Move x86 CPUID code from cpuid.hpp to cpuid.cpp (kimwalisch#150)

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.4...3.27)
 project(primesieve CXX)
-set(PRIMESIEVE_VERSION "12.3")
-set(PRIMESIEVE_SOVERSION "12.3.0")
+set(PRIMESIEVE_VERSION "12.4")
+set(PRIMESIEVE_SOVERSION "12.4.0")
 
 # Build options ######################################################
 
@@ -47,7 +47,7 @@ if(NOT isMultiConfig AND NOT CMAKE_BUILD_TYPE)
 endif()
 
 if(CMAKE_BUILD_TYPE STREQUAL "Debug")
-    set(ENABLE_ASSERT "ENABLE_ASSERT")
+    list(APPEND PRIMESIEVE_COMPILE_DEFINITIONS "ENABLE_ASSERT")
 endif()
 
 # primesieve binary source files #####################################
@@ -82,6 +82,17 @@ set(LIB_SRC src/api-c.cpp
             src/RiemannR.cpp
             src/SievingPrimes.cpp)
 
+# Check if compiler supports CPU multiarch ###########################
+
+if(WITH_MULTIARCH)
+    include("${PROJECT_SOURCE_DIR}/cmake/multiarch_x86_popcnt.cmake")
+    include("${PROJECT_SOURCE_DIR}/cmake/multiarch_avx512_vbmi2.cmake")
+
+    if(multiarch_x86_popcnt OR multiarch_avx512_vbmi2)
+        set(LIB_SRC ${LIB_SRC} src/x86/cpuid.cpp)
+    endif()
+endif()
+
 # Required includes ##################################################
 
 include(GNUInstallDirs)
@@ -107,26 +118,20 @@ if(WITH_AUTO_VECTORIZATION)
     include("${PROJECT_SOURCE_DIR}/cmake/auto_vectorization.cmake")
 endif()
 
-# Check if compiler supports x64 multiarch ###########################
-
-if(WITH_MULTIARCH)
-    include("${PROJECT_SOURCE_DIR}/cmake/multiarch_avx512_vbmi2.cmake")
-endif()
-
 # libprimesieve (shared library) #####################################
 
 find_package(Threads REQUIRED QUIET)
 
 if(BUILD_SHARED_LIBS)
     add_library(libprimesieve SHARED ${LIB_SRC})
     set_target_properties(libprimesieve PROPERTIES OUTPUT_NAME primesieve)
-    target_link_libraries(libprimesieve PRIVATE Threads::Threads ${LIBATOMIC})
+    target_link_libraries(libprimesieve PRIVATE Threads::Threads ${PRIMESIEVE_LINK_LIBRARIES})
     string(REPLACE "." ";" SOVERSION_LIST ${PRIMESIEVE_SOVERSION})
     list(GET SOVERSION_LIST 0 PRIMESIEVE_SOVERSION_MAJOR)
     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}" "${ENABLE_MULTIARCH_AVX512}")
+    target_compile_options(libprimesieve PRIVATE ${PRIMESIEVE_COMPILE_OPTIONS})
+    target_compile_definitions(libprimesieve PRIVATE ${PRIMESIEVE_COMPILE_DEFINITIONS})
 
     if(WIN32_MSVC_COMPATIBLE)
         # On Windows the shared library will be named primesieve.dll
@@ -162,9 +167,9 @@ endif()
 if(BUILD_STATIC_LIBS)
     add_library(libprimesieve-static STATIC ${LIB_SRC})
     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}" "${ENABLE_MULTIARCH_AVX512}")
+    target_link_libraries(libprimesieve-static PRIVATE Threads::Threads ${PRIMESIEVE_LINK_LIBRARIES})
+    target_compile_options(libprimesieve-static PRIVATE ${PRIMESIEVE_COMPILE_OPTIONS})
+    target_compile_definitions(libprimesieve-static PRIVATE ${PRIMESIEVE_COMPILE_DEFINITIONS})
 
     if(WITH_MSVC_CRT_STATIC)
         set_target_properties(libprimesieve-static PROPERTIES MSVC_RUNTIME_LIBRARY "MultiThreaded")
@@ -219,7 +224,7 @@ endif()
 if(BUILD_PRIMESIEVE)
     add_executable(primesieve ${BIN_SRC})
     target_link_libraries(primesieve primesieve::primesieve Threads::Threads)
-    target_compile_definitions(primesieve PRIVATE "${ENABLE_ASSERT}")
+    target_compile_definitions(primesieve PRIVATE ${PRIMESIEVE_COMPILE_DEFINITIONS})
     target_compile_features(primesieve PRIVATE cxx_auto_type)
     install(TARGETS primesieve DESTINATION ${CMAKE_INSTALL_BINDIR})
 

ChangeLog

@@ -1,3 +1,11 @@
+Changes in version 12.4, 22/06/2024
+===================================
+
+* Move x86 CPUID code from cpuid.hpp to src/x86/cpuid.cpp.
+* multiarch_x86_popcnt.cmake: Detect x86 POPCNT support.
+* CMakeLists.txt: Use CMake list for all compile time definitions.
+* CMakeLists.txt: Use CMake list for all link libraries.
+
 Changes in version 12.3, 15/04/2024
 ===================================
 

cmake/auto_vectorization.cmake

@@ -16,14 +16,14 @@ check_cxx_compiler_flag(-ftree-vectorize ftree_vectorize)
 cmake_pop_check_state()
 
 if(ftree_vectorize)
-    set(FTREE_VECTORIZE_FLAG "-ftree-vectorize")
+    list(APPEND PRIMESIEVE_COMPILE_OPTIONS "-ftree-vectorize")
 
     cmake_push_check_state()
     set(CMAKE_REQUIRED_FLAGS -Werror)
     check_cxx_compiler_flag(-fvect-cost-model=dynamic fvect_cost_model)
     cmake_pop_check_state()
 
     if(fvect_cost_model)
-        set(FVECT_COST_MODEL_FLAG "-fvect-cost-model=dynamic")
+        list(APPEND PRIMESIEVE_COMPILE_OPTIONS "-fvect-cost-model=dynamic")
     endif()
 endif()

cmake/libatomic.cmake

@@ -57,7 +57,10 @@ if(NOT atomic64)
         }"
         atomic64_with_libatomic)
 
-    if (NOT atomic64_with_libatomic)
+    if(atomic64_with_libatomic)
+        list(APPEND PRIMESIEVE_LINK_LIBRARIES "${LIBATOMIC}")
+    else()
+        set(LIBATOMIC "")
         message(FATAL_ERROR "Failed to compile std::atomic, libatomic likely not found!")
     endif()
 endif()

cmake/multiarch_avx512_vbmi2.cmake

@@ -7,7 +7,7 @@ include(CheckCXXSourceCompiles)
 include(CMakePushCheckState)
 
 cmake_push_check_state()
-set(CMAKE_REQUIRED_INCLUDES "${PROJECT_SOURCE_DIR}/include")
+set(CMAKE_REQUIRED_INCLUDES "${PROJECT_SOURCE_DIR}")
 
 check_cxx_source_compiles("
     // GCC/Clang function multiversioning for AVX512 is not needed if
@@ -20,19 +20,19 @@ check_cxx_source_compiles("
       Error: AVX512VBMI2 multiarch not needed!
     #endif
 
-    #include <primesieve/cpu_supports_avx512_vbmi2.hpp>
+    #include <src/x86/cpuid.cpp>
     #include <immintrin.h>
     #include <stdint.h>
 
     class PrimeGenerator {
     public:
         __attribute__ ((target (\"avx512f,avx512vbmi,avx512vbmi2\")))
-        void fillNextPrimes_avx512(uint64_t* primes64);
+        void fillNextPrimes_avx512_vbmi2(uint64_t* primes64);
         void fillNextPrimes_default(uint64_t* primes64);
         void fillNextPrimes(uint64_t* primes64)
         {
-            if (cpu_supports_avx512_vbmi2)
-                fillNextPrimes_avx512(primes64);
+            if (primesieve::has_cpuid_avx512_vbmi2())
+                fillNextPrimes_avx512_vbmi2(primes64);
             else
                 fillNextPrimes_default(primes64);
         }
@@ -44,7 +44,7 @@ check_cxx_source_compiles("
     }
 
     __attribute__ ((target (\"avx512f,avx512vbmi,avx512vbmi2\")))
-    void PrimeGenerator::fillNextPrimes_avx512(uint64_t* primes64)
+    void PrimeGenerator::fillNextPrimes_avx512_vbmi2(uint64_t* primes64)
     {
         __m512i bytes_0_to_7 = _mm512_setr_epi64(0, 1, 2, 3, 4, 5, 6, 7);
         __m512i base = _mm512_set1_epi64(123);
@@ -64,7 +64,7 @@ check_cxx_source_compiles("
 " multiarch_avx512_vbmi2)
 
 if(multiarch_avx512_vbmi2)
-    set(ENABLE_MULTIARCH_AVX512 "ENABLE_MULTIARCH_AVX512")
+    list(APPEND PRIMESIEVE_COMPILE_DEFINITIONS "ENABLE_MULTIARCH_AVX512_VBMI2")
 endif()
 
 cmake_pop_check_state()

cmake/multiarch_x86_popcnt.cmake

@@ -0,0 +1,53 @@
+# On x86 CPUs we need to enable the use of cpuid.cpp.
+# If cpuid.cpp compiles we assume it is a x86 CPU.
+
+include(CheckCXXSourceCompiles)
+include(CMakePushCheckState)
+
+cmake_push_check_state()
+set(CMAKE_REQUIRED_INCLUDES "${PROJECT_SOURCE_DIR}")
+
+check_cxx_source_compiles("
+    // Enable CPUID for POPCNT on x86 and x86-64 CPUs.
+    // This is required because not all x86 and x86-64 CPUs
+    // support the POPCNT instruction.
+    #if !(defined(__x86_64__) || \
+          defined(__i386__) || \
+          defined(_M_X64) || \
+          defined(_M_IX86))
+      Error: x86 POPCNT multiarch not needed!
+    #endif
+
+    // Both GCC and Clang (even Clang on Windows) define the __POPCNT__
+    // macro if the user compiles with -mpopcnt. The __POPCNT__
+    // macro is even defined if the user compiles with other flags
+    // such as -mavx or -march=native.
+    #if defined(__POPCNT__)
+        Error: x86 POPCNT multiarch not needed!
+
+    // The MSVC compiler does not support a POPCNT macro, but if the user
+    // compiles with e.g. /arch:AVX or /arch:AVX512 then MSVC defines
+    // the __AVX__ macro and POPCNT is also supported.
+    #elif defined(_MSC_VER) && defined(__AVX__)
+        Error: x86 POPCNT multiarch not needed!
+    #endif
+
+    #include <src/x86/cpuid.cpp>
+    #include <iostream>
+
+    int main()
+    {
+        if (primesieve::has_cpuid_popcnt())
+            std::cout << \"CPU supports POPCNT!\" << std::endl;
+        else
+            std::cout << \"CPU does not support POPCNT!\" << std::endl;
+
+        return 0;
+    }
+" multiarch_x86_popcnt)
+
+if(multiarch_x86_popcnt)
+    list(APPEND PRIMESIEVE_COMPILE_DEFINITIONS "ENABLE_MULTIARCH_x86_POPCNT")
+endif()
+
+cmake_pop_check_state()

include/primesieve.h

@@ -15,9 +15,9 @@
 #ifndef PRIMESIEVE_H
 #define PRIMESIEVE_H
 
-#define PRIMESIEVE_VERSION "12.3"
+#define PRIMESIEVE_VERSION "12.4"
 #define PRIMESIEVE_VERSION_MAJOR 12
-#define PRIMESIEVE_VERSION_MINOR 3
+#define PRIMESIEVE_VERSION_MINOR 4
 
 #include <primesieve/iterator.h>
 

include/primesieve.hpp

@@ -13,9 +13,9 @@
 #ifndef PRIMESIEVE_HPP
 #define PRIMESIEVE_HPP
 
-#define PRIMESIEVE_VERSION "12.3"
+#define PRIMESIEVE_VERSION "12.4"
 #define PRIMESIEVE_VERSION_MAJOR 12
-#define PRIMESIEVE_VERSION_MINOR 3
+#define PRIMESIEVE_VERSION_MINOR 4
 
 #include <primesieve/iterator.hpp>
 #include <primesieve/primesieve_error.hpp>

include/primesieve/CpuInfo.hpp

@@ -1,7 +1,7 @@
 ///
 /// @file  CpuInfo.hpp
 ///
-/// Copyright (C) 2023 Kim Walisch, <[email protected]>
+/// Copyright (C) 2024 Kim Walisch, <[email protected]>
 ///
 /// This file is distributed under the BSD License. See the COPYING
 /// file in the top level directory.
@@ -22,7 +22,6 @@ class CpuInfo
 public:
   CpuInfo();
   bool hasCpuName() const;
-  bool hasAVX512() const;
   bool hasLogicalCpuCores() const;
   bool hasL1Cache() const;
   bool hasL2Cache() const;

include/primesieve/Erat.hpp

@@ -1,7 +1,7 @@
 ///
 /// @file  Erat.hpp
 ///
-/// Copyright (C) 2023 Kim Walisch, <[email protected]>
+/// Copyright (C) 2024 Kim Walisch, <[email protected]>
 ///
 /// This file is distributed under the BSD License. See the COPYING
 /// file in the top level directory.
@@ -15,8 +15,8 @@
 #include "EratMedium.hpp"
 #include "EratBig.hpp"
 #include "macros.hpp"
-#include "intrinsics.hpp"
 #include "Vector.hpp"
+#include "ctz.hpp"
 
 #include <stdint.h>
 

include/primesieve/PrimeGenerator.hpp

@@ -27,9 +27,9 @@
     defined(__AVX512VBMI__) && \
     defined(__AVX512VBMI2__) && \
     __has_include(<immintrin.h>)
-  #define ENABLE_AVX512
+  #define ENABLE_AVX512_VBMI2
 
-#elif defined(ENABLE_MULTIARCH_AVX512) && \
+#elif defined(ENABLE_MULTIARCH_AVX512_VBMI2) && \
       __has_include(<immintrin.h>)
   #include "cpu_supports_avx512_vbmi2.hpp"
   #define ENABLE_DEFAULT
@@ -50,11 +50,11 @@ class PrimeGenerator : public Erat
 
   ALWAYS_INLINE void fillNextPrimes(Vector<uint64_t>& primes, std::size_t* size)
   {
-    #if defined(ENABLE_AVX512)
-      fillNextPrimes_avx512(primes, size);
-    #elif defined(ENABLE_MULTIARCH_AVX512)
+    #if defined(ENABLE_AVX512_VBMI2)
+      fillNextPrimes_avx512_vbmi2(primes, size);
+    #elif defined(ENABLE_MULTIARCH_AVX512_VBMI2)
       if (cpu_supports_avx512_vbmi2)
-        fillNextPrimes_avx512(primes, size);
+        fillNextPrimes_avx512_vbmi2(primes, size);
       else
         fillNextPrimes_default(primes, size);
     #else
@@ -68,13 +68,13 @@ class PrimeGenerator : public Erat
   void fillNextPrimes_default(Vector<uint64_t>& primes, std::size_t* size);
 #endif
 
-#if defined(ENABLE_AVX512) || \
-    defined(ENABLE_MULTIARCH_AVX512)
+#if defined(ENABLE_AVX512_VBMI2) || \
+    defined(ENABLE_MULTIARCH_AVX512_VBMI2)
 
-  #if defined(ENABLE_MULTIARCH_AVX512)
+  #if defined(ENABLE_MULTIARCH_AVX512_VBMI2)
     __attribute__ ((target ("avx512f,avx512vbmi,avx512vbmi2")))
   #endif
-  void fillNextPrimes_avx512(Vector<uint64_t>& primes, std::size_t* size);
+  void fillNextPrimes_avx512_vbmi2(Vector<uint64_t>& primes, std::size_t* size);
 
 #endif