GPU Standalone: unify duplicate denormal code in standalone benchmark

Author: davidrohr

GPU/GPUTracking/Standalone/Benchmark/standalone.cxx

@@ -53,9 +53,6 @@
 #include "GPUTPCGMMergedTrack.h"
 #include "GPUSettings.h"
 #include <vector>
-#if not(defined(__ARM_NEON) or defined(__aarch64__)) // ARM doesn't have SSE
-#include <xmmintrin.h>
-#endif
 
 #include "GPUO2DataTypes.h"
 #include "GPUChainITS.h"
@@ -85,23 +82,6 @@ std::atomic<uint32_t> nIteration, nIterationEnd;
 std::vector<GPUTrackingInOutPointers> ioPtrEvents;
 std::vector<GPUChainTracking::InOutMemory> ioMemEvents;
 
-void SetCPUAndOSSettings()
-{
-#if not(defined(__ARM_NEON) or defined(__aarch64__)) // ARM doesn't have SSE
-#ifdef FE_DFL_DISABLE_SSE_DENORMS_ENV                // Flush and load denormals to zero in any case
-  fesetenv(FE_DFL_DISABLE_SSE_DENORMS_ENV);
-#else
-#ifndef _MM_FLUSH_ZERO_ON
-#define _MM_FLUSH_ZERO_ON 0x8000
-#endif
-#ifndef _MM_DENORMALS_ZERO_ON
-#define _MM_DENORMALS_ZERO_ON 0x0040
-#endif
-  _mm_setcsr(_mm_getcsr() | (_MM_FLUSH_ZERO_ON | _MM_DENORMALS_ZERO_ON));
-#endif
-#endif // ARM
-}
-
 int32_t ReadConfiguration(int argc, char** argv)
 {
   int32_t qcRet = qConfigParse(argc, (const char**)argv);
@@ -740,8 +720,6 @@ int32_t main(int argc, char** argv)
 {
   std::unique_ptr<GPUReconstruction> recUnique, recUniqueAsync, recUniquePipeline;
 
-  SetCPUAndOSSettings();
-
   if (ReadConfiguration(argc, argv)) {
     return 1;
   }

GPU/GPUTracking/utils/qmaths_helpers.h

@@ -15,24 +15,31 @@
 #ifndef QMATH_HELPERS_H
 #define QMATH_HELPERS_H
 
-#if defined __has_include
-#if __has_include(<xmmintrin.h>) && __has_include(<pmmintrin.h>)
-#include <xmmintrin.h>
-#include <pmmintrin.h>
-#if defined(_MM_FLUSH_ZERO_OFF) && defined(_MM_DENORMALS_ZERO_ON)
 static void disable_denormals()
 {
-  _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
-  _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
+#if !(defined(__ARM_NEON) || defined(__aarch64__)) && __has_include(<xmmintrin.h>) // clang-format off
+   #include <xmmintrin.h>
+   #if __has_include(<pmmintrin.h>)
+     #include <pmmintrin.h>
+   #endif
+   #if defined(_MM_FLUSH_ZERO_OFF) && defined(_MM_DENORMALS_ZERO_ON)
+     _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
+     _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
+   #else
+     #ifndef _MM_FLUSH_ZERO_ON
+       #define _MM_FLUSH_ZERO_ON 0x8000
+     #endif
+     #ifndef _MM_DENORMALS_ZERO_ON
+       #define _MM_DENORMALS_ZERO_ON 0x0040
+     #endif
+     _mm_setcsr(_mm_getcsr() | (_MM_FLUSH_ZERO_ON | _MM_DENORMALS_ZERO_ON));
+   #endif
+#elif __has_include(<cfenv>)
+  #include <cfenv>
+  #ifdef FE_DFL_DISABLE_SSE_DENORMS_ENV
+    fesetenv(FE_DFL_DISABLE_SSE_DENORMS_ENV);
+  #endif
+#endif // clang-format on
 }
-#define XMM_HAS_DENORMAL_DEACTIVATE
-#endif
-#endif
-#endif
-#ifdef XMM_HAS_DENORMAL_DEACTIVATE
-#undef XMM_HAS_DENORMAL_DEACTIVATE
-#else
-static void disable_denormals() {}
-#endif
 
 #endif