Remove amrex::clz (#5539)

Author: WeiqunZhang

Src/Base/AMReX_Algorithm.H

@@ -387,155 +387,6 @@ namespace amrex
             *first = std::pow(base, stop);
         }
     }
-
-/// \cond DOXYGEN_IGNORE
-namespace detail {
-
-struct clzll_tag {};
-struct clzl_tag : clzll_tag {};
-struct clz_tag : clzl_tag {};
-
-// in gcc and clang, there are three versions of __builtin_clz taking unsigned int,
-// unsigned long, and unsigned long long inputs. Because the sizes of these data types
-// vary on different platforms, we work with fixed-width integer types.
-// these tags and overloads select the smallest version of __builtin_clz that will hold the input type
-template <typename T>
-requires (sizeof(T) <= sizeof(unsigned int))
-AMREX_FORCE_INLINE
-int builtin_clz_wrapper (clz_tag, T x) noexcept
-{
-    return static_cast<int>(__builtin_clz(x) - (sizeof(unsigned int) * CHAR_BIT - sizeof(T) * CHAR_BIT));
-}
-
-template <typename T>
-requires (sizeof(T) <= sizeof(unsigned long))
-AMREX_FORCE_INLINE
-int builtin_clz_wrapper (clzl_tag, T x) noexcept
-{
-    return static_cast<int>(__builtin_clzl(x) - (sizeof(unsigned long) * CHAR_BIT - sizeof(T) * CHAR_BIT));
-}
-
-template <typename T>
-requires (sizeof(T) <= sizeof(unsigned long long))
-AMREX_FORCE_INLINE
-int builtin_clz_wrapper (clzll_tag, T x) noexcept
-{
-    return static_cast<int>(__builtin_clzll(x) - (sizeof(unsigned long long) * CHAR_BIT - sizeof(T) * CHAR_BIT));
-}
-
-}
-/// \endcond
-
-/// \ingroup amrex_utilities
-//! Return the number of leading zeros present in the unsigned integer \p x.
-//! \pre \p x must be nonzero.
-template <class T>
-requires (std::same_as<std::remove_cvref_t<T>,std::uint8_t>  ||
-          std::same_as<std::remove_cvref_t<T>,std::uint16_t> ||
-          std::same_as<std::remove_cvref_t<T>,std::uint32_t> ||
-          std::same_as<std::remove_cvref_t<T>,std::uint64_t>)
-AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
-int clz (T x) noexcept;
-
-/// \cond DOXYGEN_IGNORE
-
-AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
-int clz_generic (std::uint8_t x) noexcept
-{
-#if !defined(__NVCOMPILER)
-    static constexpr int clz_lookup[16] = { 4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 };
-#else
-    constexpr int clz_lookup[16] = { 4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 };
-#endif
-    auto upper = x >> 4;
-    auto lower = x & 0xF;
-    return upper ? clz_lookup[upper] : 4 + clz_lookup[lower];
-}
-
-AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
-int clz_generic (std::uint16_t x) noexcept
-{
-    auto upper = std::uint8_t(x >> 8);
-    auto lower = std::uint8_t(x & 0xFF);
-    return upper ? clz(upper) : 8 + clz(lower);
-}
-
-AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
-int clz_generic (std::uint32_t x) noexcept
-{
-    auto upper = std::uint16_t(x >> 16);
-    auto lower = std::uint16_t(x & 0xFFFF);
-    return upper ? clz(upper) : 16 + clz(lower);
-}
-
-AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
-int clz_generic (std::uint64_t x) noexcept
-{
-    auto upper = std::uint32_t(x >> 32);
-    auto lower = std::uint32_t(x & 0xFFFFFFFF);
-    return upper ? clz(upper) : 32 + clz(lower);
-}
-
-/// \endcond
-
-#if defined AMREX_USE_CUDA
-
-/// \cond DOXYGEN_IGNORE
-namespace detail {
-    // likewise with CUDA, there are __clz functions that take (signed) int and long long int
-    template <typename T>
-    requires (sizeof(T) <= sizeof(int))
-    AMREX_GPU_DEVICE AMREX_FORCE_INLINE
-    int clz_wrapper (clz_tag, T x) noexcept
-    {
-        return __clz((int) x) - (sizeof(int) * CHAR_BIT - sizeof(T) * CHAR_BIT);
-    }
-
-    template <typename T>
-    requires (sizeof(T) <= sizeof(long long int))
-    AMREX_GPU_DEVICE AMREX_FORCE_INLINE
-    int clz_wrapper (clzll_tag, T x) noexcept
-    {
-        return __clzll((long long int) x) - (sizeof(long long int) * CHAR_BIT - sizeof(T) * CHAR_BIT);
-    }
-}
-/// \endcond
-
-template <class T>
-requires (std::same_as<std::remove_cvref_t<T>,std::uint8_t>  ||
-          std::same_as<std::remove_cvref_t<T>,std::uint16_t> ||
-          std::same_as<std::remove_cvref_t<T>,std::uint32_t> ||
-          std::same_as<std::remove_cvref_t<T>,std::uint64_t>)
-AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
-int clz (T x) noexcept
-{
-    AMREX_IF_ON_DEVICE((return detail::clz_wrapper(detail::clz_tag{}, x);))
-#if AMREX_HAS_BUILTIN_CLZ
-    AMREX_IF_ON_HOST((return detail::builtin_clz_wrapper(detail::clz_tag{}, x);))
-#else
-    AMREX_IF_ON_HOST((return clz_generic(x);))
-#endif
-}
-
-#else // !defined AMREX_USE_CUDA
-
-template <class T>
-requires (std::same_as<std::remove_cvref_t<T>,std::uint8_t>  ||
-          std::same_as<std::remove_cvref_t<T>,std::uint16_t> ||
-          std::same_as<std::remove_cvref_t<T>,std::uint32_t> ||
-          std::same_as<std::remove_cvref_t<T>,std::uint64_t>)
-AMREX_GPU_HOST_DEVICE AMREX_FORCE_INLINE
-int clz (T x) noexcept
-{
-#if (!AMREX_DEVICE_COMPILE && AMREX_HAS_BUILTIN_CLZ)
-    return detail::builtin_clz_wrapper(detail::clz_tag{}, x);
-#else
-    return clz_generic(x);
-#endif
-}
-
-#endif // defined AMREX_USE_CUDA
-
 }
 
 #endif

Src/Base/AMReX_Extension.H

@@ -270,13 +270,6 @@
 #   define AMREX_IF_CONSTEXPR if
 #endif
 
-#if !defined(AMREX_NO_BUILTIN_CLZ)
-#   if defined(__clang__) || defined(__GNUC__)
-#      define AMREX_HAS_BUILTIN_CLZ 1
-#   endif
-#endif
-
-
 #endif /* !BL_LANG_FORT */
 
 #endif

Tests/CLZ/CMakeLists.txt

@@ -133,7 +133,7 @@ else()
    #
    # List of subdirectories to search for CMakeLists.
    #
-   set( AMREX_TESTS_SUBDIRS Amr ArrayND AsyncOut Base CallNoinline CLZ CommType CTOParFor DeviceGlobal
+   set( AMREX_TESTS_SUBDIRS Amr ArrayND AsyncOut Base CallNoinline CommType CTOParFor DeviceGlobal
                             Enum GpuParallelReduce HeatEquation MultiBlock MultiPeriod ParmParse Parser Parser2
                             ParserUserFn Reducer ReduceToPlaneGuards ReduceToPlanePatchy Reinit RoundoffDomain SIMD
                             SmallMatrix SumBoundary TOML)