I would like to request help in configuring my Windows 10 Professional PC so that cuDNN methods run correctly. Below I’ve included a very simple CUDA program, and a Makefile. As written, the program runs correctly. The program uses CUDA methods to copy an integer to GPU memory, “sets the device to be used for GPU executions”, and declares a cudnnHandle_t object. If I uncomment the cudnnCreate method, the program will ostensibly run, but will throw a CUDA_ERROR_NOT_FOUND when run by NSight Compute.
I’ve tried:
- Toggling the “INCLUDE_DIRECTORIES” line in the Makefile (as well as the appropriate recipe).
- Adding “C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\lib\x64” to system and user PATH.
- Uninstalling completely (without visiting the registry) and reinstalling Visual Studio 2019, the GeForce Game-Ready Drivers version 442.59, and CUDA 10.2.89_441.22. I’ve added cudnn64_7.dll to “C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\bin”, cudnn.h to “C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\include”, and cudnn.lib to “C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\include”.
- Installing Visual Studio 2017, CUDA 10.1 and 10.0, and appropriate versions of cuDNN. CUDA 9.2 doesn’t come with NSight Compute, and my 2080Ti is not compatible with NVIDIA Visual Profiler 9.2.
Would you recommend other steps?
NgNet.cu
#include <cudnn.h>
int main()
{
int integer_host = 7;
int* integer_GPU;
cudaMalloc(&integer_GPU, sizeof(int));
cudaMemcpy(integer_GPU, &integer_host, sizeof(int), cudaMemcpyHostToDevice);
cudaSetDevice(0);
cudnnHandle_t cudnnHandle;
//cudnnCreate(&cudnnHandle);
return 0;
}
Makefile
# SOURCE_FILES must be in format [FILENAME][SPACE][FILENAME][SPACE][FILENAME]...
# [FILENAME] cannot be in quotes.
SOURCE_FILES = NgNet.cu
# HEADER_FILES must be in format [FILENAME][SPACE][FILENAME][SPACE][FILENAME]...
# [FILENAME] cannot be in quotes.
#HEADER_FILES =
# INCLUDE_DIRECTORIES must be in format [PATH],[PATH],[PATH]...
# [PATH] may be written without quotes if [PATH] does not contain a space.
# For uniformity, I have enclosed all paths in quotes.
#INCLUDE_DIRECTORIES = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\include"
# LIBRARIES_TO_LINK must be in the format [PARTIAL_PATH],[PARTIAL_PATH],[PARTIAL_PATH]...
# [PARTIAL_PATH] consists of path to directory of library and prefix of library filename.
# [PARTIAL_PATH] may be written without quotes.
# For uniformity, I have enclosed all partial paths in quotes.
# [PARTIAL_PATH] cannot have spaces.
# My workaround was to add "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\lib\x64" to
# Control Panel -> System and Security -> System -> Change Settings -> Advanced ->
# Environmental Variables -> System Variables -> Path -> New.
LIBRARIES_TO_LINK = "cudnn"
# PREFIX_OF_OUTPUT_FILE_NAME may be written without quotes.
# I do not write filenames with spaces.
PREFIX_OF_NAME_OF_OUTPUT_FILE = "NgNet"
# Make output file whenever a source file or a header file changes.
# Use the nvcc compiler.
# Output severe, significant, production-quality, and informational compiler warnings.
# Reference include directories (i.e., directories of header files).
# Reference directories of libraries.
# Specify the prefix of the name of the output file.
#makeNgNet: $(SOURCE_FILES) $(HEADER_FILES)
nvcc --compiler-options -W4 $(SOURCE_FILES) -I $(INCLUDE_DIRECTORIES) -l $(LIBRARIES_TO_LINK) -o $(PREFIX_OF_NAME_OF_OUTPUT_FILE)
makeNgNet: $(SOURCE_FILES)
nvcc --compiler-options -W4 $(SOURCE_FILES) -l $(LIBRARIES_TO_LINK) -o $(PREFIX_OF_NAME_OF_OUTPUT_FILE)