dev-v2 branch build failure

[Dr-Noob]

Summary

Building dev-v2 with NVIDIA backend support fails.

Version

oneDNN dev-v2 (commit ad5176c)

Environment

Intel Xeon Gold. CentOS 8.2, clang 12.0 clang version 12.0.0 (https://github.com/intel/llvm 633691327b3f2d6a2a7a8321244682e91db78e7d).

Compiler obtained from https://github.com/intel/llvm and built succesfully with:

[drnoob@xeon sycl_workspace]$ python $DPCPP_HOME/llvm/buildbot/configure.py --cuda -t release
[drnoob@xeon sycl_workspace]$ python $DPCPP_HOME/llvm/buildbot/compile.py

Steps to reproduce

Tried to build as explained in README.md

[drnoob@xeon sycl_workspace]$ git clone https://github.com/oneapi-src/oneDNN && cd oneDNN
[drnoob@xeon oneDNN]$ git checkout dev-v2
[drnoob@xeon oneDNN]$ mkdir build/ && cd build/ 
[drnoob@xeon build]$ cp ../../build.sh .
[drnoob@xeon build]$ cat build.sh
export SYCL_HOME=~/sycl_workspace
export ONEDNN_HOME=$SYCL_HOME/oneDNN/
export DPCPP_HOME=$SYCL_HOME/llvm/build/bin/

export CC=$DPCPP_HOME/clang
export CXX=$DPCPP_HOME/clang++

cmake -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP \
      -DDNNL_GPU_VENDOR=NVIDIA -G Ninja \
      -DOPENCLROOT=$SYCL_HOME/llvm/build/lib/ \
      -DOpenCL_INCLUDE_DIR=$SYCL_HOME/llvm/build/include/sycl/CL \
      -DCUBLAS_INCLUDE_DIR=/path/cuda-11.0/targets/x86_64-linux/include/ \
      -DCUBLAS_LIBRARY=/path/cuda-11.0/targets/x86_64-linux/lib/ \
      -DCUDNN_INCLUDE_DIR=/path/cuda-11.0/include/ \
      -DCUDNN_LIBRARY=/path/cuda-11.0/lib64 \
      -DTBBROOT=$SYCL_HOME/tbb/oneapi-tbb-2021.1-beta10/ ..
[drnoob@xeon build]$ ./build.sh
[drnoob@xeon build]$ ninja

Observed behavior

[drnoob@xeon build]$ ninja
[313/845] Building CXX object src/gpu/nvidia/CMakeFiles/dnnl_sycl_nvidia.dir/cudnn_lrn.cpp.o
FAILED: src/gpu/nvidia/CMakeFiles/dnnl_sycl_nvidia.dir/cudnn_lrn.cpp.o 
/home/drnoob/sycl_workspace/llvm/build/bin/clang++  -DCUDA_NO_HALF -DDNNL_DLL -DDNNL_DLL_EXPORTS -DDNNL_ENABLE_MAX_CPU_ISA -DDNNL_X64=1 -DNGEN_CPP11 -DNGEN_NEO_INTERFACE -DNGEN_NO_OP_NAMES -DNGEN_SAFE -DNGEN_WINDOWS_COMPAT -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS -I/home/drnoob/sycl_workspace/tbb/oneapi-tbb-2021.1-beta10/include -I../include -Iinclude -I../src -I/home/drnoob/sycl_workspace/llvm/build/include/sycl/CL -I/usr/local/cuda/include -I/home/modules/cuda-11.0/include -I/home/modules/cuda-11.0/targets/x86_64-linux/include -std=c++11 -fsycl -fvisibility-inlines-hidden  -Wall -Wno-unknown-pragmas -w -fvisibility=internal -msse4.1 -fPIC -Wformat -Wformat-security -fstack-protector-all   -Wno-pass-failed  -O2 -DNDEBUG -D_FORTIFY_SOURCE=2 -MD -MT src/gpu/nvidia/CMakeFiles/dnnl_sycl_nvidia.dir/cudnn_lrn.cpp.o -MF src/gpu/nvidia/CMakeFiles/dnnl_sycl_nvidia.dir/cudnn_lrn.cpp.o.d -o src/gpu/nvidia/CMakeFiles/dnnl_sycl_nvidia.dir/cudnn_lrn.cpp.o -c ../src/gpu/nvidia/cudnn_lrn.cpp
In file included from ../src/gpu/nvidia/cudnn_lrn.cpp:18:
In file included from ../src/gpu/nvidia/cudnn_lrn.hpp:23:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl.hpp:11:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/ONEAPI/atomic.hpp:11:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/ONEAPI/atomic_accessor.hpp:11:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/ONEAPI/atomic_enums.hpp:12:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/access/access.hpp:10:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/common.hpp:11:
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/cl_ext_intel.h:431:9: error: unknown type name 'cl_properties'
typedef cl_properties cl_mem_properties_intel;
        ^
In file included from ../src/gpu/nvidia/cudnn_lrn.cpp:18:
In file included from ../src/gpu/nvidia/cudnn_lrn.hpp:23:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl.hpp:11:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/ONEAPI/atomic.hpp:11:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/ONEAPI/atomic_accessor.hpp:11:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/ONEAPI/atomic_enums.hpp:12:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/access/access.hpp:10:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/common.hpp:121:
In file included from /home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/exception.hpp:15:
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:228:7: error: use of undeclared identifier 'CL_DEVICE_QUEUE_ON_DEVICE_PROPERTIES'
      CL_DEVICE_QUEUE_ON_DEVICE_PROPERTIES,
      ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:229:45: error: use of undeclared identifier 'CL_DEVICE_QUEUE_ON_HOST_PROPERTIES'
  PI_DEVICE_INFO_QUEUE_ON_HOST_PROPERTIES = CL_DEVICE_QUEUE_ON_HOST_PROPERTIES,
                                            ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:233:31: error: use of undeclared identifier 'CL_DEVICE_IL_VERSION_KHR'
  PI_DEVICE_INFO_IL_VERSION = CL_DEVICE_IL_VERSION_KHR,
                              ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:251:39: error: use of undeclared identifier 'CL_DEVICE_MAX_NUM_SUB_GROUPS'
  PI_DEVICE_INFO_MAX_NUM_SUB_GROUPS = CL_DEVICE_MAX_NUM_SUB_GROUPS,
                                      ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:253:7: error: use of undeclared identifier 'CL_DEVICE_SUB_GROUP_INDEPENDENT_FORWARD_PROGRESS'
      CL_DEVICE_SUB_GROUP_INDEPENDENT_FORWARD_PROGRESS,
      ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:287:34: error: use of undeclared identifier 'CL_QUEUE_DEVICE_DEFAULT'
  PI_QUEUE_INFO_DEVICE_DEFAULT = CL_QUEUE_DEVICE_DEFAULT,
                                 ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:290:24: error: use of undeclared identifier 'CL_QUEUE_SIZE'
  PI_QUEUE_INFO_SIZE = CL_QUEUE_SIZE
                       ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:334:34: error: use of undeclared identifier 'CL_KERNEL_MAX_SUB_GROUP_SIZE_FOR_NDRANGE'
  PI_KERNEL_MAX_SUB_GROUP_SIZE = CL_KERNEL_MAX_SUB_GROUP_SIZE_FOR_NDRANGE,
                                 ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:335:34: error: use of undeclared identifier 'CL_KERNEL_MAX_NUM_SUB_GROUPS'; did you mean 'PI_KERNEL_MAX_SUB_GROUP_SIZE'?
  PI_KERNEL_MAX_NUM_SUB_GROUPS = CL_KERNEL_MAX_NUM_SUB_GROUPS,
                                 ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:334:3: note: 'PI_KERNEL_MAX_SUB_GROUP_SIZE' declared here
  PI_KERNEL_MAX_SUB_GROUP_SIZE = CL_KERNEL_MAX_SUB_GROUP_SIZE_FOR_NDRANGE,
  ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:336:38: error: use of undeclared identifier 'CL_KERNEL_COMPILE_NUM_SUB_GROUPS'; did you mean 'PI_KERNEL_MAX_NUM_SUB_GROUPS'?
  PI_KERNEL_COMPILE_NUM_SUB_GROUPS = CL_KERNEL_COMPILE_NUM_SUB_GROUPS,
                                     ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:335:3: note: 'PI_KERNEL_MAX_NUM_SUB_GROUPS' declared here
  PI_KERNEL_MAX_NUM_SUB_GROUPS = CL_KERNEL_MAX_NUM_SUB_GROUPS,
  ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:374:30: error: use of undeclared identifier 'CL_COMMAND_SVM_FREE'
  PI_COMMAND_TYPE_SVM_FREE = CL_COMMAND_SVM_FREE,
                             ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:375:32: error: use of undeclared identifier 'CL_COMMAND_SVM_MEMCPY'
  PI_COMMAND_TYPE_SVM_MEMCPY = CL_COMMAND_SVM_MEMCPY,
                               ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:376:33: error: use of undeclared identifier 'CL_COMMAND_SVM_MEMFILL'
  PI_COMMAND_TYPE_SVM_MEMFILL = CL_COMMAND_SVM_MEMFILL,
                                ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:377:29: error: use of undeclared identifier 'CL_COMMAND_SVM_MAP'
  PI_COMMAND_TYPE_SVM_MAP = CL_COMMAND_SVM_MAP,
                            ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:378:31: error: use of undeclared identifier 'CL_COMMAND_SVM_UNMAP'
  PI_COMMAND_TYPE_SVM_UNMAP = CL_COMMAND_SVM_UNMAP
                              ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:405:33: error: use of undeclared identifier 'CL_ABGR'
  PI_IMAGE_CHANNEL_ORDER_ABGR = CL_ABGR,
                                ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:444:37: error: use of undeclared identifier 'CL_SAMPLER_MIP_FILTER_MODE'; did you mean 'PI_SAMPLER_INFO_FILTER_MODE'?
  PI_SAMPLER_INFO_MIP_FILTER_MODE = CL_SAMPLER_MIP_FILTER_MODE,
                                    ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:443:3: note: 'PI_SAMPLER_INFO_FILTER_MODE' declared here
  PI_SAMPLER_INFO_FILTER_MODE = CL_SAMPLER_FILTER_MODE,
  ^
/home/drnoob/sycl_workspace/llvm/build/bin/../include/sycl/CL/sycl/detail/pi.h:445:29: error: use of undeclared identifier 'CL_SAMPLER_LOD_MIN'
  PI_SAMPLER_INFO_LOD_MIN = CL_SAMPLER_LOD_MIN,
                            ^
fatal error: too many errors emitted, stopping now [-ferror-limit=]
20 errors generated.

Expected behavior

Expected a successful build.

[vpirogov]

@Dr-Noob, this looks like an issue with DPC++ Compiler. See discussion in intel/llvm#2617.

[dzarukin]

Sorry for the mess, answered in a wrong tracker. So I've added the work around information in the original issue which we use on our side to make it work out-of-the-box. This should resolve OpenCL build stuff once DOPENCLROOT=/path/to/compiler/root/folder (which is /compiler) is passed to CMake line.

Thanks.

[Dr-Noob]

Problem solved. Build completes successfully with your instructions. What I did was:

[drnoob@xeon sycl_workspace]$ ln -s ~/sycl_workspace/llvm/build/include/sycl/CL ~/sycl_workspace/llvm/build/include/

but also changed my build configuration because DOPENCLROOT was wrong. The working build.sh:

[drnoob@xeon build]$ cat build.sh
export SYCL_HOME=~/sycl_workspace
export ONEDNN_HOME=$SYCL_HOME/oneDNN/
export DPCPP_HOME=$SYCL_HOME/llvm/build/bin/

mkdir -p $ONEDNN_HOME/install/

export CC=$DPCPP_HOME/clang
export CXX=$DPCPP_HOME/clang++

cmake -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP \
      -DDNNL_GPU_VENDOR=NVIDIA -G Ninja \
      -DOPENCLROOT=$SYCL_HOME/llvm/build/ \
      -DCUDA_TOOLKIT_ROOT_DIR=/path/cuda-10.2 \
      -DCUBLAS_INCLUDE_DIR=/path/cuda-10.2/targets/x86_64-linux/include/ \
      -DCUBLAS_LIBRARY=/path/cuda-10.2/targets/x86_64-linux/lib/libcublas.so \
      -DCUDNN_INCLUDE_DIR=/path/cuda-10.2/include/ \
      -DCUDNN_LIBRARY=/path/cuda-10.2/targets/x86_64-linux/lib/libcudnn.so.7 \
      -DCMAKE_CXX_FLAGS=-Wl,-rpath,/path/cuda-10.2/targets/x86_64-linux/lib/ \
      -DCMAKE_C_FLAGS=-Wl,-rpath,/path/cuda-10.2/targets/x86_64-linux/lib/ \
      -DCMAKE_INSTALL_PREFIX=$ONEDNN_HOME/install/ ..

ninja

cmake --build . --target install

Note that I'm not using CUDA 11.0, but 10.2 instead. At first I installed cuDNN 8.5 and my build failed with:

[338/862] Building CXX object src/gpu/nvidia/CMakeFiles/dnnl_sycl_nvidia.dir/cudnn_inner_product.cpp.o
FAILED: src/gpu/nvidia/CMakeFiles/dnnl_sycl_nvidia.dir/cudnn_inner_product.cpp.o 
/home/drnoob/sycl_workspace/llvm/build/bin/clang++  -DCUDA_NO_HALF -DDNNL_DLL -DDNNL_DLL_EXPORTS -DDNNL_ENABLE_MAX_CPU_ISA -DDNNL_X64=1 -DNGEN_CPP11 -DNGEN_NEO_INTERFACE -DNGEN_NO_OP_NAMES -DNGEN_SAFE -DNGEN_WINDOWS_COMPAT -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS -I/opt/intel/compilers_and_libraries_2020.2.254/linux/tbb/include -I../include -Iinclude -I../src -I/home/drnoob/sycl_workspace/llvm/build/include -I/home/modules/cuda-11.0/include -I/home/modules/cuda-11.0/targets/x86_64-linux/include -Wl,-rpath,/home/modules/cuda-10.2/targets/x86_64-linux/lib/ -std=c++11 -fsycl -fvisibility-inlines-hidden  -Wall -Wno-unknown-pragmas -w -fvisibility=internal -msse4.1 -fPIC -Wformat -Wformat-security -fstack-protector-all   -Wno-pass-failed  -O2 -DNDEBUG -D_FORTIFY_SOURCE=2 -MD -MT src/gpu/nvidia/CMakeFiles/dnnl_sycl_nvidia.dir/cudnn_inner_product.cpp.o -MF src/gpu/nvidia/CMakeFiles/dnnl_sycl_nvidia.dir/cudnn_inner_product.cpp.o.d -o src/gpu/nvidia/CMakeFiles/dnnl_sycl_nvidia.dir/cudnn_inner_product.cpp.o -c ../src/gpu/nvidia/cudnn_inner_product.cpp
In file included from ../src/gpu/nvidia/cudnn_inner_product.cpp:19:
In file included from ../src/gpu/nvidia/cudnn_conv_inner_product.hpp:28:
../src/gpu/nvidia/cudnn_conv_inner_product_impl.hpp:283:13: error: unknown type name 'cudnnConvolutionFwdPreference_t'
            cudnnConvolutionFwdPreference_t algo_pref
            ^
../src/gpu/nvidia/cudnn_conv_inner_product_impl.hpp:284:23: error: use of undeclared identifier 'CUDNN_CONVOLUTION_FWD_PREFER_FASTEST'
                    = CUDNN_CONVOLUTION_FWD_PREFER_FASTEST;
                      ^
../src/gpu/nvidia/cudnn_conv_inner_product_impl.hpp:495:9: error: unknown type name 'cudnnConvolutionBwdDataPreference_t'; did you mean 'cudnnConvolutionBwdDataAlgoPerf_t'?
        cudnnConvolutionBwdDataPreference_t algo_pref
        ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
        cudnnConvolutionBwdDataAlgoPerf_t
/home/modules/cuda-11.0/include/cudnn_cnn_infer.h:304:3: note: 'cudnnConvolutionBwdDataAlgoPerf_t' declared here
} cudnnConvolutionBwdDataAlgoPerf_t;
  ^
In file included from ../src/gpu/nvidia/cudnn_inner_product.cpp:19:
In file included from ../src/gpu/nvidia/cudnn_conv_inner_product.hpp:28:
../src/gpu/nvidia/cudnn_conv_inner_product_impl.hpp:496:19: error: use of undeclared identifier 'CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST'
                = CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST;
                  ^
../src/gpu/nvidia/cudnn_conv_inner_product_impl.hpp:498:28: error: use of undeclared identifier 'cudnnGetConvolutionBackwardDataAlgorithm'; did you mean 'cudnnGetConvolutionBackwardDataAlgorithm_v7'?
        CUDNN_EXECUTE_FUNC(cudnnGetConvolutionBackwardDataAlgorithm, handle,
                           ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
                           cudnnGetConvolutionBackwardDataAlgorithm_v7
../src/gpu/nvidia/sycl_cuda_utils.hpp:403:20: note: expanded from macro 'CUDNN_EXECUTE_FUNC'
        auto err = name(__VA_ARGS__); \
                   ^
/home/modules/cuda-11.0/include/cudnn_cnn_infer.h:335:1: note: 'cudnnGetConvolutionBackwardDataAlgorithm_v7' declared here
cudnnGetConvolutionBackwardDataAlgorithm_v7(cudnnHandle_t handle,
^
In file included from ../src/gpu/nvidia/cudnn_inner_product.cpp:19:
In file included from ../src/gpu/nvidia/cudnn_conv_inner_product.hpp:28:
../src/gpu/nvidia/cudnn_conv_inner_product_impl.hpp:500:41: error: no viable conversion from 'cudnnConvolutionBwdDataAlgoPerf_t' to 'int'
                tensor_descs_[io::src], algo_pref, 0, &algo_);
                                        ^~~~~~~~~
../src/gpu/nvidia/sycl_cuda_utils.hpp:403:25: note: expanded from macro 'CUDNN_EXECUTE_FUNC'
        auto err = name(__VA_ARGS__); \
                        ^~~~~~~~~~~
/home/modules/cuda-11.0/include/cudnn_cnn_infer.h:340:55: note: passing argument to parameter 'requestedAlgoCount' here
                                            const int requestedAlgoCount,
                                                      ^
In file included from ../src/gpu/nvidia/cudnn_inner_product.cpp:19:
In file included from ../src/gpu/nvidia/cudnn_conv_inner_product.hpp:28:
../src/gpu/nvidia/cudnn_conv_inner_product_impl.hpp:650:9: error: unknown type name 'cudnnConvolutionBwdFilterPreference_t'; did you mean 'cudnnConvolutionBwdFilterAlgoPerf_t'?
        cudnnConvolutionBwdFilterPreference_t algo_pref
        ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
        cudnnConvolutionBwdFilterAlgoPerf_t
/home/modules/cuda-11.0/include/cudnn_cnn_train.h:87:3: note: 'cudnnConvolutionBwdFilterAlgoPerf_t' declared here
} cudnnConvolutionBwdFilterAlgoPerf_t;
  ^
In file included from ../src/gpu/nvidia/cudnn_inner_product.cpp:19:
In file included from ../src/gpu/nvidia/cudnn_conv_inner_product.hpp:28:
../src/gpu/nvidia/cudnn_conv_inner_product_impl.hpp:651:19: error: use of undeclared identifier 'CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST'
                = CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST;
                  ^
../src/gpu/nvidia/cudnn_conv_inner_product_impl.hpp:653:28: error: use of undeclared identifier 'cudnnGetConvolutionBackwardFilterAlgorithm'; did you mean 'cudnnGetConvolutionBackwardFilterAlgorithm_v7'?
        CUDNN_EXECUTE_FUNC(cudnnGetConvolutionBackwardFilterAlgorithm, handle,
                           ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
                           cudnnGetConvolutionBackwardFilterAlgorithm_v7
../src/gpu/nvidia/sycl_cuda_utils.hpp:403:20: note: expanded from macro 'CUDNN_EXECUTE_FUNC'
        auto err = name(__VA_ARGS__); \
                   ^
/home/modules/cuda-11.0/include/cudnn_cnn_train.h:118:1: note: 'cudnnGetConvolutionBackwardFilterAlgorithm_v7' declared here
cudnnGetConvolutionBackwardFilterAlgorithm_v7(cudnnHandle_t handle,
^
In file included from ../src/gpu/nvidia/cudnn_inner_product.cpp:19:
In file included from ../src/gpu/nvidia/cudnn_conv_inner_product.hpp:28:
../src/gpu/nvidia/cudnn_conv_inner_product_impl.hpp:655:31: error: no viable conversion from 'cudnnConvolutionBwdFilterAlgoPerf_t' to 'int'
                filter_desc_, algo_pref, 0, &algo_);
                              ^~~~~~~~~
../src/gpu/nvidia/sycl_cuda_utils.hpp:403:25: note: expanded from macro 'CUDNN_EXECUTE_FUNC'
        auto err = name(__VA_ARGS__); \
                        ^~~~~~~~~~~
/home/modules/cuda-11.0/include/cudnn_cnn_train.h:123:57: note: passing argument to parameter 'requestedAlgoCount' here
                                              const int requestedAlgoCount,
                                                        ^
10 errors generated.

I think that the README.md is misleading since says you need "cuDNN 7.6 or later", but I think that cuDNN 8 is not supported. For example, cudnnConvolutionFwdPreference_tis needed in oneDNN but got removed in cuDNN 8 (see API Reference). If this is the case, I would change README to state that cuDNN 8 won't work.

Closing the issue.