[fix-arm-support] CODEGEN: use builtin __ARM_NEON for aarch64 simd (with DanieleM)

Remove the custom __ARM_NEON__ with two extra underscores
Use 'g++ -march=armv8.2-a+simd -E -dM - < /dev/null | grep ARM' to check

Author: valassi

epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/madgraph/iolibs/template_files/gpu/MatrixElementKernels.cc

@@ -260,9 +260,8 @@ namespace mg5amcCpu
     bool ok = true;     // this is just an assumption!
     const std::string tag = "simd arch not defined";
 #endif
-#elif defined __ARM_NEON__ // consider using __BUILTIN_CPU_SUPPORTS__
+#elif defined __ARM_NEON // consider using __BUILTIN_CPU_SUPPORTS__
     bool known = false; // __builtin_cpu_supports is not supported
-    // See https://gcc.gnu.org/onlinedocs/gcc/Basic-PowerPC-Built-in-Functions-Available-on-all-Configurations.html
     // See https://stackoverflow.com/q/62783908
     // See https://community.arm.com/arm-community-blogs/b/operating-systems-blog/posts/runtime-detection-of-cpu-features-on-an-armv8-a-cpu
     bool ok = true; // this is just an assumption!

epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/madgraph/iolibs/template_files/gpu/check_sa.cc

@@ -915,7 +915,7 @@ main( int argc, char** argv )
 #else
   wrkflwtxt += "/sse4";
 #endif
-#elif defined __ARM_NEON__
+#elif defined __ARM_NEON
   wrkflwtxt += "/neon";
 #else
   wrkflwtxt += "/????"; // no path to this statement
@@ -1031,7 +1031,7 @@ main( int argc, char** argv )
 #else
               << "] ('sse4': SSE4.2, 128bit)" << cxtref << std::endl
 #endif
-#elif defined __ARM_NEON__
+#elif defined __ARM_NEON
               << "Internal loops fptype_sv    = VECTOR[" << neppV
               << "] ('sse4': ARM NEON, 128bit)" << cxtref << std::endl
 #else

epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/madgraph/iolibs/template_files/gpu/cudacpp.mk

@@ -516,6 +516,7 @@ CXXFLAGS += $(OMPFLAGS)
 # Set the build flags appropriate to each BACKEND choice (example: "make BACKEND=cppnone")
 # [NB MGONGPU_PVW512 is needed because "-mprefer-vector-width=256" is not exposed in a macro]
 # [See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96476]
+# [Use 'g++ <buildflags> -E -dM - < /dev/null' to check which #define's are enabled]
 ifeq ($(UNAME_P),ppc64le)
   ifeq ($(BACKEND),cppsse4)
     override AVXFLAGS = -D__SSE4_2__ # Power9 VSX with 128 width (VSR registers)
@@ -537,10 +538,10 @@ else ifeq ($(UNAME_P),arm)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on ARM for the moment)
   endif
 else ifeq ($(UNAME_P),aarch64)
-  ifeq ($(BACKEND),cppnone)
+  ifeq ($(BACKEND),cppnone) # +nosimd ensures __ARM_NEON is absent
     override AVXFLAGS = -march=armv8-a+nosimd
-  else ifeq ($(BACKEND),cppsse4)
-    override AVXFLAGS = -march=armv8-a+simd -D__ARM_NEON__
+  else ifeq ($(BACKEND),cppsse4) # +simd ensures __ARM_NEON is present (128 width Q/quadword registers)
+    override AVXFLAGS = -march=armv8-a+simd
   else ifeq ($(BACKEND),cppavx2)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on aarch64 for the moment)
   else ifeq ($(BACKEND),cpp512y)

epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/madgraph/iolibs/template_files/gpu/mgOnGpuConfig.h

@@ -241,7 +241,7 @@ using mgOnGpu::fptype2;
 #else
 #define MGONGPU_CPPSIMD 4
 #endif
-#elif defined __ARM_NEON__ // C++ "sse4" ARM NEON (128-bit ie 16-byte): 2 (DOUBLE) or 4 (FLOAT) [ARM default]
+#elif defined __ARM_NEON // C++ "sse4" ARM NEON (128-bit ie 16-byte): 2 (DOUBLE) or 4 (FLOAT) [ARM default]
 #ifdef MGONGPU_FPTYPE_DOUBLE
 #define MGONGPU_CPPSIMD 2
 #else