Makefile

GPU_ARCH := sm_120
NVCC := /usr/local/cuda/bin/nvcc

pair_sort_real_gpu: src/main_real.cu
	$(NVCC) -O3 -arch=$(GPU_ARCH) -std=c++20 -Xcompiler -fopenmp -lgomp -o $@ $<

pair_sort_real_cpu: src/main_real_cpu.cpp
	g++ -O3 -std=c++20 -fopenmp -lgomp -o $@ $<

pair_sort_preprocess_gpu: src/main_preprocess.cu src/mem_preprocess.cuh
	$(NVCC) -O3 -arch=$(GPU_ARCH) -std=c++20 -o $@ $<

pair_sort_full_gpu: src/main_full.cu src/count_and_plan.cu src/count_and_plan.cuh
	$(NVCC) -O3 -arch=$(GPU_ARCH) -std=c++20 -Xcompiler -fopenmp -lgomp -o $@ src/main_full.cu src/count_and_plan.cu

# ─── Static library for Rust FFI ───────────────────────────────────
# Builds libcount_and_plan.a containing the CountAndPlan implementation
# + the C ABI shim. Rust's build.rs links against this + libcudart.

CNP_OBJS := build/count_and_plan.o build/count_and_plan_c.o

build/count_and_plan.o: src/count_and_plan.cu src/count_and_plan.cuh
	@mkdir -p build
	$(NVCC) -O3 -arch=$(GPU_ARCH) -std=c++20 -Xcompiler -fPIC -dc -c -o $@ $<

build/count_and_plan_c.o: src/count_and_plan_c.cu src/count_and_plan.cuh
	@mkdir -p build
	$(NVCC) -O3 -arch=$(GPU_ARCH) -std=c++20 -Xcompiler -fPIC -dc -c -o $@ $<

# Resolve device-relocatable symbols across the two TUs into one object.
build/count_and_plan_dlink.o: $(CNP_OBJS)
	$(NVCC) -arch=$(GPU_ARCH) -dlink -Xcompiler -fPIC -o $@ $(CNP_OBJS)

build/libcount_and_plan.a: $(CNP_OBJS) build/count_and_plan_dlink.o
	@mkdir -p build
	ar rcs $@ $(CNP_OBJS) build/count_and_plan_dlink.o

libcount_and_plan: build/libcount_and_plan.a

clean:
	rm -f pair_sort_real_gpu pair_sort_real_cpu pair_sort_preprocess_gpu pair_sort_full_gpu
	rm -rf build

.PHONY: clean libcount_and_plan