
# Compiler and flags
CC := g++
CXXFLAGS := -std=c++14 -Wall -Iinclude -O3 -march=native -fopenmp
LDFLAGS  := -fopenmp
# Compiles .h files when there's been a change
DEPFLAGS := -MMD -MP

CXX ?= g++




# Directories
SRC_DIR := src
INC_DIR := include
OBJ_DIR := obj
BIN_DIR := bin
TEST_BIN := $(BIN_DIR)/tests





#
# === Executable Name ===
TARGET := $(BIN_DIR)/Flux

# === Find all .cpp files recursively in src/ ===
SRCS := $(shell find $(SRC_DIR) -name '*.cpp')

# === Convert src/foo/bar.cpp → obj/foo/bar.o ===
OBJS := $(patsubst $(SRC_DIR)/%.cpp,$(OBJ_DIR)/%.o,$(SRCS))


# === Test sources ===
TEST_SRCS := $(shell find test -name 'test_*.cpp')
TEST_OBJS := $(patsubst test/%.cpp, $(OBJ_DIR)/test/%.o, $(TEST_SRCS))

# Compiles .h files when there's been a change
DEPS := $(OBJS:.o=.d) $(TEST_OBJS:.o=.d) $(TEST_MAIN:.o=.d)
-include $(DEPS)

# The single file that defines TEST_MAIN / main()
TEST_MAIN := $(OBJ_DIR)/test/test_all.o


# === OpenMP runtime configuration (override-able) ===
OMP_PROC_BIND   ?= close        # close|spread|master
OMP_PLACES      ?= cores        # cores|threads|sockets
OMP_MAX_LEVELS  ?= 1            # 1 = no nested teams; set 2+ to allow nesting
OMP_THREADS     ?= 16           # e.g. "16" or "8,4" for nested (outer,inner)
OMP_DYNAMIC     ?= TRUE         # TRUE/FALSE: let runtime adjust threads
OMP_SCHEDULE    ?= STATIC     	# STATIC recommended for matvec/matmul
OMP_DISPLAY_ENV ?= FALSE        # TRUE to print runtime config at startup

# Export OMP defaults so child makes or tools see them (not strictly required)
export OMP_PROC_BIND
export OMP_PLACES
export OMP_MAX_LEVELS
export OMP_THREADS
export OMP_DYNAMIC
export OMP_SCHEDULE
export OMP_DISPLAY_ENV

# What belongs to "run" vs "test"
RUN_DEPS := $(OBJS:.o=.d)
TEST_DEPS := $(TEST_OBJS:.o=.d) $(TEST_MAIN:.o=.d)

RUN_ARTIFACTS  := $(TARGET) $(OBJS) $(RUN_DEPS)
TEST_ARTIFACTS := $(TEST_BIN) $(TEST_OBJS) $(TEST_MAIN) $(TEST_DEPS)


# === Default Target ===
all: $(TARGET)

# === Linking final executable ===
$(TARGET): $(OBJS)
	@mkdir -p $(dir $@)
	$(CXX) $(CXXFLAGS) -o $@ $^ $(LDFLAGS)

# === Compiling source files to object files ===
$(OBJ_DIR)/%.o: $(SRC_DIR)/%.cpp
	@mkdir -p $(dir $@)
	$(CXX) $(CXXFLAGS) $(DEPFLAGS) -c $< -o $@

# === Run with OpenMP env set only for the run ===
.PHONY: run
run: clean-test $(TARGET)
	@echo ">>> OMP_PROC_BIND=$(OMP_PROC_BIND)"
	@echo ">>> OMP_PLACES=$(OMP_PLACES)"
	@echo ">>> OMP_MAX_ACTIVE_LEVELS=$(OMP_MAX_LEVELS)"
	@echo ">>> OMP_NUM_THREADS=$(OMP_THREADS)"
	@echo ">>> OMP_DYNAMIC=$(OMP_DYNAMIC)"
	@echo ">>> OMP_SCHEDULE=$(OMP_SCHEDULE)"
	@echo ">>> OMP_DISPLAY_ENV=$(OMP_DISPLAY_ENV)"
	@OMP_PROC_BIND=$(OMP_PROC_BIND) \
	OMP_PLACES=$(OMP_PLACES) \
	OMP_MAX_ACTIVE_LEVELS=$(OMP_MAX_LEVELS) \
	OMP_NUM_THREADS="$(OMP_THREADS)" \
	OMP_DYNAMIC=$(OMP_DYNAMIC) \
	OMP_SCHEDULE=$(OMP_SCHEDULE) \
	OMP_DISPLAY_ENV=$(OMP_DISPLAY_ENV) \
	./$(TARGET)

# Handy presets
.PHONY: run-single-core
run-single-core: ## Single-level one core (good default)
	$(MAKE) run OMP_MAX_LEVELS=1 OMP_THREADS=1 OMP_PROC_BIND=close OMP_PLACES=cores

# Handy presets
.PHONY: run-multi-core
run-multi-core: ## Single-level parallel (good default)
	$(MAKE) run OMP_MAX_LEVELS=1 OMP_THREADS=16 OMP_PROC_BIND=close OMP_PLACES=cores

.PHONY: run-nested-multi-core
run-nested-multi-core: ## Two-level nested (outer,inner), adjust to your cores
	$(MAKE) run OMP_MAX_LEVELS=2 OMP_THREADS="2,8" OMP_PROC_BIND=close OMP_PLACES=cores



# Optional: print debug info
.PHONY: info
info:
	@echo "Source files: $(SRCS)"
	@echo "Object files: $(OBJS)"
	@echo "CXXFLAGS: $(CXXFLAGS)"
	@echo "LDFLAGS:   $(LDFLAGS)"


.PHONY: test
test: clean-run $(TEST_BIN)
	@echo ">>> OMP_PROC_BIND=$(OMP_PROC_BIND)"
	@echo ">>> OMP_PLACES=$(OMP_PLACES)"
	@echo ">>> OMP_MAX_ACTIVE_LEVELS=$(OMP_MAX_LEVELS)"
	@echo ">>> OMP_NUM_THREADS=$(OMP_THREADS)"
	@echo ">>> OMP_DYNAMIC=$(OMP_DYNAMIC)"
	@echo ">>> OMP_SCHEDULE=$(OMP_SCHEDULE)"
	@echo ">>> OMP_DISPLAY_ENV=$(OMP_DISPLAY_ENV)"
	@OMP_PROC_BIND=$(OMP_PROC_BIND) \
	OMP_PLACES=$(OMP_PLACES) \
	OMP_MAX_ACTIVE_LEVELS=$(OMP_MAX_LEVELS) \
	OMP_NUM_THREADS="$(OMP_THREADS)" \
	OMP_DYNAMIC=$(OMP_DYNAMIC) \
	OMP_SCHEDULE=$(OMP_SCHEDULE) \
	OMP_DISPLAY_ENV=$(OMP_DISPLAY_ENV) \
	$(TEST_BIN)

$(TEST_BIN): $(TEST_OBJS) $(TEST_MAIN)
	@mkdir -p $(BIN_DIR)
	$(CXX) $(CXXFLAGS) -o $@ $^ $(LDFLAGS)

$(OBJ_DIR)/test/%.o: test/%.cpp
	@mkdir -p $(dir $@)
	$(CXX) $(CXXFLAGS) $(DEPFLAGS) -c $< -o $@

# Clean up
.PHONY: clean
clean:
	rm -rf $(OBJ_DIR) $(BIN_DIR)

.PHONY: clean-run clean-test
clean-run:
	@echo "Cleaning run artifacts..."
	@rm -f $(RUN_ARTIFACTS)

clean-test:
	@echo "Cleaning test artifacts..."
	@rm -f $(TEST_ARTIFACTS)