dRPA-CUDA: saving

src/cuda/Makefile

@@ -1,63 +1,45 @@
 NVCC = nvcc
-NFLAGS = -O2 --compiler-options '-fPIC'
+NVFLAGS = -O2 --compiler-options '-O2 -Wall'
-NDFLAGS = --shared
 
 CC = gcc
-CFLAGS = -fPIC -O2 -Wall -g
+CFLAGS = -O2 -Wall -g
 
 FC = gfortran
 FFLAGS = -O2 -Wall -g
 
 SRC_DIR = src
 INC_DIR = include
-
-BIN_DIR = bin
 BLD_DIR = build
-$(shell mkdir -p $(BIN_DIR))
 $(shell mkdir -p $(BLD_DIR))
 
 CU_SRC = $(wildcard $(SRC_DIR)/*.cu)
 CU_OBJ = $(CU_SRC:$(SRC_DIR)/%.cu=$(BLD_DIR)/%.o)
 
-C_SRC = $(SRC_DIR)/ph_drpa.c
+C_SRC = $(wildcard $(SRC_DIR)/*.c)
-C_OBJ = $(BLD_DIR)/ph_drpa.o
+C_OBJ = $(CU_SRC:$(SRC_DIR)/%.c=$(BLD_DIR)/%.o)
 
 F_SRC = $(SRC_DIR)/cu_quack_module.f90
-F_OBJ = $(BLD_DIR)/cu_quack_module.f90
+F_OBJ = $(BLD_DIR)/cu_quack_module.o
-
-MAIN_SRC = $(SRC_DIR)/cu_quack.f90
-MAIN_OBJ = $(BLD_DIR)/cu_quack.o
-
-OUTPUT_LIB = $(BLD_DIR)/libcuquack.so
-
-CUDA_LIBS = -lcudart -lcublas
 
+OUTPUT_LIB = $(BLD_DIR)/cuda.a
 
 all: $(OUTPUT_LIB)
 
-$(OUTPUT_LIB): $(CU_OBJ) $(C_OBJ)
+$(OUTPUT_LIB): $(CU_OBJ) $(C_OBJ) $(F_OBJ)
-	$(NVCC) $(NFLAGS) $(NLDFLAGS) $^ -o $@ $(CUDA_LIBS) -I$(INC_DIR)
+	ar rcs $(OUTPUT_LIB) $(CU_OBJ) $(C_OBJ) $(F_OBJ)
 
-$(BLD_DIR)/%.o: $(SRC_DIR)/%.cu
+$(CU_OBJ): $(CU_SRC)
-	$(NVCC) $(NFLAGS) -c $< -o $@ -I$(INC_DIR)
+	$(NVCC) $(NVFLAGS) -c -o $@ $< -I$(INC_DIR)
 
 $(C_OBJ): $(C_SRC)
-	@for src in $(C_SRC); do \
+	$(CC) $(CFLAGS) -c -o $@ $< -I$(INC_DIR)
-		obj=$(BLD_DIR)/$$(basename $${src} .c).o; \
-		echo "$(CC) $(CFLAGS) -c $$src -o $$obj -I$(INC_DIR)"; \
-		$(CC) $(CFLAGS) -c $$src -o $$obj -I$(INC_DIR); \
-	done
 
 $(F_OBJ): $(F_SRC)
-	$(FC) $(FFLAGS) -c $< -o $@ -J$(BLD_DIR)
+	$(FC) $(FFLAGS) -c -o $@ $< -J$(BLD_DIR)
-
-$(MAIN_OBJ): $(MAIN_SRC)
-	$(FC) $(FFLAGS) -c $< -o $@ -J$(BLD_DIR)
-
 
 
 .PHONY: clean
 clean:
-	rm -f $(BLD_DIR)/*.o $(BLD_DIR)/*.so $(BLD_DIR)/*.mod $(BIN_DIR)/*
+	rm -f $(BLD_DIR)/*
 
 

src/cuda/src/cu_quack_module.f90

@@ -0,0 +1,26 @@
+module cu_quack_module
+
+  use, intrinsic :: iso_c_binding
+
+  implicit none
+
+  interface
+
+    ! ---
+
+    subroutine ph_drpa(nO, nBas, eps, ERI) bind(C, name = "cutc_int")
+
+      import c_int, c_double
+      integer(c_int), intent(in), value :: nO, nBas
+      real(c_double), intent(in)        :: eps(nBas)
+      real(c_double), intent(in)        :: ERI(nBas,nBas,nBas,nBas)
+
+    end subroutine ph_drpa
+
+    ! ---
+
+  end interface
+
+end module cu_quack_module
+
+

src/cuda/src/ph_drpa.c

@@ -9,8 +9,6 @@
 
 int ph_drpa(int nO, int nBas, double *h_eps, double *h_ERI) {
 
-
-
     double *d_eps;
     double *d_ERI;
 

src/make_ninja.py

@@ -81,7 +81,7 @@ elif sys.platform.lower() == "linux" or os.path.exists('/proc/version'):
     else:
         if check_compiler_exists('ifort'):
             compiler = """
-FC = ifort -qmkl=parallel -qopenmp
+FC = ifort -mkl=parallel -qopenmp
 AR = ar crs
 FFLAGS = -I$IDIR -module $IDIR -traceback -g -Ofast -xHost
 CC = icc
@@ -182,9 +182,10 @@ build_main = "\n".join([
         rule_git_clone,
 ])
 
-exe_dirs = [ "QuAcK"]
+exe_dirs = ["QuAcK"]
 lib_dirs = list(filter(lambda x: os.path.isdir(x) and \
-                                 x not in exe_dirs, os.listdir(".")))
+                                x not in ["cuda"] and \
+                                x not in exe_dirs, os.listdir(".")))
 
 def create_ninja_in_libdir(directory):
     def write_rule(f, source_file, replace):