move by pieces

LoopModels · Dec 4, 2023 · 359d087 · 359d087
1 parent 9d81388
commit 359d087
Show file tree

Hide file tree

Showing 2 changed files with 13 additions and 12 deletions.
diff --git a/benchmark/CMakeLists.txt b/benchmark/CMakeLists.txt
@@ -71,6 +71,11 @@ if(ENABLE_NATIVE_COMPILATION)
       check_cxx_compiler_flag("-mprefer-vector-width=512" VEC512)
       if(VEC512)
         target_compile_options(${PROJECT_NAME} PRIVATE -mprefer-vector-width=512)
+        if (CMAKE_CXX_COMPILER_ID MATCHES "GNU")
+          if (NOT CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+            target_compile_options(${PROJECT_NAME} PRIVATE -mtune-ctrl=avx512_move_by_pieces)
+          endif()
+        endif()
       endif()
     endif()
   endif()

diff --git a/benchmark/dual_benchmark.cpp b/benchmark/dual_benchmark.cpp
@@ -70,8 +70,8 @@ template <std::floating_point T, ptrdiff_t N> struct ManualDual<T, N, true> {
   auto grad() -> P & { return partials; }
 };
 template <typename T, ptrdiff_t N, bool B>
-[[gnu::always_inline]] constexpr auto operator*(const ManualDual<T, N, B> &a,
-                                                const ManualDual<T, N, B> &b)
+[[gnu::always_inline]] constexpr auto operator*(ManualDual<T, N, B> a,
+                                                ManualDual<T, N, B> b)
   -> ManualDual<T, N, B> {
   if constexpr ((!B) && (!std::floating_point<T>)&&(N == 2))
     return {a.value * b.value,
@@ -80,32 +80,28 @@ template <typename T, ptrdiff_t N, bool B>
   else return {a.value * b.value, a.value * b.partials + b.value * a.partials};
 }
 template <typename T, ptrdiff_t N, bool B>
-[[gnu::always_inline]] constexpr auto operator*(const ManualDual<T, N, B> &a,
-                                                const T &b)
+[[gnu::always_inline]] constexpr auto operator*(ManualDual<T, N, B> a, T b)
   -> ManualDual<T, N, B> {
   return {a.value * b, b * a.partials};
 }
 template <typename T, ptrdiff_t N, bool B>
-[[gnu::always_inline]] constexpr auto operator*(const T &a,
-                                                const ManualDual<T, N, B> &b)
+[[gnu::always_inline]] constexpr auto operator*(T a, ManualDual<T, N, B> b)
   -> ManualDual<T, N, B> {
   return {b.value * a, a * b.partials};
 }
 template <typename T, ptrdiff_t N, bool B>
-[[gnu::always_inline]] constexpr auto operator+(const ManualDual<T, N, B> &a,
-                                                const ManualDual<T, N, B> &b)
+[[gnu::always_inline]] constexpr auto operator+(ManualDual<T, N, B> a,
+                                                ManualDual<T, N, B> b)
   -> ManualDual<T, N, B> {
   return {a.value + b.value, a.partials + b.partials};
 }
 template <typename T, ptrdiff_t N, bool B>
-[[gnu::always_inline]] constexpr auto operator+(const ManualDual<T, N, B> &a,
-                                                const T &b)
+[[gnu::always_inline]] constexpr auto operator+(ManualDual<T, N, B> a, T b)
   -> ManualDual<T, N, B> {
   return {a.value + b, a.partials};
 }
 template <typename T, ptrdiff_t N, bool B>
-[[gnu::always_inline]] constexpr auto operator+(const T &a,
-                                                const ManualDual<T, N, B> &b)
+[[gnu::always_inline]] constexpr auto operator+(T a, ManualDual<T, N, B> b)
   -> ManualDual<T, N, B> {
   return {b.value + a, b.partials};
 }