Simplify preprocessor logic

kimwalisch · Apr 7, 2024 · 2097a56 · 2097a56
1 parent 55cdcce
commit 2097a56
Show file tree

Hide file tree

Showing 5 changed files with 49 additions and 44 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -112,7 +112,7 @@ endif()
 if(WITH_MULTIARCH)
     include("${PROJECT_SOURCE_DIR}/cmake/multiarch_avx512_vbmi2.cmake")
     if(multiarch_avx512_vbmi2)
-        set(MULTIARCH_AVX512 "MULTIARCH_AVX512")
+        set(ENABLE_MULTIARCH_AVX512 "ENABLE_MULTIARCH_AVX512")
     endif()
 endif()
 
@@ -129,7 +129,7 @@ if(BUILD_SHARED_LIBS)
     set_target_properties(libprimesieve PROPERTIES SOVERSION ${PRIMESIEVE_SOVERSION_MAJOR})
     set_target_properties(libprimesieve PROPERTIES VERSION ${PRIMESIEVE_SOVERSION})
     target_compile_options(libprimesieve PRIVATE ${FTREE_VECTORIZE_FLAG} ${FVECT_COST_MODEL_FLAG})
-    target_compile_definitions(libprimesieve PRIVATE "${ENABLE_ASSERT}" "${MULTIARCH_AVX512}")
+    target_compile_definitions(libprimesieve PRIVATE "${ENABLE_ASSERT}" "${ENABLE_MULTIARCH_AVX512}")
 
     if(WIN32_MSVC_COMPATIBLE)
         # On Windows the shared library will be named primesieve.dll
@@ -167,7 +167,7 @@ if(BUILD_STATIC_LIBS)
     set_target_properties(libprimesieve-static PROPERTIES OUTPUT_NAME primesieve)
     target_link_libraries(libprimesieve-static PRIVATE Threads::Threads ${LIBATOMIC})
     target_compile_options(libprimesieve-static PRIVATE ${FTREE_VECTORIZE_FLAG} ${FVECT_COST_MODEL_FLAG})
-    target_compile_definitions(libprimesieve-static PRIVATE "${ENABLE_ASSERT}" "${MULTIARCH_AVX512}")
+    target_compile_definitions(libprimesieve-static PRIVATE "${ENABLE_ASSERT}" "${ENABLE_MULTIARCH_AVX512}")
 
     if(WITH_MSVC_CRT_STATIC)
         set_target_properties(libprimesieve-static PROPERTIES MSVC_RUNTIME_LIBRARY "MultiThreaded")

diff --git a/cmake/multiarch_avx512_vbmi2.cmake b/cmake/multiarch_avx512_vbmi2.cmake
@@ -1,24 +1,38 @@
-include(CheckCXXSourceCompiles)
-
 # We use GCC/Clang's function multi-versioning for AVX512
 # support. This code will automatically dispatch to the
 # AVX512 algorithm if the CPU supports AVX512 and use the
 # default (portable) algorithm otherwise.
+
+include(CheckCXXSourceCompiles)
+
 check_cxx_source_compiles("
     #include <immintrin.h>
     #include <stdint.h>
+
+    // GCC/Clang function multiversioning for AVX512 is not needed if
+    // the user compiles with -mavx512f -mavx512vbmi -mavx512vbmi2.
+    // GCC/Clang function multiversioning generally causes a minor
+    // overhead, hence we disable it if it is not needed.
+    #if defined(__AVX512F__) && \
+        defined(__AVX512VBMI__) && \
+        defined(__AVX512VBMI2__)
+      Error: AVX512VBMI2 multiarch not needed!
+    #endif
+
     class PrimeGenerator {
         public:
         __attribute__ ((target (\"default\")))
         void fillNextPrimes(uint64_t* primes64);
         __attribute__ ((target (\"avx512f,avx512vbmi,avx512vbmi2\")))
         void fillNextPrimes(uint64_t* primes64);
     };
+
     __attribute__ ((target (\"default\")))
     void PrimeGenerator::fillNextPrimes(uint64_t* primes64)
     {
         primes64[0] = 2;
     }
+
     __attribute__ ((target (\"avx512f,avx512vbmi,avx512vbmi2\")))
     void PrimeGenerator::fillNextPrimes(uint64_t* primes64)
     {
@@ -29,6 +43,7 @@ check_cxx_source_compiles("
         vprimes = _mm512_add_epi64(base, vprimes);
         _mm512_storeu_si512(primes64, vprimes);
     }
+
     int main()
     {
         uint64_t primes[8];

diff --git a/include/primesieve/PrimeGenerator.hpp b/include/primesieve/PrimeGenerator.hpp
@@ -23,21 +23,6 @@
 #include <stdint.h>
 #include <cstddef>
 
-#if defined(MULTIARCH_AVX512)
-  // GCC/Clang function multiversioning for AVX512 is not needed if
-  // the user compiles with -mavx512f -mavx512vbmi -mavx512vbmi2.
-  // GCC/Clang function multiversioning generally causes a minor
-  // overhead, hence we disable it if it is not needed.
-  #if defined(__AVX512__) || (defined(__AVX512F__) && \
-                              defined(__AVX512VBMI__) && \
-                              defined(__AVX512VBMI2__))
-    #undef MULTIARCH_AVX512
-  #else
-    #define MULTIARCH_TARGET_DEFAULT
-    #define MULTIARCH_TARGET_AVX512
-  #endif
-#endif
-
 namespace primesieve {
 
 class PreSieve;
@@ -49,16 +34,17 @@ class PrimeGenerator : public Erat
   void fillPrevPrimes(Vector<uint64_t>& primes, std::size_t* size);
   static uint64_t maxCachedPrime();
 
-#if defined(MULTIARCH_TARGET_DEFAULT)
-  __attribute__ ((target ("default")))
-#endif
-  void fillNextPrimes(Vector<uint64_t>& primes, std::size_t* size);
-
-#if defined(MULTIARCH_TARGET_AVX512)
+#if defined(ENABLE_MULTIARCH_AVX512)
+  #define ENABLE_MULTIARCH_DEFAULT
   __attribute__ ((target ("avx512f,avx512vbmi,avx512vbmi2")))
   void fillNextPrimes(Vector<uint64_t>& primes, std::size_t* size);
 #endif
 
+#if defined(ENABLE_MULTIARCH_DEFAULT)
+  __attribute__ ((target ("default")))
+#endif
+  void fillNextPrimes(Vector<uint64_t>& primes, std::size_t* size);
+
 private:
   bool isInit_ = false;
   uint64_t low_ = 0;

diff --git a/src/PrimeGenerator.cpp b/src/PrimeGenerator.cpp
@@ -35,21 +35,25 @@
 #include <algorithm>
 #include <limits>
 
-// x86-64 AVX512
-#if __has_include(<immintrin.h>) && \
-   (defined(__AVX512__) || (defined(__AVX512F__) && \
-                            defined(__AVX512VBMI__) && \
-                            defined(__AVX512VBMI2__)))
+#if defined(ENABLE_MULTIARCH_AVX512) && \
+    __has_include(<immintrin.h>)
   #include <immintrin.h>
-  #define HAS_AVX512_VBMI2
 
-// GCC/Clang function multiversioning
-#elif defined(MULTIARCH_TARGET_AVX512) && \
-    __has_include(<immintrin.h>)
+#elif defined(__AVX512F__) && \
+      defined(__AVX512VBMI__) && \
+      defined(__AVX512VBMI2__) && \
+      __has_include(<immintrin.h>)
+  #include <immintrin.h>
+  #define ENABLE_AVX512
+
+#elif defined(_MSC_VER) && \
+      defined(__AVX512__) && \
+      __has_include(<immintrin.h>)
   #include <immintrin.h>
+  #define ENABLE_AVX512
 
-#else // Default portable algorithm
-  #define DEFAULT_CPU_ARCH
+#else
+  #define ENABLE_DEFAULT
 #endif
 
 namespace {
@@ -406,8 +410,8 @@ void PrimeGenerator::fillPrevPrimes(Vector<uint64_t>& primes,
   }
 }
 
-#if defined(DEFAULT_CPU_ARCH) || \
-    defined(MULTIARCH_TARGET_DEFAULT)
+#if defined(ENABLE_DEFAULT) || \
+    defined(ENABLE_MULTIARCH_DEFAULT)
 
 /// This method is used by iterator::next_prime().
 /// This method stores only the next few primes (~ 1000) in the
@@ -416,7 +420,7 @@ void PrimeGenerator::fillPrevPrimes(Vector<uint64_t>& primes,
 /// this reason iterator::next_prime() runs up to 2x faster
 /// than iterator::prev_prime().
 ///
-#if defined(MULTIARCH_TARGET_DEFAULT)
+#if defined(ENABLE_MULTIARCH_DEFAULT)
   __attribute__ ((target ("default")))
 #endif
 void PrimeGenerator::fillNextPrimes(Vector<uint64_t>& primes,
@@ -475,8 +479,8 @@ void PrimeGenerator::fillNextPrimes(Vector<uint64_t>& primes,
 
 #endif
 
-#if defined(HAS_AVX512_VBMI2) || \
-    defined(MULTIARCH_TARGET_AVX512)
+#if defined(ENABLE_AVX512) || \
+    defined(ENABLE_MULTIARCH_AVX512)
 
 /// This algorithm converts 1 bits from the sieve array into primes
 /// using AVX512. The algorithm is a modified version of the AVX512
@@ -491,7 +495,7 @@ void PrimeGenerator::fillNextPrimes(Vector<uint64_t>& primes,
 /// benchmarks this algorithm ran about 10% faster than the default
 /// fillNextPrimes() algorithm which uses __builtin_ctzll().
 ///
-#if defined(MULTIARCH_TARGET_AVX512)
+#if defined(ENABLE_MULTIARCH_AVX512)
   __attribute__ ((target ("avx512f,avx512vbmi,avx512vbmi2")))
 #endif
 void PrimeGenerator::fillNextPrimes(Vector<uint64_t>& primes,

diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
@@ -5,6 +5,6 @@ foreach(file ${files})
     get_filename_component(binary_name ${file} NAME_WE)
     add_executable(${binary_name} ${file})
     target_link_libraries(${binary_name} primesieve::primesieve)
-    target_compile_definitions(${binary_name} PRIVATE "${ENABLE_ASSERT}" "${MULTIARCH_AVX512}")
+    target_compile_definitions(${binary_name} PRIVATE "${ENABLE_ASSERT}" "${ENABLE_MULTIARCH_AVX512}")
     add_test(NAME ${binary_name} COMMAND ${binary_name})
 endforeach()