Why can I skip "device code linking" with a non-nvcc linker? - makefile

Question

If I run the following:
c++ -c --std=c++11 $(includes) -o src/main.o src/main.cpp
nvcc -c -m64 -arch=sm_30 --std=c++11 $(includes) -o src/kernels/add.o src/kernels/add.cu
ar qc src/kernels/libkernels.a src/kernels/add.o
ranlib src/kernels/libkernels.a
c++ -o program -L/usr/local/cuda/lib64 src/main.o src/kernels/libkernels.a -lcudart -lcudadevrt
It works. Shouldn't it fail because I didn't perform a -dlink phase? The Parallel4All blog entry on separate compilation says:
When you use nvcc to link, there is nothing special to do: replace your normal compiler command with nvcc and it will take care of all the necessary steps. However, you may choose to use a compiler driver other than nvcc (such as g++) for the final link step. Since your CPU compiler will not know how to link CUDA device code, you’ll have to add a step in your build to have nvcc link the CUDA device code, using the nvcc option –dlink.
nvcc –arch=sm_20 –dlink v3.o particle.o main.o –o gpuCode.o
This links all the device object code and places it into gpuCode.o. Note that this does not link the CPU object code. In fact, the CPU object code in v3.o, particle.o, and main.o is discarded in this step. To complete the link to an executable, we can use ld or g++.
g++ gpuCode.o main.o particle.o v3.o –lcudart –o app
Does the use of a .a library somehow make up for the lack of "device code linking"?
PS - I'm using CUDA 8.0.61 on Linux Mint 18.2

Answer 1

Device code linking is not required in all scenarios. (This must be true, because prior to CUDA 5.0 there was no device code linking.)
Device code linking is required in a number of scenarios, the most typical being when linking of device code must occur across different compilation units. This means that device code in one module (,file,compilation unit) calls device code in another module (,file, compilation unit).
I can tell for a fact that this scenario is not present in your case, because there is exactly one module (,file, compilation unit) of yours that contains any device code:
nvcc -c -m64 -arch=sm_30 --std=c++11 $(includes) -o src/kernels/add.o src/kernels/add.cu
^^
only one file here
I know this to be true, because any attempt to compile any device code by an ordinary host-code compiler other than nvcc will throw syntax errors. Since this is not happening in your case, and you have only one file that could possibly have device code, you cannot possibly have a scenario where device-code linking is required, so your method works.
Even if you had for example multiple .cu files, you might still not need device code linking if no device code in a file called device code (or referenced device variables) in another file.
The relevant section of the nvcc manual covers the topic of device code linking in more detail.