Add a new Makefile for builds with embedded-clang

LBM/Makefile.custom-clang

@@ -0,0 +1,147 @@
+#===============================================================================
+# User Options
+#===============================================================================
+
+CUDA_PATH ?= /usr/local/cuda-12.9
+
+ENZYME_PATH ?= $(HOME)/git/Reactant/enzyme/buildRelease/Enzyme/ClangEnzyme-22.so
+CLANG_PATH ?= $(HOME)/git/Enzyme/llvm-project/buildRelease/bin/clang++
+LIB_RAISE_PATH ?= $(HOME)/git/Enzyme-JaX/bazel-bin/libRaise.so
+REACTANT_EXTRA_PATH ?= $(HOME)/git/Reactant.jl/deps/ReactantExtra/bazel-bin/
+
+export REACTANT_DEBUG=1
+export PATH := $(CUDA_PATH)/bin:$(PATH)
+export POLYGEIST_GPU_KERNEL_BLOCK_SIZE = 128
+export LD_LIBRARY_PATH := $(REACTANT_EXTRA_PATH):$(LD_LIBRARY_PATH)
+
+OPTIMIZE    ?= no # yes
+
+NEWCACHE    ?= yes
+AA          ?= no
+PHISTRUCT   ?= no
+FORWARD     ?= yes
+COALESE     ?= yes
+CACHELICM   ?= yes
+ABI         ?= no
+ALLOCATOR   ?= no
+VERIFY      ?= no
+
+DEBUG       = no
+PROFILE     = no
+SM_VERSION  = 120
+
+#===============================================================================
+# Program name & source code list
+#===============================================================================
+
+program = lbm
+
+source = args.c lbm.cu lbm_kernel.cu parboil_cuda.c main.cc
+
+#===============================================================================
+# Sets Flags
+#===============================================================================
+
+# Standard Flags
+CFLAGS := -mllvm -max-heap-to-stack-size=1000000 -I $(CUDA_PATH)/include -I .
+# -mllvm -max-heap-to-stack-size=-1
+# -Rpass=attributor -mllvm -debug -mllvm -debug-only=attributor
+
+
+CC := $(CLANG_PATH)
+CFLAGS += -ffast-math --cuda-path=$(CUDA_PATH) -L$(CUDA_PATH)/lib64 --cuda-gpu-arch=sm_$(SM_VERSION)  -std=c++11 -Xclang -add-plugin -Xclang enzyme -mllvm -raising-plugin-path -mllvm $(LIB_RAISE_PATH) -mllvm -reactant-backend -mllvm cuda -x cuda -fno-exceptions
+
+# Linker Flags
+LDFLAGS = -lcudart_static "-ldl" "-lrt" -lpthread -lm 
+# -L$(REACTANT_EXTRA_PATH) # -lReactantExtra
+
+# Debug Flags
+ifeq ($(DEBUG),yes)
+  CFLAGS += -g
+  LDFLAGS += -g
+endif
+
+ifeq ($(ALLOCATOR),yes)
+  CFLAGS += -DALLOCATOR
+endif
+
+ifeq ($(VERIFY),yes)
+  CFLAGS += -DVERIFY
+endif
+
+ifeq ($(ABI),yes)
+  CFLAGS += -DABI
+else
+  CFLAGS += -DSIZE=8
+endif
+
+# Profiling Flags
+ifeq ($(PROFILE),yes)
+  CFLAGS += -pg
+  LDFLAGS  += -pg
+endif
+
+# Optimization Flags
+ifeq ($(OPTIMIZE),yes)
+  CFLAGS += -O3
+endif
+
+
+# Optimization Flags
+ifeq ($(FORWARD),yes)
+  CFLAGS += 
+else
+  CFLAGS += -DALLOW_AD=1 
+endif
+
+
+# Optimization Flags
+# ifeq ($(NEWCACHE),yes)
+#   CFLAGS += -mllvm -enzyme-new-cache=1 -mllvm -enzyme-mincut-cache=1
+#   ifeq ($(ABI),yes)
+#     CFLAGS += -DSIZE=20
+#   endif
+# else
+#   CFLAGS += -mllvm -enzyme-new-cache=0 -mllvm -enzyme-mincut-cache=0
+#   ifeq ($(ABI),yes)
+#     CFLAGS += -DSIZE=80
+#   endif
+# endif
+
+# ifeq ($(AA),yes)
+#   CFLAGS += -mllvm -enzyme-aggressive-aa=1
+# else
+#   CFLAGS += -mllvm -enzyme-aggressive-aa=0
+# endif
+
+
+# ifeq ($(PHISTRUCT),yes)
+#   CFLAGS += -mllvm -enzyme-phi-restructure=1
+# else
+#   CFLAGS += -mllvm -enzyme-phi-restructure=0
+# endif
+
+# ifeq ($(COALESE),yes)
+#   CFLAGS += -mllvm -enzyme-coalese
+# endif
+
+# ifeq ($(CACHELICM),yes)
+#   CFLAGS +=  -mllvm -enzyme-loop-invariant-cache=1
+# else
+#   CFLAGS +=  -mllvm -enzyme-loop-invariant-cache=0
+# endif
+
+
+
+#===============================================================================
+# Targets to Build
+#===============================================================================
+
+$(program): $(source) Makefile
+	$(CC) $(CFLAGS) args.c lbm.cu lbm_kernel.cu parboil_cuda.c main.cc $(LDFLAGS) -o $@
+
+clean:
+	rm -rf $(program) *.o *.re_export
+
+run: $(program)
+	./$(program) 10 -o out.txt