klas2 (SIMD CPU) + epoch1/epoch2

epoch1/cuda/ee_mumu/SubProcesses/Makefile

@@ -1,16 +1,62 @@
-LIBDIR   = ../../lib
 TOOLSDIR = ../../../../../tools
 TESTDIR  = ../../../../../test
 INCFLAGS = -I. -I../../src -I$(TOOLSDIR)
 MODELLIB = model_sm
 OPTFLAGS = -O3 # this ends up in CUFLAGS too (should it?), cannot add -Ofast or -ffast-math here
 OMPFLAGS?= -fopenmp
-CXXFLAGS = $(OPTFLAGS) -std=c++11 $(INCFLAGS) $(USE_NVTX) -Wall -Wshadow -Wextra $(OMPFLAGS) $(MGONGPU_CONFIG)
+CXXFLAGS = $(OPTFLAGS) -std=c++17 $(INCFLAGS) $(USE_NVTX) -Wall -Wshadow -Wextra $(OMPFLAGS) $(MGONGPU_CONFIG)
 CXXFLAGS+= -ffast-math # see issue #117
 ###CXXFLAGS+= -Ofast # performance is not different from --fast-math
 LIBFLAGS = -L$(LIBDIR) -l$(MODELLIB)
 CXX     ?= g++
 
+# AVX choice (example: "make AVX=none")
+ifneq ($(findstring clang++,$(CXX)),)
+override AVX = none
+$(warning Using AVX='$(AVX)' for clang builds)
+else ifneq ($(AVX),)
+###$(info Using AVX='$(AVX)' according to user input)
+else ifneq ($(shell grep -c avx512vl /proc/cpuinfo),0)
+override AVX = 512y
+###$(info Using AVX='$(AVX)' as no user input exists)
+else
+override AVX = avx2
+$(warning Using AVX='$(AVX)' as no user input exists and host does not support avx512vl)
+endif
+###$(info AVX=$(AVX))
+
+# Set the build flags appropriate to each AVX
+# [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]
+ifeq ($(AVX),sse4)
+override AVXFLAGS = -march=nehalem # SSE4.2 with 128 width (xmm registers)
+else ifeq ($(AVX),avx2)
+override AVXFLAGS = -march=haswell # AVX2 with 256 width (ymm registers)
+else ifeq ($(AVX),512y)
+override AVXFLAGS = -march=skylake-avx512 -mprefer-vector-width=256 # AVX512 with 256 width (ymm registers) [DEFAULT!]
+else ifeq ($(AVX),512z)
+override AVXFLAGS = -march=skylake-avx512 -DMGONGPU_PVW512 # AVX512 with 512 width (zmm registers)
+else ifneq ($(AVX),none)
+$(error Unknown AVX='$(AVX)': only 'none', 'sse4', 'avx2', '512y' and '512z' are supported)
+endif
+
+# For the moment, use AVXFLAGS everywhere: eventually, use them only in encapsulated implementations
+CXXFLAGS+= $(AVXFLAGS)
+
+# Build tag (defines target and path to the optional build directory)
+override TAG = $(AVX)
+
+# Build directory: current directory by default, or build.$(TAG) if USEBUILDDIR==1
+ifeq ($(USEBUILDDIR),1)
+override BUILDDIR = build.$(TAG)
+override LIBDIR   = ../../lib/$(BUILDDIR)
+else
+override BUILDDIR = .
+override LIBDIR   = ../../lib
+endif
+###$(info BUILDDIR=$(BUILDDIR))
+$(info Building in BUILDDIR=$(BUILDDIR) for tag=$(TAG))
+
 # If CUDA_HOME is not set, try to set it from the location of nvcc
 ifndef CUDA_HOME
   NVCC ?= $(shell which nvcc 2>/dev/null)
@@ -37,8 +83,8 @@ ifneq ($(wildcard $(CUDA_HOME)/bin/nvcc),)
   ###CUFLAGS+= --maxrregcount 128 # improves throughput: 7.3E8 (16384 32 12) up to 7.6E8 (65536 128 12)
   ###CUFLAGS+= --maxrregcount 96 # degrades throughput: 4.1E8 (16384 32 12) up to 4.5E8 (65536 128 12)
   ###CUFLAGS+= --maxrregcount 64 # degrades throughput: 1.7E8 (16384 32 12) flat at 1.7E8 (65536 128 12)
-  cu_main     = gcheck.exe
-  cu_objects  = gCPPProcess.o
+  cu_main     = $(BUILDDIR)/gcheck.exe
+  cu_objects  = $(BUILDDIR)/gCPPProcess.o
 else
   # No cuda. Switch cuda compilation off and go to common random numbers in C++
   NVCC       := $(warning CUDA_HOME is not set or is invalid. Export CUDA_HOME to compile with cuda)
@@ -51,97 +97,142 @@ endif
 
 GTESTLIBDIR = $(TESTDIR)/googletest/build/lib/
 GTESTLIBS   = $(GTESTLIBDIR)/libgtest.a $(GTESTLIBDIR)/libgtest_main.a
+
 MAKEDEBUG=
 
-cxx_main=check.exe
-cxx_objects=CPPProcess.o
+cxx_main=$(BUILDDIR)/check.exe
+cxx_objects=$(BUILDDIR)/CPPProcess.o
+
+testmain=$(BUILDDIR)/runTest.exe
 
 # Assuming uname is available, detect if architecture is power
 UNAME_P := $(shell uname -p)
 ifeq ($(UNAME_P),ppc64le)
     CUFLAGS+= -Xcompiler -mno-float128
 endif
 
-all: ../../src $(cu_main) $(cxx_main) runTest.exe
+all.$(TAG): $(BUILDDIR)/build.tag_$(TAG) ../../src $(cu_main) $(cxx_main) $(testmain)
+
+override oldtags=`find $(BUILDDIR) -maxdepth 1 -name 'build.tag_*' ! -name 'build.tag_$(TAG)'`
+$(BUILDDIR)/build.tag_$(TAG):
+	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir $(BUILDDIR)"; mkdir $(BUILDDIR); fi
+	@if [ "$(oldtags)" != "" ]; then echo -e "Cannot build for tag=$(TAG) as old builds exist for other tags:\n$(oldtags)\nPlease run 'make clean' first or consider using 'make USEBUILDDIR=1 AVX=$(AVX)'"; exit 1; fi
+	@touch $(BUILDDIR)/build.tag_$(TAG)
 
 debug: OPTFLAGS   = -g -O0 -DDEBUG2
 debug: CUOPTFLAGS = -G
 debug: MAKEDEBUG := debug
-debug: all
+debug: all.$(TAG)
 
 $(LIBDIR)/lib$(MODELLIB).a: ../../src/*.h ../../src/*.cc
-	$(MAKE) -C ../../src $(MAKEDEBUG)
+	$(MAKE) -C ../../src AVX=$(AVX) $(MAKEDEBUG)
 
-gcheck.o: gcheck.cu *.h ../../src/*.h ../../src/*.cu
+$(BUILDDIR)/gcheck.o: gcheck.cu *.h ../../src/*.h ../../src/*.cu
+	@if [ ! -d $(BUILDDIR) ]; then mkdir $(BUILDDIR); fi
 	$(NVCC) $(CPPFLAGS) $(CUFLAGS) -c $< -o $@
 
-%.o : %.cu *.h ../../src/*.h
+$(BUILDDIR)/%.o : %.cu *.h ../../src/*.h
+	@if [ ! -d $(BUILDDIR) ]; then mkdir $(BUILDDIR); fi
 	$(NVCC) $(CPPFLAGS) $(CUFLAGS) -c $< -o $@
 
-%.o : %.cc *.h ../../src/*.h
+#$(BUILDDIR)/CPPProcess.o : CPPProcess.cc *.h ../../src/*.h
+#	@if [ ! -d $(BUILDDIR) ]; then mkdir $(BUILDDIR); fi
+#	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(AVXFLAGS) $(CUINC) -c $< -o $@
+
+$(BUILDDIR)/%.o : %.cc *.h ../../src/*.h
+	@if [ ! -d $(BUILDDIR) ]; then mkdir $(BUILDDIR); fi
 	$(CXX) $(CPPFLAGS) $(CXXFLAGS) $(CUINC) -c $< -o $@
 
-$(cu_main): gcheck.o $(LIBDIR)/lib$(MODELLIB).a $(cu_objects)
+$(cu_main): $(BUILDDIR)/gcheck.o $(LIBDIR)/lib$(MODELLIB).a $(cu_objects)
 	$(NVCC) $< -o $@ $(cu_objects) $(CUARCHFLAGS) $(LIBFLAGS) $(CULIBFLAGS)
 
-$(cxx_main): check.o $(LIBDIR)/lib$(MODELLIB).a $(cxx_objects)
+$(cxx_main): $(BUILDDIR)/check.o $(LIBDIR)/lib$(MODELLIB).a $(cxx_objects)
 	$(CXX) $< -o $@ $(cxx_objects) $(CPPFLAGS) $(CXXFLAGS) -ldl -pthread $(LIBFLAGS) $(CULIBFLAGS)
 
-runTest.o:   $(GTESTLIBS)
-runTest.exe: $(GTESTLIBS)
-runTest.exe: INCFLAGS += -I$(TESTDIR)/googletest/googletest/include/
-runTest.exe: INCFLAGS += -I$(TESTDIR)/include/
-runTest.exe: LIBFLAGS += -L$(GTESTLIBDIR)/ -lgtest -lgtest_main
-runTest.exe: runTest.o $(TESTDIR)/src/MadgraphTest.o $(TESTDIR)/include/*.h
-runTest.exe: cxx_objects += runTest.o $(TESTDIR)/src/MadgraphTest.o
-runTest.exe: cu_objects  += runTest_cu.o
+$(BUILDDIR)/runTest.o: $(GTESTLIBS)
+$(testmain): $(GTESTLIBS)
+$(testmain): INCFLAGS += -I$(TESTDIR)/googletest/googletest/include
+$(testmain): INCFLAGS += -I$(TESTDIR)/include
+$(testmain): LIBFLAGS += -L$(GTESTLIBDIR) -lgtest -lgtest_main
+$(testmain): $(BUILDDIR)/runTest.o $(TESTDIR)/src/MadgraphTest.o $(TESTDIR)/include/*.h
+$(testmain): cxx_objects += $(BUILDDIR)/runTest.o $(TESTDIR)/src/MadgraphTest.o
+$(testmain): cu_objects  += $(BUILDDIR)/runTest_cu.o
 ifneq ($(findstring clang++,$(CXX)),)
-runTest.exe: LIBFLAGS += -L$(patsubst %bin/clang++,%lib,$(CXX))
+$(testmain): LIBFLAGS += -L$(patsubst %bin/clang++,%lib,$(CXX))
 endif
 
-testxxx.o:   $(GTESTLIBS)
-testxxx.o:   testxxx_cc_ref.txt
-runTest.exe: testxxx.o
-runTest.exe: cxx_objects += testxxx.o # Comment out this line to skip the test of xxx functions
+$(BUILDDIR)/testxxx.o:   $(GTESTLIBS)
+$(BUILDDIR)/testxxx.o:   testxxx_cc_ref.txt
+$(testmain): $(BUILDDIR)/testxxx.o
+$(testmain): cxx_objects += $(BUILDDIR)/testxxx.o # Comment out this line to skip the test of xxx functions
 
 ifeq ($(NVCC),)
 # Link only runTest.o
-runTest.exe: $(LIBDIR)/lib$(MODELLIB).a $(cxx_objects) $(GTESTLIBS)
+$(testmain): $(LIBDIR)/lib$(MODELLIB).a $(cxx_objects) $(GTESTLIBS) 
 	$(CXX) -o $@ $(cxx_objects) $(CPPFLAGS) $(CXXFLAGS) -ldl -pthread $(LIBFLAGS) $(CULIBFLAGS)
 else
 # Link both runTest.o and runTest_cu.o
-runTest.exe runTest_cu.o &: runTest.cc $(LIBDIR)/lib$(MODELLIB).a $(cxx_objects) $(cu_objects) $(GTESTLIBS)
-	$(NVCC) -o runTest_cu.o -c -x cu runTest.cc $(CPPFLAGS) $(CUFLAGS)
+$(testmain) $(BUILDDIR)/runTest_cu.o &: runTest.cc $(LIBDIR)/lib$(MODELLIB).a $(cxx_objects) $(cu_objects) $(GTESTLIBS)
+	$(NVCC) -o $(BUILDDIR)/runTest_cu.o -c -x cu runTest.cc $(CPPFLAGS) $(CUFLAGS)
 	$(NVCC) -o $@ $(cxx_objects) $(cu_objects) $(CPPFLAGS) $(CUFLAGS) -ldl $(LIBFLAGS) $(CULIBFLAGS) -lcuda -lgomp
 endif
 
 $(GTESTLIBS):
 	$(MAKE) -C $(TESTDIR)
 
-check: runTest.exe
-	./runTest.exe
+check: $(testmain)
+	$(testmain)
 
 .PHONY: clean
 
 clean:
-	make -C ../../src clean
-	rm -f *.o *.exe
+	make -C ../../src AVX=$(AVX) clean
+	rm -f $(BUILDDIR)/build.tag*
+ifneq ($(BUILDDIR),.)
+	rm -rf $(BUILDDIR)
+else
+	rm -f $(BUILDDIR)/*.o $(BUILDDIR)/*.exe
+endif
+
+avxall:
+	@echo
+	make USEBUILDDIR=1 AVX=none
+	@echo
+	make USEBUILDDIR=1 AVX=sse4
+	@echo
+	make USEBUILDDIR=1 AVX=avx2
+	@echo
+	make USEBUILDDIR=1 AVX=512y
+	@echo
+	make USEBUILDDIR=1 AVX=512z
 
-cleanall: clean
+cleanall:
+	@echo
+	make clean
+	@echo
+	make USEBUILDDIR=1 AVX=none clean; make -C ../../src USEBUILDDIR=1 AVX=none clean
+	@echo
+	make USEBUILDDIR=1 AVX=sse4 clean; make -C ../../src USEBUILDDIR=1 AVX=sse4 clean
+	@echo
+	make USEBUILDDIR=1 AVX=avx2 clean; make -C ../../src USEBUILDDIR=1 AVX=avx2 clean
+	@echo
+	make USEBUILDDIR=1 AVX=512y clean; make -C ../../src USEBUILDDIR=1 AVX=512y clean
+	@echo
+	make USEBUILDDIR=1 AVX=512z clean; make -C ../../src USEBUILDDIR=1 AVX=512z clean
 
-distclean: clean
+distclean: cleanall
 	make -C $(TOOLSDIR) clean
 	make -C $(TESTDIR) clean
 
 memcheck: $(cu_main)
-	/usr/local/cuda/bin/cuda-memcheck --check-api-memory-access yes --check-deprecated-instr yes --check-device-heap yes --demangle full --language c --leak-check full --racecheck-report all --report-api-errors all --show-backtrace yes --tool memcheck --track-unused-memory yes ./gcheck.exe 2 32 2
+	/usr/local/cuda/bin/cuda-memcheck --check-api-memory-access yes --check-deprecated-instr yes --check-device-heap yes --demangle full --language c --leak-check full --racecheck-report all --report-api-errors all --show-backtrace yes --tool memcheck --track-unused-memory yes $(BUILDDIR)/gcheck.exe 2 32 2
 
 perf: force
 	make clean && make
-	time ./gcheck.exe -p 16348 32 12 && date
+	time $(BUILDDIR)/gcheck.exe -p 16348 32 12 && date
 
 test: force
-	./gcheck.exe -v 1 32 1
+	$(BUILDDIR)/gcheck.exe -v 1 32 1
 
 info:
 	@hostname

epoch1/cuda/ee_mumu/SubProcesses/Memory.h

@@ -6,20 +6,16 @@
  */
 
 #ifndef MEMORY_H
-#define MEMORY_H
+#define MEMORY_H 1
 
+#include "mgOnGpuConfig.h"
 #include "mgOnGpuTypes.h"
+#include "mgOnGpuVectors.h"
 
 #include <memory>
 
-template<typename T = fptype>
-struct CudaHstDeleter {
-  void operator()(T* mem) {
-    checkCuda( cudaFreeHost( mem ) );
-  }
-};
-
 #ifdef __CUDACC__
+
 template<typename T = fptype>
 struct CudaDevDeleter {
   void operator()(T* mem) {
@@ -28,31 +24,38 @@ struct CudaDevDeleter {
 };
 
 template<typename T = fptype>
-using unique_ptr_dev = std::unique_ptr<T, CudaDevDeleter<T>>;
-
-template<typename T = fptype>
-unique_ptr_dev<T> devMakeUnique(std::size_t N) {
+std::unique_ptr<T, CudaDevDeleter<T>> devMakeUnique(std::size_t N) {
   T* tmp = nullptr;
   checkCuda( cudaMalloc( &tmp, N * sizeof(T) ) );
   return std::unique_ptr<T, CudaDevDeleter<T>>{ tmp };
 }
 
 template<typename T = fptype>
-using unique_ptr_host = std::unique_ptr<T[], CudaHstDeleter<T>>;
+struct CudaHstDeleter {
+  void operator()(T* mem) {
+    checkCuda( cudaFreeHost( mem ) );
+  }
+};
 
 template<typename T = fptype>
-unique_ptr_host<T> hstMakeUnique(std::size_t N) {
+std::unique_ptr<T[], CudaHstDeleter<T>> hstMakeUnique(std::size_t N) {
   T* tmp = nullptr;
   checkCuda( cudaMallocHost( &tmp, N * sizeof(T) ) );
   return std::unique_ptr<T[], CudaHstDeleter<T>>{ tmp };
 };
+
 #else
-template<typename T = fptype>
-using unique_ptr_host = std::unique_ptr<T[]>;
-template<typename T = fptype>
-std::unique_ptr<T[]> hstMakeUnique(std::size_t N) { return std::unique_ptr<T[]>{ new T[N] }; };
-#endif
 
+template<typename T = fptype> inline
+std::unique_ptr<T[]> hstMakeUnique(std::size_t N) { return std::unique_ptr<T[]>{ new T[N]() }; };
+
+#ifdef MGONGPU_CPPSIMD
+
+template<> inline
+std::unique_ptr<fptype_v[]> hstMakeUnique(std::size_t N) { return std::unique_ptr<fptype_v[]>{ new fptype_v[N/neppV]() }; };
 
+#endif
+
+#endif
 
 #endif /* MEMORY_H */