Maskrcnn-benchmark: Build error under Mac OS X 10.13.6 (no instance of function template "THCCeilDiv" matches)

[zhangew]

🐛 Bug

To Reproduce

Steps to reproduce the behavior:

1. Follow the install.md, and when arrive at the last step for Mac OS X, run it:
2. MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py build develop
3. 4 errors occured:

/github/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cuda/ROIPool_cuda.cu(129): error: no instance of function template "THCCeilDiv" matches the argument list
argument types are: (long long, long)

/github/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cuda/ROIPool_cuda.cu(129): error: no instance of overloaded function "std::min" matches the argument list
argument types are: (<error-type>, long)

/github/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cuda/ROIPool_cuda.cu(176): error: no instance of function template "THCCeilDiv" matches the argument list
argument types are: (int64_t, long)

/github/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cuda/ROIPool_cuda.cu(176): error: no instance of overloaded function "std::min" matches the argument list
argument types are: (<error-type>, long)

Expected behavior

No build errors.

Environment

Collecting environment information...
PyTorch version: 1.0.0a0+cb79e1b
Is debug build: No
CUDA used to build PyTorch: 9.2

OS: Mac OSX 10.13.6
GCC version: Could not collect
CMake version: version 3.12.4

Python version: 3.6
Is CUDA available: Yes
CUDA runtime version: 9.2.148
GPU models and configuration: Nvidia Gforce GT750M
Nvidia driver version: 387.10.10.10.40.108
cuDNN version: Probably one of the following:
/usr/local/cuda/lib/libcudnn.7.dylib
/usr/local/cuda/lib/libcudnn.dylib
/usr/local/cuda/lib/libcudnn_static.a

Versions of relevant libraries:
[pip] numpy (1.14.0)
[pip] numpydoc (0.7.0)
[pip] torch (1.0.0a0+cb79e1b)
[pip] torchvision (0.2.1)
[conda] mkl 2018.0.1 hfbd8650_4
[conda] mkl-service 1.1.2 py36h7ea6df4_4
[conda] torch 1.0.0a0+cb79e1b
[conda] torch 0.5.0a0+a24163a
[conda] torchvision 0.2.1

Additional context

No.

[fmassa]

Hum, this is interesting, I wonder if this is something that was changed recently in PyTorch.

Can you try replacing
https://github.com/facebookresearch/maskrcnn-benchmark/blob/61ffdb3803db562c23d883439aee16e599c051e6/maskrcnn_benchmark/csrc/cuda/ROIPool_cuda.cu#L129
by

dim3 grid(std::min(THCCeilDiv((long)output_size, 512L), 4096L));

and similarly in the other places and report back?

[zhangew]

@fmassa , thanks!
I've tried your modification, added a (long) into the line, before output_size. It compiled, with no error or warning.
I'll try this method to the other places.

[zhangew]

Finally, I built it successfully.
Modified 4 lines: ROIPool_cuda.cu(Ln129, Ln176), ROIAlign_cuda.cu(Ln275, Ln320).
BTW, when linking ,another error reported:

clang++ -bundle -undefined dynamic_lookup -L/anaconda3/lib -arch x86_64 -L/anaconda3/lib -arch x86_64 -arch x86_64 build/temp.macosx-10.7-x86_64-3.6/Users/enwei/github/maskrcnn-benchmark/maskrcnn_benchmark/csrc/vision.o build/temp.macosx-10.7-x86_64-3.6/Users/enwei/github/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cpu/ROIAlign_cpu.o build/temp.macosx-10.7-x86_64-3.6/Users/enwei/github/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cpu/nms_cpu.o build/temp.macosx-10.7-x86_64-3.6/Users/enwei/github/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cuda/nms.o build/temp.macosx-10.7-x86_64-3.6/Users/enwei/github/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cuda/ROIPool_cuda.o build/temp.macosx-10.7-x86_64-3.6/Users/enwei/github/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cuda/ROIAlign_cuda.o -L/usr/local/cuda/lib64 -lcudart -o build/lib.macosx-10.7-x86_64-3.6/maskrcnn_benchmark/_C.cpython-36m-darwin.so
ld: warning: directory not found for option '-L/usr/local/cuda/lib64'
ld: library not found for -lcudart
clang: error: linker command failed with exit code 1 (use -v to see invocation)
error: command 'clang++' failed with exit status 1

I found libcudart.dylib in /usr/local/cuda/lib, but there is no lib64 directory, so I created a symbolic link to lib, named 'lib64'.

The last 2 lines of the building process are:
Processing dependencies for maskrcnn-benchmark==0.1
Finished processing dependencies for maskrcnn-benchmark==0.1

[fmassa]

Great, thanks for letting me know. I'll try to see what's the best way to fix the error you faced.

[TomHeaven]

The adding 'long' modification works for me. I'm using Pytorch 1.0.