#include <eve/detail/remove_garbage.hpp>
 
 namespace eve::detail
 {
@@ -39,11 +39,7 @@ namespace eve::detail
     else  if constexpr( c == category::float32x4  )
     {
       // Clean up garbage if needed
-      if constexpr(N::value == 2) v = bit_cast( slide_right ( bit_cast(v,as<wide<T,fixed<4>>>())
-                                                            , index<2>
-                                                            )
-                                              , as(v)
-                                              );
+      v = slide_garbage(v);
 
       if constexpr( current_api >= sse3 )
       {
@@ -63,11 +59,7 @@ namespace eve::detail
     else  if constexpr( c == category::int32x4 || c == category::uint32x4 )
     {
       // Clean up garbage if needed
-      if constexpr(N::value == 2) v = bit_cast( slide_right ( bit_cast(v,as<wide<T,fixed<4>>>())
-                                                            , index<2>
-                                                            )
-                                              , as(v)
-                                              );
+      v = slide_garbage(v);
 
       constexpr auto shuf =  _MM_SHUFFLE(1, 0, 3, 2);
       __m128i sum64;
@@ -81,14 +73,7 @@ namespace eve::detail
     else  if constexpr( c == category::uint8x16 || c == category::int8x16 )
     {
       // Clean up garbage if needed
-      if constexpr(N::value < 16)
-      {
-        v = bit_cast( slide_right ( bit_cast(v,as<wide<T,fixed<16>>>())
-                                  , index<16-N::value>
-                                  )
-                    , as(v)
-                    );
-      }
+      v = slide_garbage(v);
 
       // https://stackoverflow.com/questions/36998538/fastest-way-to-horizontally-sum-sse-unsigned-byte-vector
       __m128i vsum = _mm_sad_epu8(v, _mm_setzero_si128());

diff --git a/include/eve/detail/remove_garbage.hpp b/include/eve/detail/remove_garbage.hpp
@@ -9,6 +9,7 @@
 
 #include <eve/traits.hpp>
 #include <eve/detail/function/bit_cast.hpp>
+#include <eve/module/core/regular/slide_right.hpp>
 #include <utility>
 
 namespace eve::detail
@@ -29,6 +30,24 @@ namespace eve::detail
       v = bit_cast(bit_cast(v,as<as_wide_t<v_t,ec_t>>()) & exact_mask, as(v));
     }
 
+    return v;
+  }
+
+ template<typename W> auto slide_garbage(W v) noexcept
+  {
+    using v_t   = element_type_t<W>;
+    using c_t   = cardinal_t<W>;
+    using ec_t  = expected_cardinal_t<v_t, typename W::abi_type>;
+
+    if constexpr(c_t::value < ec_t::value)
+    {
+      v = bit_cast( slide_right ( bit_cast(v,as<wide<v_t,ec_t>>())
+                                , index<ec_t::value - c_t::value>
+                                )
+                  , as(v)
+                  );
+    }
+
     return v;
   }
 }
diff --git a/include/eve/module/core/regular/impl/slide_left.hpp b/include/eve/module/core/regular/impl/slide_left.hpp
@@ -7,7 +7,6 @@
 //==================================================================================================
 #pragma once
 
-#include <eve/module/core/regular/slide_right.hpp>
 #include <eve/concept/vectorized.hpp>
 #include <eve/detail/abi.hpp>
 #include <eve/detail/meta.hpp>