CUDA 编译
0x00
Prior to the 5.0 release, CUDA did not support separate compilation, so CUDA code could not call device functions or access variables across files. Such compilation is referred to as whole program compilation. We have always supported the separate compilation of host code, it was just the device CUDA code that needed to all be within one file. Starting with CUDA 5.0, separate compilation of device code is supported, but the old whole program mode is still the default, so there are new options to invoke separate compilation.
CUDA5.0版本前没有独立编译的支持,所以CUDA的代码不能跨文件访问函数和变量。当时称为[整体编译],值得注意的是,无论CUDA什么版本,都支持host端的独立编译(例如g++编译器),在这种情况下,cuda的代码要放在一个文件里。从5.0以后,开始支持device端的独立编译(nvcc编译器)。默认还是整体编译,使用独立编译需要设置nvcc参数。
关于CUDA的编译过程可以参考这个文档,https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#using-separate-compilation-in-cuda
0x01
举个例子,在cmu15418的作业中,有写好的makefile,如下
1
2 EXECUTABLE := cudaSaxpy
3
4 CU_FILES := saxpy.cu
5
6 CU_DEPS :=
7
8 CC_FILES := main.cpp
9
10 ###########################################################
11
12 ARCH=$(shell uname | sed -e 's/-.*//g')
13
14 OBJDIR=objs
15 CXX=g++ -m64
16 CXXFLAGS=-O3 -Wall
17 ifeq ($(ARCH), Darwin)
18 # Building on mac
19 LDFLAGS=-L/usr/local/depot/cuda-8.0/lib/ -lcudart
20 else
21 # Building on Linux
22 # LDFLAGS=-L/usr/local/cuda-11.2/lib64/ -lcudart
23 LDFLAGS=-L/data/cuda/cuda-11.1/cuda/lib64 -lcudart
24 endif
25 NVCC=nvcc
26 NVCCFLAGS=-O3 -m64 #--gpu-architecture compute_35
27
28
29 OBJS=$(OBJDIR)/main.o $(OBJDIR)/saxpy.o
30
31
32 .PHONY: dirs clean
33
34 default: $(EXECUTABLE)
35
36 dirs:
37 mkdir -p $(OBJDIR)/
38
39 clean:
40 rm -rf $(OBJDIR) *.ppm *~ $(EXECUTABLE)
41
42 $(EXECUTABLE): dirs $(OBJS)
43 $(CXX) $(CXXFLAGS) -o $@ $(OBJS) $(LDFLAGS)
44
45 $(OBJDIR)/%.o: %.cpp
46 $(CXX) $< $(CXXFLAGS) -c -o $@
47
48 $(OBJDIR)/%.o: %.cu
49 $(NVCC) $< $(NVCCFLAGS) -c -o $@
default -> executable -> objs -> objdir/main.o & objdir/saxpy.o
objdir/main.o -> main.cpp & main.cu
main.cpp -> g++ -m64 main.cpp -O3 -Wall -c -o main.o
objdir/saxpy.o -> saxpy.cpp & saxpy.cu
saxpy.cu -> nvcc saxpy.cu -O3 -m64 -c -o saxpy.o
dirs objs/main.o objs/saxpy.o -> g++ -m64 -o cudaSaxpy objs/main.o objs/saxpy.o -L/data/cuda/cuda-11.1/cuda/lib64 -lcudart
关于makefile中特殊符号的定义:
https://stackoverflow.com/questions/3220277/what-do-the-makefile-symbols-and-mean