Use CUdA and CudNN with Matlab

Hi,

I have installed Cuda and CudNN on Ubuntu 14.04. On my terminal the command 'nvcc --version' returns:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2016 NVIDIA Corporation
Built on Tue_Jan_10_13:22:03_CST_2017
Cuda compilation tools, release 8.0, V8.0.61

But when I try on Matlab R2018a:

results = coder.checkGpuInstall('full')

It returns:

Host CUDA Environment    : FAILED (The simple NVCC command 'nvcc --version' failed to execute successfully. GPU Code will not be able to be compiled. Ensure that the 'nvcc' program is installed with the CUDA SDK.)
  Runtime   : FAILED (No Runtime library can be found. Ensure that the libraries are installed with the CUDA SDK.)
  cuFFT     : FAILED (No cuFFT library can be found. Ensure that the libraries are installed with the CUDA SDK.)
  cuSOLVER  : FAILED (No cuSOLVER library can be found. Ensure that the libraries are installed with the CUDA SDK.)
  cuBLAS    : FAILED (No cuBLAS library can be found. Ensure that the libraries are installed with the CUDA SDK.)
------------------------------------------------------------------------
nvcc -c -rdc=true -Xcompiler -fPIC,-ansi,-fexceptions,-fno-omit-frame-pointer,-pthread -Xcudafe "--diag_suppress=unsigned_compare_with_zero --diag_suppress=useless_type_qualifier_on_return_type" -D_GNU_SOURCE -DMATLAB_MEX_FILE -Wno-deprecated-declarations    -arch sm_35  -I "/usr/local/MATLAB/R2018a/simulink/include" -I "/usr/local/MATLAB/R2018a/toolbox/shared/simtargets" -I "/tmp/tpe67dcabf_65c7_44ce_b747_7199284cf224/codegen/mex/gpuSimpleTest" -I "/tmp/tpe67dcabf_65c7_44ce_b747_7199284cf224" -I "./interface" -I "/usr/local/MATLAB/R2018a/extern/include" -I "." "gpuSimpleTest_data.cu"
/bin/sh: 1: nvcc: not found
gmake: *** [gpuSimpleTest_data.o] Error 127

------------------------------------------------------------------------
??? Build error: C compiler produced errors. See the Build Log for further details.

Code generation failed: View Error Report
Code Generation          : FAILED (Test GPU code generation failed with the following error 'emlc:compilationError'.)
Compatible GPU           : FAILED (The compute capability '3.0' of the selected GPU '1' is not supported by GPU Coder. Generated GPU MEX execution will not be available.)
cuDNN Environment        : FAILED (A 'NVIDIA_CUDNN' environment variable was not found. Set 'NVIDIA_CUDNN' to point to the root directory of a NVIDIA cuDNN installation.)
Jetson TK1 Environment   : FAILED (A 'NVIDIA_CUDA_TK1' environment variable was not found. This environment variable needs to be set for cross-compilation.)
Jetson TX1 Environment   : FAILED (A 'NVIDIA_CUDA_TX1' environment variable was not found. This environment variable needs to be set for cross-compilation.)
Jetson TX2 Environment   : FAILED (A 'NVIDIA_CUDA_TX2' environment variable was not found. This environment variable needs to be set for cross-compilation.)

results = 

    struct with fields:

          host: 0
           tk1: 0
           tx1: 0
           tx2: 0
           gpu: 0
       codegen: 0
      codeexec: 0
         cudnn: 0

Do you know how to make NIVIDIA softwares compatible with Matlab2018a ?

Answer 1

[Joss Knight]

Check some of these environment variables to make sure they're pointing to the right place.

[Pietro Cicuta]

Tried but failing! posted a detailed continuation of thread...

Answer 2

[Pietro Cicuta]

Hi - I have a similar problem. cuda-10.0 istalled, and cudnn 7.5 installed in Ubuntu 18.04. The two seem to work fine from the terminal. But in Matlab I am failing to get the cuDNN environment to be set successfully. I have the cudnn.h in usr/include.... this does not work.... I have the cudnn_v7.h in usr/include/x86_64-linux-gnu this also does not work, I get the error from coder.checkGpuInstall('full') that unable to find cuDNN header files.

This has cost me already hours of random searching and tweaking... has anyone seen this and solved it? Maybe I'm doing something trivially stupid.

[Pietro Cicuta]

I'm still lost.... and even more puzzled.

If I set

setenv('NVIDIA_CUDNN', '/usr')

then run

code.checkGpuInstall('full')

I get the error

CuDNN Environment : FAILED (Unable to find cuDNN library in directory 'usr/lib64'...)

but that was not the env I said! In fact, I don't have a folder usr/lib64.

my cudnn.h is in usr/include.

but if I write

setenv('NVIDIA_CUDNN', '/usr/include')

then the error is

CuDNN Environment : FAILED (Unable to find cuDNN library in directory 'usr/include/include'...)

(i.e. doubling up the include....)

what's going on?!

[Saurabh Sinalkar]

Hi, I am facing the same problem faced by Pietro Cicuta,can someone please help on this

[Pietro Cicuta]

Come on Mathworks please fix this! My CUDA and CUDNN work fine from the linux command line! Seems like a bug in the setenv function.

[Nithin M]

Hi,

Got the same problem but solved it. May be helpful to someone and hence posting

your cuda installation would be mostly in usr/local/cuda-version

You need to give this path in setenv

eg: setenv('NVIDIA_CUDNN,'/usr/local/cuda-11');

Answer 3

[Miguel Medellin]

I had a different but somewhat related problem to the one discussed here.

It would help if you could describe what the directory structure of your CUDA/CuDNN install looks like.

For example, my CUDA directory is located in /usr/local/cuda and it has this kind of directory structure:

ls /usr/local/cuda
LICENSE NVIDIA_SLA_cuDNN_Support.txt README bin doc extras include lib64 nvml nvvm share src targets tools version.txt

In my case, I would have run setenv('NVIDIA_CUDNN', '/usr/local/cuda') and it should have found the lib64 and include folders it assumes are present in that directory.

If you don't have that directory structure, it will probably fail to find it.

The reason it is looking for the lib64 folder is because it looking for the shared object library files:

ls /usr/local/cuda/lib64
libOpenCL.so            libcudnn.so.7          libcusolver.so          libnppicc_static.a     libnppim.so.9.0.176   libnvToolsExt.so.1
libOpenCL.so.1          libcudnn.so.7.6.0      libcusolver.so.9.0      libnppicom.so          libnppim_static.a     libnvToolsExt.so.1.0.0
libOpenCL.so.1.0        libcudnn_static.a      libcusolver.so.9.0.176  libnppicom.so.9.0      libnppist.so          libnvblas.so
libOpenCL.so.1.0.0      libcufft.so            libcusolver_static.a    libnppicom.so.9.0.176  libnppist.so.9.0      libnvblas.so.9.0
libaccinj64.so          libcufft.so.9.0        libcusparse.so          libnppicom_static.a    libnppist.so.9.0.176  libnvblas.so.9.0.176
libaccinj64.so.9.0      libcufft.so.9.0.176    libcusparse.so.9.0      libnppidei.so          libnppist_static.a    libnvblas.so.9.0.480
libaccinj64.so.9.0.176  libcufft_static.a      libcusparse.so.9.0.176  libnppidei.so.9.0      libnppisu.so          libnvgraph.so
libcublas.so            libcufftw.so           libcusparse_static.a    libnppidei.so.9.0.176  libnppisu.so.9.0      libnvgraph.so.9.0
libcublas.so.9.0        libcufftw.so.9.0       libnppc.so              libnppidei_static.a    libnppisu.so.9.0.176  libnvgraph.so.9.0.176
libcublas.so.9.0.176    libcufftw.so.9.0.176   libnppc.so.9.0          libnppif.so            libnppisu_static.a    libnvgraph_static.a
libcublas.so.9.0.480    libcufftw_static.a     libnppc.so.9.0.176      libnppif.so.9.0        libnppitc.so          libnvrtc-builtins.so
libcublas_device.a      libcuinj64.so          libnppc_static.a        libnppif.so.9.0.176    libnppitc.so.9.0      libnvrtc-builtins.so.9.0
libcublas_static.a      libcuinj64.so.9.0      libnppial.so            libnppif_static.a      libnppitc.so.9.0.176  libnvrtc-builtins.so.9.0.176
libcudadevrt.a          libcuinj64.so.9.0.176  libnppial.so.9.0        libnppig.so            libnppitc_static.a    libnvrtc.so
libcudart.so            libculibos.a           libnppial.so.9.0.176    libnppig.so.9.0        libnpps.so            libnvrtc.so.9.0
libcudart.so.9.0        libcurand.so           libnppial_static.a      libnppig.so.9.0.176    libnpps.so.9.0        libnvrtc.so.9.0.176
libcudart.so.9.0.176    libcurand.so.9.0       libnppicc.so            libnppig_static.a      libnpps.so.9.0.176    stubs
libcudart_static.a      libcurand.so.9.0.176   libnppicc.so.9.0        libnppim.so            libnpps_static.a
libcudnn.so             libcurand_static.a     libnppicc.so.9.0.176    libnppim.so.9.0        libnvToolsExt.so

The reason it is looking for the include folder is because it assumes cudnn.h among other things is in there:

ls /usr/local/cuda/include
CL                            curand_kernel.h                host_config.h                             nvToolsExtSync.h
builtin_types.h               curand_lognormal.h             host_defines.h                            nvblas.h
channel_descriptor.h          curand_mrg32k3a.h              library_types.h                           nvfunctional
common_functions.h            curand_mtgp32.h                math_constants.h                          nvgraph.h
cooperative_groups.h          curand_mtgp32_host.h           math_functions.h                          nvml.h
cooperative_groups_helpers.h  curand_mtgp32_kernel.h         math_functions.hpp                        nvrtc.h
crt                           curand_mtgp32dc_p_11213.h      math_functions_dbl_ptx3.h                 sm_20_atomic_functions.h
cuComplex.h                   curand_normal.h                math_functions_dbl_ptx3.hpp               sm_20_atomic_functions.hpp
cublas.h                      curand_normal_static.h         mma.h                                     sm_20_intrinsics.h
cublasXt.h                    curand_philox4x32_x.h          npp.h                                     sm_20_intrinsics.hpp
cublas_api.h                  curand_poisson.h               nppcore.h                                 sm_30_intrinsics.h
cublas_v2.h                   curand_precalc.h               nppdefs.h                                 sm_30_intrinsics.hpp
cuda.h                        curand_uniform.h               nppi.h                                    sm_32_atomic_functions.h
cudaEGL.h                     cusolverDn.h                   nppi_arithmetic_and_logical_operations.h  sm_32_atomic_functions.hpp
cudaGL.h                      cusolverRf.h                   nppi_color_conversion.h                   sm_32_intrinsics.h
cudaProfiler.h                cusolverSp.h                   nppi_compression_functions.h              sm_32_intrinsics.hpp
cudaVDPAU.h                   cusolverSp_LOWLEVEL_PREVIEW.h  nppi_computer_vision.h                    sm_35_atomic_functions.h
cuda_device_runtime_api.h     cusolver_common.h              nppi_data_exchange_and_initialization.h   sm_35_intrinsics.h
cuda_fp16.h                   cusparse.h                     nppi_filtering_functions.h                sm_60_atomic_functions.h
cuda_fp16.hpp                 cusparse_v2.h                  nppi_geometry_transforms.h                sm_60_atomic_functions.hpp
cuda_gl_interop.h             device_atomic_functions.h      nppi_linear_transforms.h                  sm_61_intrinsics.h
cuda_occupancy.h              device_atomic_functions.hpp    nppi_morphological_operations.h           sm_61_intrinsics.hpp
cuda_profiler_api.h           device_double_functions.h      nppi_statistics_functions.h               sobol_direction_vectors.h
cuda_runtime.h                device_double_functions.hpp    nppi_support_functions.h                  surface_functions.h
cuda_runtime_api.h            device_functions.h             nppi_threshold_and_compare_operations.h   surface_functions.hpp
cuda_surface_types.h          device_functions.hpp           npps.h                                    surface_indirect_functions.h
cuda_texture_types.h          device_functions_decls.h       npps_arithmetic_and_logical_operations.h  surface_indirect_functions.hpp
cuda_vdpau_interop.h          device_launch_parameters.h     npps_conversion_functions.h               surface_types.h
cudalibxt.h                   device_types.h                 npps_filtering_functions.h                texture_fetch_functions.h
cudnn.h                       driver_functions.h             npps_initialization.h                     texture_fetch_functions.hpp
cufft.h                       driver_types.h                 npps_statistics_functions.h               texture_indirect_functions.h
cufftXt.h                     dynlink_cuda.h                 npps_support_functions.h                  texture_indirect_functions.hpp
cufftw.h                      dynlink_cuda_cuda.h            nppversion.h                              texture_types.h
curand.h                      dynlink_cuviddec.h             nvToolsExt.h                              thrust
curand_discrete.h             dynlink_nvcuvid.h              nvToolsExtCuda.h                          vector_functions.h
curand_discrete2.h            fatBinaryCtl.h                 nvToolsExtCudaRt.h                        vector_functions.hpp
curand_globals.h              fatbinary.h                    nvToolsExtMeta.h                          vector_types.h

If your directory does not look like that at all - it might mean you installed CUDA/CuDNN in some unforseen way - it may require reinstalling it in some more conventional manner.

You say you don't have a lib64 folder - do you have a lib folder or lib32 folder instead? If you do, I would try and figure out why - but you might be able to try create the folder manually and copying the contents of your lib / lib32 folder and placing it there just to see if it works. If it doesn't, it seems to suggest there is something unconventional/incorrect about your installation.

Answer 4

[nabila elsawy]

Hi,

i have same problem with cudnn header

can you help me?

>> coder.checkGpuInstall('full')

Host CUDA Environment : PASSED

Runtime : PASSED

cuFFT : PASSED

cuSOLVER : PASSED

cuBLAS : PASSED

Code Generation : PASSED

Compatible GPU : PASSED

Code Execution : PASSED

cuDNN Environment : FAILED (No cuDNN header could be found in directory "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\cuDNN\"\include. Ensure that the cuDNN headers are installed with the specified cuDNN SDK.)

Jetson TK1 Environment : FAILED (Jetson Cross-compilation is not supported on this platform. It is only supported on Linux operating systems.)

Jetson TX1 Environment : FAILED (Jetson Cross-compilation is not supported on this platform. It is only supported on Linux operating systems.)

Jetson TX2 Environment : FAILED (Jetson Cross-compilation is not supported on this platform. It is only supported on Linux operating systems.)

ans =

struct with fields:

host: 1

tk1: 0

tx1: 0

tx2: 0

gpu: 1

codegen: 1

codeexec: 1

cudnn: 0

[Miguel Medellin]

I would double check your path since that is what the error is complaining about.

First verify that location is in fact correct (you should see a bunch of .h/.hpp files in the 'include' sub folder and perhaps some other folders)

If the directory exists, then double check how you are passing the directory to MATLAB - I notice in the output it says:

FAILED (No cuDNN header could be found in directory "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\cuDNN\ " \include.

It might the the double quotation mark between 'cuDNN\' and '\include' is throwing off how MATLAB searches for the path. Perhaps try using single quotes? (If I'm not mistaken the 'include' folder is appended to the path you provided which is why it appears like this)