Where should I put cblas.h so that my build of mxnet will find it?

Question

I'm trying to install mxnet by cloning a git repository as follows:

git clone --recursive https://github.com/dmlc/mxnet

However, when I try to build it by dropping into the mxnet directory and running make as follows:

make -j $(nproc) USE_OPENCV=1 USE_BLAS=openblas USE_CUDA=1 USE_CUDA_PATH=/usr/local/cuda USE_CUDNN=1

I get errors indicating that cblas.h cannot be found - e.g.

/home/me/mxnet/3rdparty/mshadow/mshadow/./base.h:162:14: fatal error: cblas.h: No such file or directory
 #include <cblas.h>

The output of make is a series of nvcc compilation commands that seem to be as follows:

/usr/local/cuda/bin/nvcc -std=c++11 -Xcompiler -D_FORCE_INLINES -O3 -ccbin /home/me/anaconda2/envs/deepnets/bin/x86_64-conda_cos6-linux-gnu-c++  -gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=[sm_75,compute_75] --fatbin-options -compress-all -Xcompiler "-DMSHADOW_FORCE_STREAM -Wall -Wsign-compare -O3 -DNDEBUG=1 -I/home/me/mxnet/3rdparty/mshadow/ -I/home/me/mxnet/3rdparty/dmlc-core/include -fPIC -I/home/me/mxnet/3rdparty/tvm/nnvm/include -I/home/me/mxnet/3rdparty/dlpack/include -I/home/me/mxnet/3rdparty/tvm/include -Iinclude -funroll-loops -Wno-unused-parameter -Wno-unknown-pragmas -Wno-unused-local-typedefs -msse3 -mf16c -I/usr/local/cuda/include -DMSHADOW_USE_CBLAS=1 -DMSHADOW_USE_MKL=0 -I/home/me/mxnet/3rdparty/mkldnn/build/install/include -DMSHADOW_RABIT_PS=0 -DMSHADOW_DIST_PS=0 -DMSHADOW_USE_PASCAL=0 -DMXNET_USE_MKLDNN=1 -DUSE_MKL=1 -I/home/me/mxnet/src/operator/nn/mkldnn/ -I/home/me/mxnet/3rdparty/mkldnn/build/install/include -DMXNET_USE_OPENCV=1 -I/usr/local/include/opencv -I/usr/local/include -fopenmp -DMXNET_USE_OPERATOR_TUNING=1 -DMXNET_USE_LAPACK -DMSHADOW_USE_CUDNN=1  -I/home/me/mxnet/3rdparty/cub -DMXNET_ENABLE_CUDA_RTC=1 -DMXNET_USE_NCCL=0 -DMXNET_USE_LIBJPEG_TURBO=0" --generate-dependencies -MT build/src/operator/tensor/elemwise_unary_op_trig_gpu.o src/operator/tensor/elemwise_unary_op_trig.cu >build/src/operator/tensor/elemwise_unary_op_trig_gpu.d

If I break out (what I believe) to be all the include directories, I can pick out these directories:

-I/home/me/mxnet/3rdparty/mshadow/
-I/home/me/mxnet/3rdparty/dmlc-core/include -fPIC
-I/home/me/mxnet/3rdparty/tvm/nnvm/include
-I/home/me/mxnet/3rdparty/dlpack/include
-I/home/me/mxnet/3rdparty/tvm/include
-Iinclude -funroll-loops -Wno-unused-parameter -Wno-unknown-pragmas -Wno-unused-local-typedefs -msse3 -mf16c
-I/usr/local/cuda/include -DMSHADOW_USE_CBLAS=1 -DMSHADOW_USE_MKL=0
-I/home/me/mxnet/3rdparty/mkldnn/build/install/include -DMSHADOW_RABIT_PS=0 -DMSHADOW_DIST_PS=0 -DMSHADOW_USE_PASCAL=0 -DMXNET_USE_MKLDNN=1 -DUSE_MKL=1
-I/home/me/mxnet/src/operator/nn/mkldnn/
-I/home/me/mxnet/3rdparty/mkldnn/build/install/include -DMXNET_USE_OPENCV=1
-I/usr/local/include/opencv
-I/usr/local/include -fopenmp -DMXNET_USE_OPERATOR_TUNING=1 -DMXNET_USE_LAPACK -DMSHADOW_USE_CUDNN=1
-I/home/me/mxnet/3rdparty/cub

When I look for cblas.h, I get see these copies:

(deepnets) me@Chanticleer:~/mxnet$ sudo find /home -name 'cblas.h'
[sudo] password for me: 
/home/me/anaconda2/pkgs/openblas-0.3.3-h9ac9557_1001/include/cblas.h
/home/me/anaconda2/pkgs/lapack-3.6.1-ha44fe06_2/include/cblas.h
/home/me/anaconda2/pkgs/openblas-0.3.5-h9ac9557_1001/include/cblas.h
/home/me/anaconda2/envs/deepnets/include/cblas.h
/home/me/mxnet/julia/deps/cblas.h
(deepnets) me@Chanticleer:~/mxnet$ sudo find /usr -name 'cblas.h'
/usr/include/atlas/cblas.h
/usr/include/openblas/cblas.h
/usr/include/cblas.h

Clearly none of these seem to be where the Makefile has been directed to look. Do I need to change the Makefile in some way (how?), or do I need to put a copy of cblas.h in some other directory, or at least a soft link to it. If so, which directory and how?

Answer 1

If you're compiling with nvcc, you can just use the last include file, since it's in your include directory already, like this:

nvcc -I/usr/include <other stuff here...>

You don't have to do anything as drastic as manually copying; in fact I would really recommend against it because you're going to have to remember to re-do this when there's an update.

To answer your question about the makefile, if you were going to add the declaration I would say it's pretty common to see it being added to the $(CPPFLAGS) variable, which stands for C/C++ Preprocessor Flags. However, I would do this instead:

INCLUDE_DIRS  = -I/usr/include

COMPILE.cpp   = $(CXX) $(CXXFLAGS) $(CPPFLAGS) $(INCLUDE_DIRS) $(TARGET_ARCH) -c
OUTPUT_OPTION = -o $@    

%.o: %.cpp
    $(COMPILE.cpp) $(OUTPUT_OPTION) $^

The (proper) convention is for your users to be able to customize the $(CXX), $(CXXFLAGS), and $(CPPFLAGS) variables. Essentially, you want them to pick and customize their compiler, flags, and build options. Since you obviously need this include file, however (or you wouldn't have gotten this error), I would recommend setting it completely separately to avoid any confusion.

Anyways, good luck, man. I hope this helps.