Fix arm support for Linux

.github/workflows/c-cpp.yml

@@ -41,6 +41,24 @@ jobs:
       run: make BACKEND=${{ matrix.backend }} FPTYPE=${{ matrix.precision }} -C ${{ matrix.folder }}
     - name: make test
       run: make BACKEND=${{ matrix.backend }} FPTYPE=${{ matrix.precision }} -C ${{ matrix.folder }} -f cudacpp.mk test
+  CPU_ARM:
+    runs-on: ubuntu-24.04-arm
+    strategy:
+      matrix:
+        folder: [ epochX/cudacpp/ee_mumu.mad/SubProcesses/P1_epem_mupmum , epochX/cudacpp/gg_ttgg.mad/SubProcesses/P1_gg_ttxgg ]
+        precision: [ d , f , m ]
+        backend: [ cppnone, cppsse4 ]
+      fail-fast: false
+    steps:
+    - uses: actions/checkout@v2
+    - name: github PR info
+      run: date; echo github.event.pull_request.head.sha='${{ github.event.pull_request.head.sha }}'
+    - name: make info
+      run: make BACKEND=${{ matrix.backend }} FPTYPE=${{ matrix.precision }} -C ${{ matrix.folder }} -f cudacpp.mk info
+    - name: make
+      run: make BACKEND=${{ matrix.backend }} FPTYPE=${{ matrix.precision }} -C ${{ matrix.folder }}
+    - name: make test
+      run: make BACKEND=${{ matrix.backend }} FPTYPE=${{ matrix.precision }} -C ${{ matrix.folder }} -f cudacpp.mk test
   CPU_MAC:
     runs-on: macos-latest
     env:

epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/MadtRex/makefiles/cudacpp_driver.mk

@@ -41,6 +41,7 @@ UNAME_S := $(shell uname -s)
 # Detect architecture (x86_64, ppc64le...)
 UNAME_P := $(shell uname -p)
 ###$(info UNAME_P='$(UNAME_P)')
+UNAME_M := $(shell uname -m)
 
 #-------------------------------------------------------------------------------
 
@@ -60,7 +61,7 @@ endif
 ifeq ($(BACKEND),cppauto)
   ifeq ($(UNAME_P),ppc64le)
     override BACKEND = cppsse4
-  else ifeq ($(UNAME_P),arm)
+  else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
     override BACKEND = cppsse4
   else ifeq ($(wildcard /proc/cpuinfo),)
     override BACKEND = cppnone
@@ -415,6 +416,7 @@ 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]
+# [Use 'g++ <buildflags> -E -dM - < /dev/null' to check which #define's are enabled]
 # [See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96476]
 ifeq ($(UNAME_P),ppc64le)
   ifeq ($(BACKEND),cppsse4)
@@ -426,16 +428,30 @@ ifeq ($(UNAME_P),ppc64le)
   else ifeq ($(BACKEND),cpp512z)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on PowerPC for the moment)
   endif
-else ifeq ($(UNAME_P),arm)
-  ifeq ($(BACKEND),cppsse4)
-    override AVXFLAGS = -D__SSE4_2__ # ARM NEON with 128 width (Q/quadword registers)
+else ifeq ($(UNAME_M),arm64) # ARM on Apple silicon
+  ifeq ($(BACKEND),cppnone) # this internally undefines __ARM_NEON
+    override AVXFLAGS = -DMGONGPU_NOARMNEON
+  else ifeq ($(BACKEND),cppsse4) # __ARM_NEON is always defined on Apple silicon
+    override AVXFLAGS =
   else ifeq ($(BACKEND),cppavx2)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on ARM for the moment)
   else ifeq ($(BACKEND),cpp512y)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on ARM for the moment)
   else ifeq ($(BACKEND),cpp512z)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on ARM for the moment)
   endif
+else ifeq ($(UNAME_M),aarch64) # ARM on Linux
+  ifeq ($(BACKEND),cppnone) # +nosimd ensures __ARM_NEON is absent
+    override AVXFLAGS = -march=armv8-a+nosimd
+  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)
+    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on aarch64 for the moment)
+  else ifeq ($(BACKEND),cpp512z)
+    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on aarch64 for the moment)
+  endif
 else ifneq ($(shell $(CXX) --version | grep ^nvc++),) # support nvc++ #531
   ifeq ($(BACKEND),cppnone)
     override AVXFLAGS = -mno-sse3 # no SIMD
@@ -713,7 +729,7 @@ bld512z:
 ifeq ($(UNAME_P),ppc64le)
 ###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4
 bldavxs: bldnone bldsse4
-else ifeq ($(UNAME_P),arm)
+else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
 ###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4
 bldavxs: bldnone bldsse4
 else

epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/MadtRex/makefiles/cudacpp_runner.mk

@@ -41,6 +41,7 @@ UNAME_S := $(shell uname -s)
 # Detect architecture (x86_64, ppc64le...)
 UNAME_P := $(shell uname -p)
 ###$(info UNAME_P='$(UNAME_P)')
+UNAME_M := $(shell uname -m)
 
 #-------------------------------------------------------------------------------
 
@@ -60,7 +61,7 @@ endif
 ifeq ($(BACKEND),cppauto)
   ifeq ($(UNAME_P),ppc64le)
     override BACKEND = cppsse4
-  else ifeq ($(UNAME_P),arm)
+  else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
     override BACKEND = cppsse4
   else ifeq ($(wildcard /proc/cpuinfo),)
     override BACKEND = cppnone
@@ -259,6 +260,7 @@ 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]
+# [Use 'g++ <buildflags> -E -dM - < /dev/null' to check which #define's are enabled]
 # [See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96476]
 ifeq ($(UNAME_P),ppc64le)
   ifeq ($(BACKEND),cppsse4)
@@ -270,16 +272,30 @@ ifeq ($(UNAME_P),ppc64le)
   else ifeq ($(BACKEND),cpp512z)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on PowerPC for the moment)
   endif
-else ifeq ($(UNAME_P),arm)
-  ifeq ($(BACKEND),cppsse4)
-    override AVXFLAGS = -D__SSE4_2__ # ARM NEON with 128 width (Q/quadword registers)
+else ifeq ($(UNAME_M),arm64) # ARM on Apple silicon
+  ifeq ($(BACKEND),cppnone) # this internally undefines __ARM_NEON
+    override AVXFLAGS = -DMGONGPU_NOARMNEON
+  else ifeq ($(BACKEND),cppsse4) # __ARM_NEON is always defined on Apple silicon
+    override AVXFLAGS =
   else ifeq ($(BACKEND),cppavx2)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on ARM for the moment)
   else ifeq ($(BACKEND),cpp512y)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on ARM for the moment)
   else ifeq ($(BACKEND),cpp512z)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on ARM for the moment)
   endif
+else ifeq ($(UNAME_M),aarch64) # ARM on Linux
+  ifeq ($(BACKEND),cppnone) # +nosimd ensures __ARM_NEON is absent
+    override AVXFLAGS = -march=armv8-a+nosimd
+  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)
+    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on aarch64 for the moment)
+  else ifeq ($(BACKEND),cpp512z)
+    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on aarch64 for the moment)
+  endif
 else ifneq ($(shell $(CXX) --version | grep ^nvc++),) # support nvc++ #531
   ifeq ($(BACKEND),cppnone)
     override AVXFLAGS = -mno-sse3 # no SIMD
@@ -729,7 +745,7 @@ bld512z:
 ifeq ($(UNAME_P),ppc64le)
 ###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4
 bldavxs: bldnone bldsse4
-else ifeq ($(UNAME_P),arm)
+else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
 ###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4
 bldavxs: bldnone bldsse4
 else

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

@@ -250,25 +250,22 @@ namespace mg5amcCpu
     bool known = true;
     bool ok = __builtin_cpu_supports( "vsx" );
     const std::string tag = "powerpc vsx (128bit as in SSE4.2)";
-#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!
-    const std::string tag = "arm neon (128bit as in SSE4.2)";
 #elif defined( __x86_64__ ) || defined( __i386__ )
     bool known = true;
     bool ok = __builtin_cpu_supports( "sse4.2" );
     const std::string tag = "nehalem (SSE4.2)";
 #else // AV FIXME! Added by OM for Mac, should identify the correct __xxx__ flag that should be targeted
+    // DM now we have an explicit NEON target for ARM
+    bool known = false; // __builtin_cpu_supports is not supported
+    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__
     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!
     const std::string tag = "arm neon (128bit as in SSE4.2)";
-#endif
 #else
     bool known = true;
     bool ok = true;

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

@@ -912,13 +912,13 @@ main( int argc, char** argv )
 #elif defined __SSE4_2__
 #ifdef __PPC__
   wrkflwtxt += "/ppcv";
-#elif defined __ARM_NEON__
-  wrkflwtxt += "/neon";
 #else
   wrkflwtxt += "/sse4";
 #endif
+#elif defined __ARM_NEON
+  wrkflwtxt += "/neon";
 #else
-  wrkflwtxt += "/????";                                           // no path to this statement
+  wrkflwtxt += "/????"; // no path to this statement
 #endif
   // -- Has cxtype_v::operator[] bracket with non-const reference?
 #if defined MGONGPU_CPPSIMD
@@ -1028,11 +1028,12 @@ main( int argc, char** argv )
               << "Internal loops fptype_sv    = VECTOR[" << neppV
 #ifdef __PPC__
               << "] ('sse4': PPC VSX, 128bit)" << cxtref << std::endl
-#elif defined __ARM_NEON__
-              << "] ('sse4': ARM NEON, 128bit)" << cxtref << std::endl
 #else
               << "] ('sse4': SSE4.2, 128bit)" << cxtref << std::endl
 #endif
+#elif defined __ARM_NEON
+              << "Internal loops fptype_sv    = VECTOR[" << neppV
+              << "] ('sse4': ARM NEON, 128bit)" << cxtref << std::endl
 #else
 #error Internal error: unknown SIMD build configuration
 #endif

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

@@ -41,6 +41,7 @@ UNAME_S := $(shell uname -s)
 # Detect architecture (x86_64, ppc64le...)
 UNAME_P := $(shell uname -p)
 ###$(info UNAME_P='$(UNAME_P)')
+UNAME_M := $(shell uname -m)
 
 #-------------------------------------------------------------------------------
 
@@ -60,7 +61,7 @@ endif
 ifeq ($(BACKEND),cppauto)
   ifeq ($(UNAME_P),ppc64le)
     override BACKEND = cppsse4
-  else ifeq ($(UNAME_P),arm)
+  else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
     override BACKEND = cppsse4
   else ifeq ($(wildcard /proc/cpuinfo),)
     override BACKEND = cppnone
@@ -516,6 +517,7 @@ 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 ($(BACKEND),cppsse4)
     override AVXFLAGS = -D__SSE4_2__ # Power9 VSX with 128 width (VSR registers)
@@ -526,16 +528,30 @@ ifeq ($(UNAME_P),ppc64le)
   else ifeq ($(BACKEND),cpp512z)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on PowerPC for the moment)
   endif
-else ifeq ($(UNAME_P),arm)
-  ifeq ($(BACKEND),cppsse4)
-    override AVXFLAGS = -D__SSE4_2__ # ARM NEON with 128 width (Q/quadword registers)
+else ifeq ($(UNAME_M),arm64) # ARM on Apple silicon
+  ifeq ($(BACKEND),cppnone) # this internally undefines __ARM_NEON
+    override AVXFLAGS = -DMGONGPU_NOARMNEON
+  else ifeq ($(BACKEND),cppsse4) # __ARM_NEON is always defined on Apple silicon
+    override AVXFLAGS =
   else ifeq ($(BACKEND),cppavx2)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on ARM for the moment)
   else ifeq ($(BACKEND),cpp512y)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on ARM for the moment)
   else ifeq ($(BACKEND),cpp512z)
     $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on ARM for the moment)
   endif
+else ifeq ($(UNAME_M),aarch64) # ARM on Linux
+  ifeq ($(BACKEND),cppnone) # +nosimd ensures __ARM_NEON is absent
+    override AVXFLAGS = -march=armv8-a+nosimd
+  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)
+    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on aarch64 for the moment)
+  else ifeq ($(BACKEND),cpp512z)
+    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'cppnone' and 'cppsse4' are supported on aarch64 for the moment)
+  endif
 else ifneq ($(shell $(CXX) --version | grep ^nvc++),) # support nvc++ #531
   ifeq ($(BACKEND),cppnone)
     override AVXFLAGS = -mno-sse3 # no SIMD
@@ -1092,7 +1108,7 @@ bld512z:
 ifeq ($(UNAME_P),ppc64le)
 ###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4
 bldavxs: bldnone bldsse4
-else ifeq ($(UNAME_P),arm)
+else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
 ###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4
 bldavxs: bldnone bldsse4
 else

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

@@ -16,6 +16,7 @@ endif
 # Basic uname helpers (if not already set)
 UNAME_S ?= $(shell uname -s)
 UNAME_P ?= $(shell uname -p)
+UNAME_M ?= $(shell uname -m)
 
 # Enable the C preprocessor https://gcc.gnu.org/onlinedocs/gfortran/Preprocessing-Options.html
 FFLAGS+= -cpp
@@ -114,7 +115,7 @@ $(LIBS): .libs
 	touch $@
 
 $(CUDACPP_BUILDDIR)/.cudacpplibs:
-	$(MAKE) -f $(CUDACPP_MAKEFILE)
+	$(MAKE) VERBOSE=1 -f $(CUDACPP_MAKEFILE)
 	touch $@
 
 # Remove per-library recipes from makefile to avoid duplicate sub-makes
@@ -225,7 +226,7 @@ madevent_%_link:
 # Cudacpp bldall targets
 ifeq ($(UNAME_P),ppc64le)
   bldavxs: bldnone bldsse4
-else ifeq ($(UNAME_P),arm)
+else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
   bldavxs: bldnone bldsse4
 else
   bldavxs: bldnone bldsse4 bldavx2 bld512y bld512z

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

@@ -7,10 +7,14 @@ THISDIR = $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
 
 # Host detection
 UNAME_S := $(shell uname -s)
+UNAME_P := $(shell uname -p)
+UNAME_M := $(shell uname -m)
 
-# Only add AVX2/FMA on non-mac hosts
+# Only add AVX2/FMA on non-mac and non-ARM hosts
 ifeq ($(UNAME_S),Darwin)
   GTEST_CMAKE_FLAGS :=
+else ifeq ($(UNAME_M),aarch64)
+  GTEST_CMAKE_FLAGS :=
 else
   GTEST_CMAKE_FLAGS := -DCMAKE_CXX_FLAGS="-mavx2 -mfma"
 endif

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

@@ -214,6 +214,11 @@ namespace mgOnGpu
 using mgOnGpu::fptype;
 using mgOnGpu::fptype2;
 
+// Undefine ARM_NEON (hack for cppnone on Apple silicon ARM)
+#ifdef MGONGPU_NOARMNEON
+#undef __ARM_NEON
+#endif
+
 // C++ SIMD vectorization width (this will be used to set neppV)
 #ifdef MGONGPUCPP_GPUIMPL // CUDA and HIP implementations have no SIMD
 #undef MGONGPU_CPPSIMD
@@ -235,7 +240,13 @@ using mgOnGpu::fptype2;
 #else
 #define MGONGPU_CPPSIMD 8
 #endif
-#elif defined __SSE4_2__ // C++ "sse4" SSE4.2 (128-bit ie 16-byte): 2 (DOUBLE) or 4 (FLOAT) [Power9 and ARM default]
+#elif defined __SSE4_2__ // C++ "sse4" SSE4.2 (128-bit ie 16-byte): 2 (DOUBLE) or 4 (FLOAT) [Power9 default]
+#ifdef MGONGPU_FPTYPE_DOUBLE
+#define MGONGPU_CPPSIMD 2
+#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]
 #ifdef MGONGPU_FPTYPE_DOUBLE
 #define MGONGPU_CPPSIMD 2
 #else