added EXTRA_FLAGS variable to CUDA Makefile to provide the freedom to… #28
Conversation
… specify debug flags or gencode flags
|
You don't need to predefine |
|
might be, but that wouldn't have switched off `-O3` that was hard coded
in the Makefile stub.
|
|
That's true. Are you able to put the |
|
While I am at it, I'll align the cuda make to OpenMP.cmake although I'd
recommend using `?=` where possible. OpenMP.cmake does always set
variables directly where `?=` might offer more flexiblity.
I'll finish the CUDA.make and thus leave the door open for CUDA-with-clang.
|
|
What flexibility are you seeing with |
|
With this:
```
CXXFLAGS?=-O3 -std=c++11
GPUCXX?=nvcc
cuda-stream: main.cpp CUDAStream.cu
$(GPUCXX) $(CXXFLAGS) -DCUDA $^ $(EXTRA_FLAGS) -o $@
```
I can pitch in anything on the command line and the Makefile stays quite
concise:
```
$ CXXFLAGS="-g -G -std=c++11" make -f CUDA.cmake #produces debug output
in case I screwed up
$ make -f CUDA.cmake #produces the default -O3 binary
```
Just saying -
|
|
Sure, but the same is true without the You can still override |
|
NB. I still need to dig the GNU make manual.
Ha, learned something again! I checked your example with a handful of
GNU make builds:
**with `?=`**
```
$ cat testme.make
CXXFLAGS?=-std=c++11 -O3
cuda-stream: main.cpp CUDAStream.cu
nvcc ${CXXFLAGS} -DCUDA $^ $(EXTRA_FLAGS) -o $@
```
using `?=` is agnostic of the position of the variable definition:
```
$ CXXFLAGS=-bla make -f testme.make -n
nvcc -bla -DCUDA main.cpp CUDAStream.cu -o cuda-stream
$ make -f testme.make -n CXXFLAGS=-bla
nvcc -bla -DCUDA main.cpp CUDAStream.cu -o cuda-stream
```
**without `?=`**
```
$ cat testme.make
CXXFLAGS=-std=c++11 -O3
cuda-stream: main.cpp CUDAStream.cu
nvcc ${CXXFLAGS} -DCUDA $^ $(EXTRA_FLAGS) -o $@
```
Now, watch this:
~~~
$ make -f testme.make -n CXXFLAGS=-bla
nvcc -bla -DCUDA main.cpp CUDAStream.cu -o cuda-stream
$ CXXFLAGS=-bla make -f testme.make -n
nvcc -std=c++11 -O3 -DCUDA main.cpp CUDAStream.cu -o cuda-stream
~~~
See the last call? the updated value of CXXFLAGS is not used.
|
|
we didn't come to a conclusion on this, did we? |
|
Thanks for your investigation into But for the purposes of this pull request, can you pull out the |
make variable to cope with clang as CUDA compiler as well
|
done |
|
Merged, thanks for the contribution. |
|
Created Issue #30 for switching to |
… specify debug flags or gencode flags,
in more detail, I ran into trouble when trying to benchmark more than
arraysize = 32*1024*1024. This is due to the fact that by default, not-gencodeflags or similar are setup for the nvcc call of the cuda GPU stream. This defaults to an SM that doesn't support an arraysize of this magnitude and hence theinit_arraysfails with error code0xb Invalid arguments. We had the discussion about gencodes before in #8, but the new Makefile based build system doesn't reflect the outcome of #8.