cmakelist update

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.0)
 
-project(training_data_loader)
+project(training_data_loader CXX)
 
 if(NOT DEFINED CMAKE_BUILD_TYPE)
     set(CMAKE_BUILD_TYPE RelWithDebInfo)
@@ -13,6 +13,44 @@ set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-g -O3 -march=native -DNDEBUG")
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED 17)
 
+include(CheckCXXCompilerFlag)
+
+# Function to check if the CPU is AMD and not older than family 25
+function(check_amd_family_25_or_newer)
+    execute_process(
+        COMMAND awk "/^vendor_id/{{print \$3; exit}}" /proc/cpuinfo
+        OUTPUT_VARIABLE VENDOR_ID
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    execute_process(
+        COMMAND awk "/^cpu family/{{print \$4; exit}}" /proc/cpuinfo
+        OUTPUT_VARIABLE CPU_FAMILY
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+
+    if(VENDOR_ID STREQUAL "AuthenticAMD" AND CPU_FAMILY GREATER_EQUAL "25")
+        set(CPU_AMD_FAMILY_25_OR_NEWER TRUE PARENT_SCOPE)
+    else()
+        set(CPU_AMD_FAMILY_25_OR_NEWER FALSE PARENT_SCOPE)
+    endif()
+endfunction()
+
+# Call the function to check if the CPU is AMD and not older than family 25
+check_amd_family_25_or_newer()
+
+# Check for BMI2 support
+check_cxx_compiler_flag("-mbmi2" COMPILER_SUPPORTS_BMI2)
+if(COMPILER_SUPPORTS_BMI2)
+    if(CPU_AMD_FAMILY_25_OR_NEWER OR NOT VENDOR_ID STREQUAL "AuthenticAMD")
+        message(STATUS "Adding BMI2 support")
+        add_definitions(-DHAS_BMI2)
+    else()
+        message(STATUS "CPU is older than AMD family 25, not adding BMI2 support")
+    endif()
+else()
+    message(FATAL_ERROR "Compiler does not support BMI2")
+endif()
+
 add_library(training_data_loader SHARED training_data_loader.cpp)
 
 find_package(Threads REQUIRED)

lib/nnue_training_data_formats.h

@@ -48,7 +48,7 @@ THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #include <mutex>
 #include <random>
 
-#ifdef __BMI2__
+#ifdef HAS_BMI2
 #include <immintrin.h> // _pdep_u64
 #endif
 
@@ -267,7 +267,7 @@ namespace chess
 
     inline int nthSetBitIndex(std::uint64_t v, std::uint64_t n)
     {
-    #ifdef __BMI2__
+    #ifdef HAS_BMI2
         return intrin::msb(_pdep_u64(1ULL << n, v));
     #endif