ROOT := .. COMPONENT := libffwd include ../build.mk LIB := ffwd STATIC := $(OUTDIR)/lib$(LIB).a SHARED := $(OUTDIR)/lib$(LIB).$(SHARED_EXT) CFLAGS += -std=c11 -pthread $(BACKEND_CFLAGS) CPPFLAGS += -I. SHARED_FLAGS += -pthread # Restrict the shared library's exported symbols to the public ABI. SHARED_LDFLAGS += $(EXPORT_FLAGS) # Tokenizer, config, safetensors reader, and L2/cosine helpers: every backend. COMMON_SRCS := config.c vec.c tokenize.c bpe.c sentencepiece.c wordpiece.c safetensors.c # CPU math kernels and SIMD dispatch. generic.c is always built; avx.c and # neon.c each guard their entire body by arch (__AVX2__ / __ARM_NEON), so the # off-target file compiles to an empty translation unit. KERNEL_SRCS := act.c gemm.c norm.c rope.c attn.c threadpool.c generic.c avx.c neon.c BLAS_SRCS := backend_blas.c model.c model_blas.c forward.c late.c late_blas.c alloc.c \ $(COMMON_SRCS) $(KERNEL_SRCS) MLX_SRCS := backend_mlx.c mlx.c $(COMMON_SRCS) CUDA_SRCS := backend_cuda.c model.c late.c alloc.c $(COMMON_SRCS) $(KERNEL_SRCS) ifeq ($(BACKEND),mlx) SRCS := $(MLX_SRCS) CU_SRCS := else ifeq ($(BACKEND),cuda) SRCS := $(CUDA_SRCS) CU_SRCS := cuda.cu else SRCS := $(BLAS_SRCS) CU_SRCS := endif OBJS := $(call objs,$(SRCS) $(CU_SRCS)) DEPS := $(call deps,$(OBJS)) .PHONY: all clean all: $(STATIC) $(SHARED) $(eval $(call static_archive,$(STATIC),$(OBJS))) $(eval $(call shared_library,$(SHARED),$(OBJS),,$(BACKEND_LDFLAGS) -lm)) # The export map is generated from the public ABI (the FFWD_API declarations in # ffwd.h) instead of being committed, so it cannot drift from the header. It is # regenerated when the header or the generator changes, and the shared library # relinks when it does. $(EXPORT_FILE): ffwd.h $(ROOT)/devtools/gen_exports.py python3 $(ROOT)/devtools/gen_exports.py --outdir $(ROOT) $(SHARED): $(EXPORT_FILE) # Removes only this MODE/BACKEND's objects and artifacts; the root `make clean` # removes the whole build/ tree. clean: $(RM) -r $(OBJDIR) $(STATIC) $(SHARED) $(SHARED).dSYM -include $(DEPS)