Posts on Visionbike - Personal Blog of CV | DSP | ML notes http://visionbike.github.io/posts/ Recent content in Posts on Visionbike - Personal Blog of CV | DSP | ML notes http://visionbike.github.io/images/apple-touch-icon.png http://visionbike.github.io/images/apple-touch-icon.png Hugo -- gohugo.io en Mon, 21 Aug 2023 14:57:01 +0800 Build OpenCV with CUDA on Windows 11 http://visionbike.github.io/posts/build-opencv-cuda-on-windows/ Mon, 21 Aug 2023 14:57:01 +0800 http://visionbike.github.io/posts/build-opencv-cuda-on-windows/ Since pre-built OpenCV binaries do not include CUDA modules, this post is a tutorial for building OpenCV with CUDA on Windows 11. The obvious dvantage of OpenCV CUDA is boosting performance of most functions, you can find evidence here. 1. Prerequisite There are a couples of softwares or libraries having been downloaded and installed before getting started: Install the Visual Studio Community 2022 and select Desktop development with C++ workload. Since pre-built OpenCV binaries do not include CUDA modules, this post is a tutorial for building OpenCV with CUDA on Windows 11. The obvious dvantage of OpenCV CUDA is boosting performance of most functions, you can find evidence here.

1. Prerequisite

There are a couples of softwares or libraries having been downloaded and installed before getting started:

  1. Install the Visual Studio Community 2022 and select Desktop development with C++ workload.

  2. Download the sources for OpenCV from GitHub by cloning the repositories (opencv and opencv_contrib).

OpenCV CUDA folder

After downloading, you can indicate the OpenCV’s version you want to. For the current version, you can run the following command in Command Prompt at OpenCV’s repositories:

git checkout tags/4.8.0

  1. Install the latest stable version (not release candidate -rc) of CMake.

  2. Install the latest version of NVIDIA CUDA Toolkit and add PATH. You can follow this tutorial.

The latest CUDA Toolkit version

At the time writting this post, the latest NVIDIA CUDA Toolkit version 12.2 still makes somes error when building with OpenCV 4.8.0, like error C2666: 'operator !=': overloaded functions have similar conversions. Therefore, if meeting such problem, try installing older version.

  1. Register an account and download the latest verson of NVIDIA DNN CUDA backend for the version of CUDA. Extract the downloaded .zip file and copy bin, include and lib directories to your CUDA installation directory, i.e., C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\vxx.x.

  2. Register an account and download the latest version of NVIDIA VIdeo Codec SDK. Extract the downloaded .zip file and copy the contents inside Interface and Lib to include and lib directories inside the CUDA installation directory.

  3. Optional - Download and install the latest version of Gstreamer.

  4. Download and install the lastest version of mambaforge to call OpenCV CUDA routines from python.

2. Create virtual environtment from Mambaforge

To bind OpenCV for python3 without any conflit package installation, you should create an new virtual environment for the installation.

mamba create -n opencv-cuda python=3.10

Activate the created environment and install numpy package.

mamba activate opencv-cuda
mamaba install numpy

3. Build OpenCV with CMake

Preparation

Create a build folder with your OpenCV extracted folders.

OpenCV CUDA folder

Edit the prioeiry of Python3 installation in OpenCVDetectPython.cmake file inside opencv-x.x.x/cmake folder.

Python3 Prior

Build GUI Build Configuration

Open Cmake GUI and provide the paths to the OpenCV and target build folders.

CMake GUI

Hit Configure and select x64 for the Optional platform for generator, then hit finish to start the configuration.

Once the configuration is done, edit the following parameters:

Parameter Value
CMAKE_INSTALL_PREFIX path of opencv installation
ENABLE_FAST_MATH ✅
WITH_CUDA ✅
BUILD_opencv_world ✅
BUILD_opencv_python3 ✅
OPENCV_DNN_CUDA ✅
OPENCV_EXTRA_MODULES_PATH path of modules directory in opencv_contrib-x.x.x
OPENCV_PYTHON3_VERSION ✅
PYTHON3_EXECUTABLE path of python3 executable in virtual env, i.e., C:/Users/ntthi/mambaforge/envs/opencv-cuda/python.exe
PYTHON3_INCLUDE_DIR path of include folder in the virtual env, i.e., C:/Users/ntthi/mambaforge/envs/opencv-cuda/include
PYTHON3_LIBRARY path of .lib file in the virtual env, i.e., C:/Users/ntthi/mambaforge/envs/opencv-cuda/libs/python310.lib
PYTHON3_NUMPY_INCLUDE_DIRS path of numpy in the virtual env, i.e., C:/Users/ntthi/mambaforge/envs/opencv-cuda/Lib/site-pakages/numpy/core/include
PYTHON3_PACKAGES_PATH path of site-packages in the virtual env, i.e., C:/Users/ntthi/mambaforge/envs/opencv-cuda/Lib/site-pakages

Note that the path separator hase to be “/” , not “".

Hit Configure again again and check edit more parameters:

Parameter Value
CUDA_FAST_MATH ✅
CUDA_ARCH_BIN version of computing capability, i.e., 8.6
WITH_CUBLAS ✅
WITH_CUDNN ✅
WITH_CUFFT ✅

The CUDA_ARCH_BIN corresponding to your GPU is the value found in the left column of the GPU support table. For instance, “8.6” fir the RTX 3070 Ti.

If you do not want to create shared lib and make sure the opencv python libraries is installed, edit the following parameters:

Parameter Value
BUILD_SHARED_LIBS 🔳
OPENCV_FORCE_PYTHON_LIBS ✅

Hit Configure at the last time and then hit Generate.

Build the project with Visual Studio

Open project OpenCV.sln created in the build folder. Go to Tools > Options…, then uncheck the last parameter in Projects and Solutions > Web Projects.

This setting may help to prevent the ImportError: DLL load failed while importing cv2: The specified module could not be found. error.

To build the OpenCV project, change Debug mode to Release. In the solution explorer expand CMakeTargets, right-click ALL_BUILD and select Build. This will take about an hour.

OpenCV build

Then repeat the step for INSTALL (below ALL_BUILD). Check for error in the two building steps. If everything is fine, you are done.

Check Installation and Troubleshooting

To verify the Python installation, activate the virtual environment for OpenCV install and try this code:

import cv2
print(cv2.__version__)
print(cv2.cuda.getCudaEnabledDeviceCount())

If it works, congratulations you are good to go!

If you meets the problem ImportError: DLL load failed while importing cv2: The specified module could not be found., it may lack the library’s binaries. One solution is to edit config.py in C:/Users/ntthi/mambaforge/envs/opencv-cuda/Lib/site-packages/cv2.

import os

BINARIES_PATHS = [
    os.path.join('C:/opencv-cuda-4.8.0', 'x64/vc17/bin'),
    os.path.join(os.getenv('CUDA_PATH', 'C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.1'), 'bin'),
    os.path.join('C:/gstreamer/1.0/msvc_x86_64', 'bin'),
] + BINARIES_PATHS

These binary paths are from installed OpenCV, CUDA and Gstreamer (if installed).

For other bugs and problems, I refer you to the chrismeunier and Bowley’s troubleshooting tutorial.

Reference

]]> Setting Up NVIDIA CUDA on WSL 2 http://visionbike.github.io/posts/setitng-up-nvidia-cuda-on-wsl2/ Thu, 13 Jul 2023 16:44:33 +0800 http://visionbike.github.io/posts/setitng-up-nvidia-cuda-on-wsl2/ Windows Subsystem for Linux (WSL) is a compatibility layer provided by Microsoft that allows you to run a Linux environment directly on WIndows. It will bring Linux applications, utilities and Bash command-line tool directly on Windows without the overhead of traditional virtual machine or dualboot setup. 1. Prequisites The Windows must be upgraded to Windows 10 version 2004 and higher or Windows 11 to install WSL by a single command line. Windows Subsystem for Linux (WSL) is a compatibility layer provided by Microsoft that allows you to run a Linux environment directly on WIndows. It will bring Linux applications, utilities and Bash command-line tool directly on Windows without the overhead of traditional virtual machine or dualboot setup.

1. Prequisites

The Windows must be upgraded to Windows 10 version 2004 and higher or Windows 11 to install WSL by a single command line. Open Command Prompt in Administrator mode, then run the following command:

wsl --install --no-distribution

The command will setup WSL version 2 by default without any Linux distribution. Reboot the machine to apply changes.

Open a Command Prompt window and install latest version Ubuntu distro:

wsl --install -d Ubuntu

The installation takes a few minutes and you need to create user name and password.

WSL-Ubuntu Installation

Display available WSL Linux distros

If you want to install another Linux distro, you can run following command to display a list of available WSL Linux distros:

wsl --list --online

For the first start, we should to install several Linux updates. The process will depend on the speed of the internet, so be patient if it slow!

sudo apt update -y && sudo apt upgrade -y
Issue: /usr/lib/wsl/lib/libcuda.so.1 is not a symbolic link

If your machine has integrated NVIDIA GPU, you may meet the /usr/lib/wsl/lib/libcuda.so.1 is not a symbolic link issue when updating the system.

In the WSL Ubuntu side, run following commands:

cd /usr/lib/wsl/lib
sudo rm libcuda.so
sudo rm libcuda.so.1
sudo ln -s libcuda.so.1.1 libcuda.so
sudo ln -s libcuda.so.1.1 libcuda.so.1

Now reinstall libc-bin package.

sudo apt reinstall libc-bin

In the Windows side, run following comands in Administrator mode:

cd C:\Windows\System32\lxss\lib
del libcuda.so
del libcuda.so.1
mklink libcuda.so libcuda.so.1.1
mklink libcuda.so.1 libcuda.so.1.1

Reboot the machine and run the sudo ldconfig command in WSL Ubuntu side to verify the problem is solved.

2. Installing Mamba

Mamba is a command-line interfacer (CLI) to manage conda’s environemts. For mamba configuration, please refer to conda documentation.

For the fresh installation, you can install Miniforge distribution >= Miniforge3-22.3.1.0. Miniforge comes with the popular conda-forge channel preconfigured, but you can modify the configuration to use any channel you like.

For Unix-like platforms (MAC OS and Linux), download the installer using curl or wget or your favorite program and run the script

curl -L -O "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
bash Miniforge3-$(uname)-$(uname -m).sh

After successful installation, you can use use the mamba commands as described in this user guide.

Working with Conda Environment

The command using mamba is similar to the conda command.

  • Create new environment
mamba create -n <ENV_NAME> python=<VERSION>
  • Activate an environment
mamba activate <ENV_NAME>
  • Deactivate environment
mamba deactivate
  • Delete an environment
mamba env remove -n <ENV_NAME>
  • Show all created environments
mamba env list

Remember to activate an environment first, do not install any packages in your base environment!

  • Install python packages
mamba install <PKG_NAME>[=VERSION] [-c <CHANNEL_NAME>]

When installing a package, you can optionally indicate specific additional channel that the packages are posted by community. The conda-forge is one of most common additional channels.

  • Delete packages
mamba remove <PKG_NAME>
  • Show all installed packages in the virtual environment
mamba list [-n <ENV_NAME>]

2. Installing NVIDIA CUDA Toolkit for WSL 2

Note that you only need to install NVIDIA Driver for Windows. Do not install any Linux Driver in WSL.

The latest NVIDIA Windows GPU Driver will fully support WSL 2. With CUDA support in the driver, existing applications compiled on a Linux system for the same target GPU can run unmodified within the WSL environment. Once NVIDIA Windows GPU Driver is installed in the system, the CUDA driver will be stubbed inside the WSL 2 as libcuda.so. Therefore, you only use a separate CUDA Toolkit for WSL 2 which does not contain the NVIDIA Linux GPU Driver.

To get started with running CUDA on WSL 2, you need to instal NVIDIA Driver on Windows 11 with a compatible GeForce or NVIDIA RTX/Quadro card from here. Once a Windows NVIDIA GPU driver is installed on the system, CUDA becomes available within WSL 2. The CUDA driver installed on Windows host will be stubbed inside the WSL 2 as libcuda.so.

However, the NVIDIA CUDA Toolkit for WSL need to be install for further use.

First, remove the old GPG key:

sudo apt-key del 7fa2af80

Install the NVIDIA CUDA Toolkit for WSL-Ubuntu by following the step in this link

Download CUDA Toolkit for WSL 2.

NVIDIA CUDA Toolkit Installation

Some actions must be taken after the installation before the CUDA Toolkit and Driver can be used. The PATH and LD_LIBRARY_PATH variables need to be updated.

Open ~/.bashrc by running nano ~/.bashrc, add two lines at the end of the file:

export PATH=/usr/local/cuda-12.4/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda-12.4/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

Then you run source ~/.bashrc for activation changes without reboot.

Verify CUDA Toolkit Installation

You can verify the installation by running the following command:

nvcc --version

If the installation was successful, you should see the CUDA version information displayed.

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2024 NVIDIA Corporation
Built on Thu_Mar_28_02:18:24_PDT_2024
Cuda compilation tools, release 12.4, V12.4.131
Build cuda_12.4.r12.4/compiler.34097967_0

It is important to verify that the NVIDIA CUDA Toolkit can find and communicate correctly with CUDA-compatible hardware. To do this, you need to compile and run some sample programs.

CUDA samples are located in https://github.com/nvidia/cuda-samples. To use the samples, clone the project, build the samples in cuda-samples directory using make command and run them following the instruction on the Github page.

To verify a correct configuration of the hardware and software, it is highly recommended that you build and run the deviceQuery sample program.

./deviceQuery

If CUDA is installed and configured correctly, the output should look similar as below:

./deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "NVIDIA GeForce RTX 3070 Ti Laptop GPU"
  CUDA Driver Version / Runtime Version          12.4 / 12.4
  CUDA Capability Major/Minor version number:    8.6
  Total amount of global memory:                 8192 MBytes (8589410304 bytes)
  (046) Multiprocessors, (128) CUDA Cores/MP:    5888 CUDA Cores
  GPU Max Clock rate:                            1410 MHz (1.41 GHz)
  Memory Clock rate:                             7001 Mhz
  Memory Bus Width:                              256-bit
  L2 Cache Size:                                 4194304 bytes
  Maximum Texture Dimension Size (x,y,z)         1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
  Maximum Layered 1D Texture Size, (num) layers  1D=(32768), 2048 layers
  Maximum Layered 2D Texture Size, (num) layers  2D=(32768, 32768), 2048 layers
  Total amount of constant memory:               65536 bytes
  Total amount of shared memory per block:       49152 bytes
  Total shared memory per multiprocessor:        102400 bytes
  Total number of registers available per block: 65536
  Warp size:                                     32
  Maximum number of threads per multiprocessor:  1536
  Maximum number of threads per block:           1024
  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
  Max dimension size of a grid size    (x,y,z): (2147483647, 65535, 65535)
  Maximum memory pitch:                          2147483647 bytes
  Texture alignment:                             512 bytes
  Concurrent copy and kernel execution:          Yes with 1 copy engine(s)
  Run time limit on kernels:                     Yes
  Integrated GPU sharing Host Memory:            No
  Support host page-locked memory mapping:       Yes
  Alignment requirement for Surfaces:            Yes
  Device has ECC support:                        Disabled
  Device supports Unified Addressing (UVA):      Yes
  Device supports Managed Memory:                Yes
  Device supports Compute Preemption:            Yes
  Supports Cooperative Kernel Launch:            Yes
  Supports MultiDevice Co-op Kernel Launch:      No
  Device PCI Domain ID / Bus ID / location ID:   0 / 1 / 0
  Compute Mode:
     < Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 12.4, CUDA Runtime Version = 12.4, NumDevs = 1
Result = PASS

By running the bandwidthTest program, you can ensure that the system and CUDA-capable device are able to communicate correctly.

./bandwidthText

The output shoud be here.

[CUDA Bandwidth Test] - Starting...
Running on...

 Device 0: NVIDIA GeForce RTX 3070 Ti Laptop GPU
 Quick Mode

 Host to Device Bandwidth, 1 Device(s)
 PINNED Memory Transfers
   Transfer Size (Bytes)        Bandwidth(GB/s)
   32000000                     24.5

 Device to Host Bandwidth, 1 Device(s)
 PINNED Memory Transfers
   Transfer Size (Bytes)        Bandwidth(GB/s)
   32000000                     26.3

 Device to Device Bandwidth, 1 Device(s)
 PINNED Memory Transfers
   Transfer Size (Bytes)        Bandwidth(GB/s)
   32000000                     323.6

Result = PASS

NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.

For CUDA graphic programs, WSL 2 currently does not support GL, Vulcan. Hence, you cannot run CUDA graphic programs.

2. Installing NVIDIA CuDNN for WSL 2

Since cuDNN version 9 can exist independently with previous version cuDNN, you can refer here and follow the steps for your target platform.

NVIDIA cuDNN Installation

Verify cuDNN Installation

To verify cuDNN is install and is running properly, you can download and compile cuDNN samples. These samples can be downloaded by running:

sudo apt install -y libcudnn9-samples

The samples is located in /usr/src/cudnn_samples_v9 directory, copy the directory to $HOME and set the permission.

sudo cp -r cudnn_samples_v9/ $HOME
sudo chmod -R a+rwx $HOME/cudnn_samples_v9

For the mnistCUDNN sample, install FreeImage package and compile the sample.

sudo apt install -y libfreeimage3 libfreeimage-dev
cd $HOME/cudnn_samples_v9/mnistCUDNN
make clean && make
./mnistCUDNN

The result should be as:

Executing: mnistCUDNN
cudnnGetVersion() : 90100 , CUDNN_VERSION from cudnn.h : 90100 (9.1.0)
Host compiler version : GCC 11.4.0

There are 1 CUDA capable devices on your machine :
device 0 : sms 46  Capabilities 8.6, SmClock 1410.0 Mhz, MemSize (Mb) 8191, MemClock 7001.0 Mhz, Ecc=0, boardGroupID=0
Using device 0

Testing single precision
Loading binary file data/conv1.bin
Loading binary file data/conv1.bias.bin
Loading binary file data/conv2.bin
Loading binary file data/conv2.bias.bin
Loading binary file data/ip1.bin
Loading binary file data/ip1.bias.bin
Loading binary file data/ip2.bin
Loading binary file data/ip2.bias.bin
Loading image data/one_28x28.pgm
Performing forward propagation ...
Testing cudnnGetConvolutionForwardAlgorithm_v7 ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: -1.000000 time requiring 178432 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: -1.000000 time requiring 184784 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: -1.000000 time requiring 2057744 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnFindConvolutionForwardAlgorithm ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: 0.009216 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: 0.009216 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: 0.096256 time requiring 178432 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: 0.115712 time requiring 2057744 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: 0.168960 time requiring 184784 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: 0.546016 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnGetConvolutionForwardAlgorithm_v7 ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: -1.000000 time requiring 2450080 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: -1.000000 time requiring 4656640 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: -1.000000 time requiring 1433120 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnFindConvolutionForwardAlgorithm ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: 0.035840 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: 0.058368 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: 0.066560 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: 0.093184 time requiring 1433120 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: 0.109600 time requiring 4656640 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: 0.192512 time requiring 2450080 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Resulting weights from Softmax:
0.0000000 0.9999399 0.0000000 0.0000000 0.0000561 0.0000000 0.0000012 0.0000017 0.0000010 0.0000000
Loading image data/three_28x28.pgm
Performing forward propagation ...
Testing cudnnGetConvolutionForwardAlgorithm_v7 ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: -1.000000 time requiring 178432 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: -1.000000 time requiring 184784 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: -1.000000 time requiring 2057744 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnFindConvolutionForwardAlgorithm ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: 0.008192 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: 0.009216 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: 0.026624 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: 0.061248 time requiring 2057744 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: 0.073728 time requiring 184784 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: 0.106496 time requiring 178432 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnGetConvolutionForwardAlgorithm_v7 ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: -1.000000 time requiring 2450080 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: -1.000000 time requiring 4656640 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: -1.000000 time requiring 1433120 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnFindConvolutionForwardAlgorithm ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: 0.031520 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: 0.038912 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: 0.066400 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: 0.091136 time requiring 1433120 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: 0.129024 time requiring 4656640 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: 0.145408 time requiring 2450080 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Resulting weights from Softmax:
0.0000000 0.0000000 0.0000000 0.9999288 0.0000000 0.0000711 0.0000000 0.0000000 0.0000000 0.0000000
Loading image data/five_28x28.pgm
Performing forward propagation ...
Resulting weights from Softmax:
0.0000000 0.0000008 0.0000000 0.0000002 0.0000000 0.9999820 0.0000154 0.0000000 0.0000012 0.0000006

Result of classification: 1 3 5

Test passed!

Testing half precision (math in single precision)
Loading binary file data/conv1.bin
Loading binary file data/conv1.bias.bin
Loading binary file data/conv2.bin
Loading binary file data/conv2.bias.bin
Loading binary file data/ip1.bin
Loading binary file data/ip1.bias.bin
Loading binary file data/ip2.bin
Loading binary file data/ip2.bias.bin
Loading image data/one_28x28.pgm
Performing forward propagation ...
Testing cudnnGetConvolutionForwardAlgorithm_v7 ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: -1.000000 time requiring 178432 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: -1.000000 time requiring 184784 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: -1.000000 time requiring 2057744 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnFindConvolutionForwardAlgorithm ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: 0.009216 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: 0.010240 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: 0.017408 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: 0.068512 time requiring 2057744 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: 0.070784 time requiring 178432 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: 0.103424 time requiring 184784 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnGetConvolutionForwardAlgorithm_v7 ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: -1.000000 time requiring 1433120 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: -1.000000 time requiring 1536 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: -1.000000 time requiring 64000 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: -1.000000 time requiring 4656640 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: -1.000000 time requiring 2450080 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnFindConvolutionForwardAlgorithm ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: 0.079872 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: 0.079872 time requiring 1536 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: 0.101376 time requiring 64000 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: 0.101440 time requiring 4656640 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: 0.105472 time requiring 2450080 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: 0.126976 time requiring 1433120 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Resulting weights from Softmax:
0.0000001 1.0000000 0.0000001 0.0000000 0.0000563 0.0000001 0.0000012 0.0000017 0.0000010 0.0000001
Loading image data/three_28x28.pgm
Performing forward propagation ...
Testing cudnnGetConvolutionForwardAlgorithm_v7 ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: -1.000000 time requiring 178432 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: -1.000000 time requiring 184784 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: -1.000000 time requiring 2057744 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnFindConvolutionForwardAlgorithm ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: 0.010240 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: 0.023552 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: 0.064480 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: 0.087008 time requiring 2057744 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: 0.103424 time requiring 184784 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: 0.138144 time requiring 178432 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnGetConvolutionForwardAlgorithm_v7 ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: -1.000000 time requiring 1433120 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: -1.000000 time requiring 1536 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: -1.000000 time requiring 64000 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: -1.000000 time requiring 4656640 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: -1.000000 time requiring 2450080 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Testing cudnnFindConvolutionForwardAlgorithm ...
^^^^ CUDNN_STATUS_SUCCESS for Algo 0: 0.083904 time requiring 0 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 7: 0.090112 time requiring 1433120 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 5: 0.094976 time requiring 4656640 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 4: 0.095232 time requiring 2450080 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 2: 0.111616 time requiring 64000 memory
^^^^ CUDNN_STATUS_SUCCESS for Algo 1: 0.186368 time requiring 1536 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 6: -1.000000 time requiring 0 memory
^^^^ CUDNN_STATUS_NOT_SUPPORTED for Algo 3: -1.000000 time requiring 0 memory
Resulting weights from Softmax:
0.0000000 0.0000000 0.0000000 1.0000000 0.0000000 0.0000714 0.0000000 0.0000000 0.0000000 0.0000000
Loading image data/five_28x28.pgm
Performing forward propagation ...
Resulting weights from Softmax:
0.0000000 0.0000008 0.0000000 0.0000002 0.0000000 1.0000000 0.0000154 0.0000000 0.0000012 0.0000006

Result of classification: 1 3 5

Test passed!

3. Installing NVIDIA TensorRT (Optional)

NVIDIA TensorRT is a C++ library that facilitates high-performance inference NVIDIA graphic processing units (GPUs). TensorRT takes a trained network, which consists of a network definition and a set of trained parameters, and produces a highly optimized runtime engine that performs inference for that network.

TensorRT provides APIs via C++ and Python that help to express deep learning model via the Network Definition API or load a pre-defined model via the ONNX parser that allow TensorRT to optimize and run them on the NVIDIA GPU. TensorRT also include optional high speed mixed precision capabilities with difference NVIDIA architectures.

You can download TensorRT local repo that matches the Ubuntu version and CPU architecture of your machine from here.

Install the Debian local package.

os="ubuntuxx04"
tag="10.x.x-cuda-x.x"
sudo dpkg -i nv-tensorrt-local-repo-${os}-${tag}_1.0-1_amd64.deb
sudo cp /var/nv-tensorrt-local-repo-${os}-${tag}/*-keyring.gpg /usr/share/keyrings/
sudo apt-get update

Replace ubuntuxx04, 10.x.x and cuda-x.x with your specific OS version, TensorRT version, and CUDA version. The full TensorRT C++ and Python runtimes can be installed by:

sudo apt install -y tensorrt
Verify TensorRT Installation

To verify the installation, run the command:

dpkg-query -W tensorrt

The ouput should be as following:

tensorrt        10.0.0.6-1+cuda12.4

The samples is located in /usr/src/tensorrt directory, copy the directory to $HOME and set the permission.

sudo cp -r tensorrt/ $HOME
sudo chmod -R a+rwx $HOME/tensorrt

For the sampleOnnxMNIST sample, compile the sample and run the program.

cd $HOME/tensorrt/samples/sampleOnnxMNIST
make clean && make
cd ../../bin
./sampleOnnxMNIST

The output should be as following:

&&&& RUNNING TensorRT.sample_onnx_mnist [TensorRT v100000] # ./sample_onnx_mnist
[04/15/2024-18:24:17] [I] Building and running a GPU inference engine for Onnx MNIST
[04/15/2024-18:24:17] [I] [TRT] [MemUsageChange] Init CUDA: CPU +19, GPU +0, now: CPU 21, GPU 1091 (MiB)
[04/15/2024-18:24:19] [I] [TRT] [MemUsageChange] Init builder kernel library: CPU +1765, GPU +312, now: CPU 1922, GPU 1403 (MiB)
[04/15/2024-18:24:19] [I] [TRT] ----------------------------------------------------------------
[04/15/2024-18:24:19] [I] [TRT] Input filename:   ../data/mnist/mnist.onnx
[04/15/2024-18:24:19] [I] [TRT] ONNX IR version:  0.0.3
[04/15/2024-18:24:19] [I] [TRT] Opset version:    8
[04/15/2024-18:24:19] [I] [TRT] Producer name:    CNTK
[04/15/2024-18:24:19] [I] [TRT] Producer version: 2.5.1
[04/15/2024-18:24:19] [I] [TRT] Domain:           ai.cntk
[04/15/2024-18:24:19] [I] [TRT] Model version:    1
[04/15/2024-18:24:19] [I] [TRT] Doc string:
[04/15/2024-18:24:19] [I] [TRT] ----------------------------------------------------------------
[04/15/2024-18:24:19] [I] [TRT] Local timing cache in use. Profiling results in this builder pass will not be stored.
[04/15/2024-18:24:21] [I] [TRT] Detected 1 inputs and 1 output network tensors.
[04/15/2024-18:24:21] [I] [TRT] Total Host Persistent Memory: 26272
[04/15/2024-18:24:21] [I] [TRT] Total Device Persistent Memory: 0
[04/15/2024-18:24:21] [I] [TRT] Total Scratch Memory: 0
[04/15/2024-18:24:21] [I] [TRT] [BlockAssignment] Started assigning block shifts. This will take 6 steps to complete.
[04/15/2024-18:24:21] [I] [TRT] [BlockAssignment] Algorithm ShiftNTopDown took 0.02021ms to assign 3 blocks to 6 nodes requiring 32256 bytes.
[04/15/2024-18:24:21] [I] [TRT] Total Activation Memory: 31744
[04/15/2024-18:24:21] [I] [TRT] Total Weights Memory: 26152
[04/15/2024-18:24:21] [I] [TRT] Engine generation completed in 1.46118 seconds.
[04/15/2024-18:24:21] [I] [TRT] [MemUsageStats] Peak memory usage of TRT CPU/GPU memory allocators: CPU 0 MiB, GPU 5 MiB
[04/15/2024-18:24:21] [I] [TRT] [MemUsageStats] Peak memory usage during Engine building and serialization: CPU: 3052 MiB
[04/15/2024-18:24:21] [I] [TRT] Loaded engine size: 0 MiB
[04/15/2024-18:24:21] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +0, now: CPU 0, GPU 0 (MiB)
[04/15/2024-18:24:21] [I] Input:
[04/15/2024-18:24:21] [I] @@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@%=#@@@@@%=%@@@@@@@@@@
@@@@@@@           %@@@@@@@@@
@@@@@@@           %@@@@@@@@@
@@@@@@@#:-#-.     %@@@@@@@@@
@@@@@@@@@@@@#    #@@@@@@@@@@
@@@@@@@@@@@@@    #@@@@@@@@@@
@@@@@@@@@@@@@:  :@@@@@@@@@@@
@@@@@@@@@%+==   *%%%%%%%%%@@
@@@@@@@@%                 -@
@@@@@@@@@#+.          .:-%@@
@@@@@@@@@@@*     :-###@@@@@@
@@@@@@@@@@@*   -%@@@@@@@@@@@
@@@@@@@@@@@*   *@@@@@@@@@@@@
@@@@@@@@@@@*   @@@@@@@@@@@@@
@@@@@@@@@@@*   #@@@@@@@@@@@@
@@@@@@@@@@@*   *@@@@@@@@@@@@
@@@@@@@@@@@*   *@@@@@@@@@@@@
@@@@@@@@@@@*   @@@@@@@@@@@@@
@@@@@@@@@@@*   @@@@@@@@@@@@@
@@@@@@@@@@@@+=#@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@

[04/15/2024-18:24:21] [I] Output:
[04/15/2024-18:24:21] [I]  Prob 0  0.0000 Class 0:
[04/15/2024-18:24:21] [I]  Prob 1  0.0000 Class 1:
[04/15/2024-18:24:21] [I]  Prob 2  0.0000 Class 2:
[04/15/2024-18:24:21] [I]  Prob 3  0.0000 Class 3:
[04/15/2024-18:24:21] [I]  Prob 4  0.0000 Class 4:
[04/15/2024-18:24:21] [I]  Prob 5  0.0000 Class 5:
[04/15/2024-18:24:21] [I]  Prob 6  0.0000 Class 6:
[04/15/2024-18:24:21] [I]  Prob 7  1.0000 Class 7: **********
[04/15/2024-18:24:21] [I]  Prob 8  0.0000 Class 8:
[04/15/2024-18:24:21] [I]  Prob 9  0.0000 Class 9:
[04/15/2024-18:24:21] [I]
&&&& PASSED TensorRT.sample_onnx_mnist [TensorRT v100000] # ./sample_onnx_mnist

If you are using TensorRT Python API, make sure CUDA-Python is installed in your system or your virtual environment.

mamba create -n tensorrt_env pip wheel
mamba activate tensorrt_env
  • Installing from PyPI, TensorRT runtime wheels
pip install cuda-python
pip install --pre --upgrade tensorrt
pip install --pre --upgrade tensorrt_lean tensorrt_dispatch

To verify the installation is working, use python code.

import tensorrt
print(tensorrt.__version__)
assert(tensorrt.Builder(tensorrt.Logger()))

import tensorrt_lean as trtl
print(trtl.__version__)
assert trtl.Runtime(trt.Logger())

import tensorrt_dispatch as trtd
print(trtd.__version__)
assert trtd.Runtime(trt.Logger())

The tensorrt/samples directory also provides the samples that utilize with the TensorRT, you can follow instructions from README.md to run samples.

3. Set up Python environment using Mambaforge

Python environment allows to manage separatelly different installations of Python and modules. It is useful when you have many projects running different version of Python and modules. It also help to manage installed modules for publish or reproduce.

There are different ways to create a python virtual environment, including built-in venv, Conda and Anaconda.

Conda is a packaging tool and installer that aims to handle library dependencies outside of the Python packages as well as the Python packages themselves. For non preinstalled package manager, Miniconda, an installation of Conda, will be a good option.

Anaconda is an installation of Conda that comes pre-loaded with a bunch of packages for scientific computing, i.e., numpy, matplotlib, scipy, etc. It also comes with IDE, Jupyter notebooks out of the box. This is helpful for beginers, but doesn’t give much control.

Mamba is a package manager which can be used with Python. Unlike Conda, it uses the C/C++ implementation to speed up the package installation. Read more about mamba in here. To install mamba, access its repo and pick the Mabaforge installer for your operating system.

Remember to run conda init at the end of your installation in your bash to activate the mamba command.

Mambaforge install

Conclusion

By following these steps and installing the required software, you will have an CUDA-ready environment in WSL for further machine learnin/deep learning applications. This environment will provide the necessary tools and libraries for GPU-accelerated computing and Python package management.

Reference

]]> Setting Up NVIDIA CUDA on Windows 11 http://visionbike.github.io/posts/setting-up-nvidia-cuda-on-windows-11/ Thu, 06 Jul 2023 12:41:36 +0800 http://visionbike.github.io/posts/setting-up-nvidia-cuda-on-windows-11/ In this post, I will cover how to setup NVIDIA CUDA on Windows 11. To ensure a smooth setup process, it is crucial to follow the following steps: NVIDIA Drivers. Microsoft Visual Studio 2022 Community. NVIDIA CUDA Toolkit. NVIDIA cuDNN. TensorRT (optional). Miniforge (optional). 1. System Requirements To use CUDA, make sure your machine has a CUDA-capable GPU inside and the Microsoft Windows 11 should be updated 21H2 version. You can verify if your machine has CUDA-supported GPU through Display Adapters section in the Windows Device Manager. In this post, I will cover how to setup NVIDIA CUDA on Windows 11. To ensure a smooth setup process, it is crucial to follow the following steps:

  • NVIDIA Drivers.
  • Microsoft Visual Studio 2022 Community.
  • NVIDIA CUDA Toolkit.
  • NVIDIA cuDNN.
  • TensorRT (optional).
  • Miniforge (optional).

1. System Requirements

To use CUDA, make sure your machine has a CUDA-capable GPU inside and the Microsoft Windows 11 should be updated 21H2 version.

You can verify if your machine has CUDA-supported GPU through Display Adapters section in the Windows Device Manager. Here you will find the vendor name and model of your GPU.

device manager

You also need to install Microsft Visual Studio 2022 Community as the native compilter for x86_64 application. The download link os the latest version is here and install the Desktop development with C++ workload.

Visual Studio Installation

If there is any NVIDIA CUDA Toolkit installed before, you need to uninstall before proceeding further, following these steps:

  1. Open the Settings > Apps > Installed Apps.
  2. Scroll down and find NVIDIA CUDA applications.

NVIDIA apps list

  1. Click to “…” button on the right and uninstall all NVIDIA GPU drivers and any associated software.

NVIDIA app uninstallation

Then you need to install NVIDIA driver to communicate your computer with NVIDIA devices. You can find the suitable driver from this website.

NVIDIA Driver Installation

After the installation is complete, reboot your system.

2. Installing Mamba

Mamba is a command-line interfacer (CLI) to manage conda’s environemts. For mamba configuration, please refer to conda documentation.

For the fresh installation, you can install Miniforge distribution >= Miniforge3-22.3.1.0. Miniforge comes with the popular conda-forge channel preconfigured, but you can modify the configuration to use any channel you like.

Installation

Follow the instaltion prompts, taking note of options to Create start menu shortcuts and Add Miniforge3 to my PATH environment variable.

Miniforge Installation

After successful installation, you can use use the mamba commands as described in this user guide.

Working with Conda Environment

The command using mamba is similar to the conda command.

  • Create new environment
mamba create -n &lt;ENV_NAME&gt; python=&lt;VERSION&gt;
  • Activate an environment
mamba activate &lt;ENV_NAME&gt;
  • Deactivate environment
mamba deactivate
  • Delete an environment
mamba env remove -n &lt;ENV_NAME&gt;
  • Show all created environments
mamba env list

Remember to activate an environment first, do not install any packages in your base environment!

  • Install python packages
mamba install &lt;PKG_NAME&gt;[=VERSION] [-c &lt;CHANNEL_NAME&gt;]

When installing a package, you can optionally indicate specific additional channel that the packages are posted by community. The conda-forge is one of most common additional channels.

  • Delete packages
mamba remove &lt;PKG_NAME&gt;
  • Show all installed packages in the virtual environment
mamba list [-n &lt;ENV_NAME&gt;]

Post-installation

After installation, you make sure that the Anaconda is not the default configured channel, seeing this.

DO NOT install anything into the base environment as this might break your installation. See here for details.

The most convenient way to use the Mamba will be via the Miniforge Prompt installed in the Start Menu.

3. Installing NVIDIA CUDA Toolkit

You can visit the NVIDIA Developer website for CUDA Toolkit TO get the latest version of NVIDIA CUDA Toolkit. For previous versions, you can check from the Archive of Previous CUDA Releases page.

On the dowload page, you choose the appropriate version based on your system.

Download CUDA Toolkit

Then, you locate the downloaded installer file and double-click on it to start the installation process. Follow the on-screen instructions provided by the installer.

NVIDIA CUDA Toolkit install

Once the installation is completed, you check environment variables CUDA_PATH and PATH to ensure that your system recognizes NVIDIA CUDA Toolkit.

NVIDIA CUDA Path 1

NVIDIA CUDA Path 2

Verify CUDA Toolkit Installation

You can verify the installation by running the following command in command prompt.

nvcc --version

If the installation was successful, you should see the CUDA version information displayed.

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2024 NVIDIA Corporation
Built on Tue_Feb_27_16:28:36_Pacific_Standard_Time_2024
Cuda compilation tools, release 12.4, V12.4.99
Build cuda_12.4.r12.4/compiler.33961263_0

It is important to verify that the NVIDIA CUDA Toolkit can find and communicate correctly with CUDA-compatible hardware. To do this, you need to compile and run some sample programs.

CUDA samples are located in https://github.com/nvidia/cuda-samples. To use the samples, clone the project, build the samples in cuda-samples directory using MVSC 2022 compiler and run them following the instruction on the Github page.

To verify a correct configuration of the hardware and software, it is highly recommended that you build and run the deviceQuery sample program.

deviceQuery.exe

If CUDA is installed and configured correctly, the output should look similar as below:

deviceQuery.exe Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "NVIDIA GeForce RTX 3070 Ti Laptop GPU"
  CUDA Driver Version / Runtime Version          12.4 / 12.4
  CUDA Capability Major/Minor version number:    8.6
  Total amount of global memory:                 8192 MBytes (8589410304 bytes)
  (046) Multiprocessors, (128) CUDA Cores/MP:    5888 CUDA Cores
  GPU Max Clock rate:                            1410 MHz (1.41 GHz)
  Memory Clock rate:                             7001 Mhz
  Memory Bus Width:                              256-bit
  L2 Cache Size:                                 4194304 bytes
  Maximum Texture Dimension Size (x,y,z)         1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
  Maximum Layered 1D Texture Size, (num) layers  1D=(32768), 2048 layers
  Maximum Layered 2D Texture Size, (num) layers  2D=(32768, 32768), 2048 layers
  Total amount of constant memory:               65536 bytes
  Total amount of shared memory per block:       49152 bytes
  Total shared memory per multiprocessor:        102400 bytes
  Total number of registers available per block: 65536
  Warp size:                                     32
  Maximum number of threads per multiprocessor:  1536
  Maximum number of threads per block:           1024
  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
  Max dimension size of a grid size    (x,y,z): (2147483647, 65535, 65535)
  Maximum memory pitch:                          2147483647 bytes
  Texture alignment:                             512 bytes
  Concurrent copy and kernel execution:          Yes with 1 copy engine(s)
  Run time limit on kernels:                     Yes
  Integrated GPU sharing Host Memory:            No
  Support host page-locked memory mapping:       Yes
  Alignment requirement for Surfaces:            Yes
  Device has ECC support:                        Disabled
  CUDA Device Driver Mode (TCC or WDDM):         WDDM (Windows Display Driver Model)
  Device supports Unified Addressing (UVA):      Yes
  Device supports Managed Memory:                Yes
  Device supports Compute Preemption:            Yes
  Supports Cooperative Kernel Launch:            Yes
  Supports MultiDevice Co-op Kernel Launch:      No
  Device PCI Domain ID / Bus ID / location ID:   0 / 1 / 0
  Compute Mode:
     < Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 12.4, CUDA Runtime Version = 12.4, NumDevs = 1
Result = PASS

By running the bandwidthTest program, you can ensure that the system and CUDA-capable device are able to communicate correctly.

bandwidthText.exe

The output shoud be here.

[CUDA Bandwidth Test] - Starting...
Running on...

 Device 0: NVIDIA GeForce RTX 3070 Ti Laptop GPU
 Quick Mode

 Host to Device Bandwidth, 1 Device(s)
 PINNED Memory Transfers
   Transfer Size (Bytes)        Bandwidth(GB/s)
   32000000                     11.3

 Device to Host Bandwidth, 1 Device(s)
 PINNED Memory Transfers
   Transfer Size (Bytes)        Bandwidth(GB/s)
   32000000                     13.9

 Device to Device Bandwidth, 1 Device(s)
 PINNED Memory Transfers
   Transfer Size (Bytes)        Bandwidth(GB/s)
   32000000                     361.1

Result = PASS

NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.

To see a graphical representation, you can run the particles sample program.

particles.exe

NVIDIA CUDA Particle Sample

The installed NVIDIA CUDA Toolkit provides the necessary libraries, compilers, and tools for developing and running CUDA-accelerated applications and machine learning models.

4. Installing NVIDIA cuDNN

cuDNN (CUDA Deep Neural Network Library) is a GPU-accelerated library specifically designed and colaborated with NVIDA CUDA Toolkit to accelerate deep neural network computations. By utilizing cuDNN, deep learning frameworks can leverage the parallel processing capabilities of NVIDIA GPUs, leading to significant speed improvements in training and inference of deep neural networks.

You can visit NVIDIA Developer website for cuDNN for the latest version. You will need to register or log in to your NVIDIA Developer account in order to access the cuDNN download files. If you don’t have an account, you can create one for free.

Once you are logged in, choose the appropriate cuDNN version based on the NVIDIA CUDA Toolkit version and operating system. There are two main installation options:

  1. Graphical installation (executable): the graphical installer bundles the available per-CUDA cuDNN verions in one package.
  2. Tarball installation (zip): per-CUDA cuDNN versions are provided as saperate tarballs (zip). These .zip archives do not replace the graphical installer and are not meant for general consumption, as they are not installers. These zip archives can be found at this.

Select one of two options for installing cuDNN. In this post, I will install cuDNN via the Tarball installation option.

Once download the zip archive, you unzip the cuDNN package.

NVIDIA cuDNN Unzip

Copy the following files from the unzipped package into the NVIDIA cuDNN directory created by yourself.

  • Copy bin\cudnn*.h to C:Program Files\NVIDIA\CUDNN\vx.y\bin.
  • Copy include\cudnn*.h to C:\Program Files\NVIDIA\CUDNN\vx.y\include.
  • Copy lib\x64\cudnn*.lib to C:\Program Files\NVIDIA\CUDNN\vx.y\lib.

You must replace x.y with your specific cuDNN version.

Set the environment variable to point to where cuDNN is located and add bin directory path to PATH environment variable.

NVIDIA cuDNN Path

For upgrading cuDNN, the remove the path to the directory containing cuDNN from PATH environment variable.

Verify cuDNN Installation

The cuDNN samples can be found here and download and extract the tar.xz archive.

NVIDIA cuDNN Samples

Since this is cross-platform LINUX samples, you need to install CMAKE and use it to compile the cuDNN samples.

Inside the cuda_sample_vx directory (x as the cuDNN version), make the comment to line 21 # add_subdirectory(mnistCUDNN).

Run the following command in the command prompt.

mkdir build && cd build
cmake -G "Visual Studio 17 2022" -D cuDNN_INCLUDE_DIR="C:\Program Files\NVIDIA\CUDNN\v9.0\include" -D cuDNN_LIBRARY_DIR="C:\Program Files\NVIDIA\CUDNN\v9.0\lib" ..
cmake --build . --config Release

By running the conv_sample program, you can ensure the cuDNN work in your system.

.\conv_sample\Release\conv_sample.exe

The result should be as following:

Executing: conv_sample.exe
Using format CUDNN_TENSOR_NCHW (for INT8x4 and INT8x32 tests use CUDNN_TENSOR_NCHW_VECT_C)
Testing single precision
====USER DIMENSIONS====
input dims are 1, 32, 4, 4
filter dims are 32, 32, 1, 1
output dims are 1, 32, 4, 4
====PADDING DIMENSIONS====
padded input dims are 1, 32, 4, 4
padded filter dims are 32, 32, 1, 1
padded output dims are 1, 32, 4, 4
Testing conv
^^^^ CUDA : elapsed = 0.0127218 sec,
Test PASSED
Testing half precision (math in single precision)
====USER DIMENSIONS====
input dims are 1, 32, 4, 4
filter dims are 32, 32, 1, 1
output dims are 1, 32, 4, 4
====PADDING DIMENSIONS====
padded input dims are 1, 32, 4, 4
padded filter dims are 32, 32, 1, 1
padded output dims are 1, 32, 4, 4
Testing conv
^^^^ CUDA : elapsed = 0.0029165 sec,
Test PASSED

For Visual Studio project, add cuDNN by following steps:

  • Right-click on the project name in Solution Explorer and choose **Properties.
  • Click VC++ Directories and append C:\Program Files\NVIDIA\CUDNN\v9.x\include to the Include Direcotries field.
  • Click Linker > General and append C:\Program Files\NVIDIA\CUDNN\v9.x\lib to the Additional Library Directories field.
  • Click Linker > Input and append cudnn.h to the Additional Dependencies field and click OK.

5. Installing NVIDIA TensorRT (Optional)

NVIDIA TensorRT is a C++ library that facilitates high-performance inference NVIDIA graphic processing units (GPUs). TensorRT takes a trained network, which consists of a network definition and a set of trained parameters, and produces a highly optimized runtime engine that performs inference for that network.

TensorRT provides APIs via C++ and Python that help to express deep learning model via the Network Definition API or load a pre-defined model via the ONNX parser that allow TensorRT to optimize and run them on the NVIDIA GPU.

TensorRT also include optional high speed mixed precision capabilities with difference NVIDIA architectures.

You can download the TensorRT at here. For Windows architecture, there is only zip archive installation.

Unzip the zip archive and copy files in lib, include direcotries to C:\Program Files\NVIDIA\TensorRT\v10.0 directory created by yourself. Then, you add lib directory path to PATH environment variable.

NVIDIA TensorRT Path

Verify TensorRT Installation

Inside the zip archive also include the sample programs. To verify the installation is working, you should open a Visual Studio file from one of the samples, such as sampleOnnxMNIST.In the project, ensure that following is presented in the Visual Studio Solution project properties:

  • Add C:\Program Files\NVIDIA\TensorRT\v10.0\lib to PATH and VC++ Directories > Executable Directories.
  • Add C:\Program Files\NVIDIA\TensorRT\v10.0\include to C/C++ > General > Additional Directories.
  • Add nvinfer.lib and .lib files that that the projects requires to Linker > Input > Additional Dependencies.

Compile the source code and run the example.

.\bin\sample_onnx_mnist.exe

The output should be as following:

&&&& RUNNING TensorRT.sample_onnx_mnist [TensorRT v100000] # .\sample_onnx_mnist.exe
[04/08/2024-18:12:10] [I] Building and running a GPU inference engine for Onnx MNIST
[04/08/2024-18:12:10] [I] [TRT] [MemUsageChange] Init CUDA: CPU +109, GPU +0, now: CPU 18227, GPU 1091 (MiB)
[04/08/2024-18:12:18] [I] [TRT] [MemUsageChange] Init builder kernel library: CPU +2597, GPU +310, now: CPU 21109, GPU 1401 (MiB)
[04/08/2024-18:12:18] [I] [TRT] ----------------------------------------------------------------
[04/08/2024-18:12:18] [I] [TRT] Input filename:   ../data/mnist/mnist.onnx
[04/08/2024-18:12:18] [I] [TRT] ONNX IR version:  0.0.3
[04/08/2024-18:12:18] [I] [TRT] Opset version:    8
[04/08/2024-18:12:18] [I] [TRT] Producer name:    CNTK
[04/08/2024-18:12:18] [I] [TRT] Producer version: 2.5.1
[04/08/2024-18:12:18] [I] [TRT] Domain:           ai.cntk
[04/08/2024-18:12:18] [I] [TRT] Model version:    1
[04/08/2024-18:12:18] [I] [TRT] Doc string:
[04/08/2024-18:12:18] [I] [TRT] ----------------------------------------------------------------
[04/08/2024-18:12:18] [I] [TRT] Local timing cache in use. Profiling results in this builder pass will not be stored.
[04/08/2024-18:12:20] [I] [TRT] Detected 1 inputs and 1 output network tensors.
[04/08/2024-18:12:20] [I] [TRT] Total Host Persistent Memory: 26400
[04/08/2024-18:12:20] [I] [TRT] Total Device Persistent Memory: 0
[04/08/2024-18:12:20] [I] [TRT] Total Scratch Memory: 0
[04/08/2024-18:12:20] [I] [TRT] [BlockAssignment] Started assigning block shifts. This will take 6 steps to complete.
[04/08/2024-18:12:20] [I] [TRT] [BlockAssignment] Algorithm ShiftNTopDown took 0.8572ms to assign 3 blocks to 6 nodes requiring 32256 bytes.
[04/08/2024-18:12:20] [I] [TRT] Total Activation Memory: 31744
[04/08/2024-18:12:20] [I] [TRT] Total Weights Memory: 26152
[04/08/2024-18:12:20] [I] [TRT] Engine generation completed in 1.68333 seconds.
[04/08/2024-18:12:20] [I] [TRT] [MemUsageStats] Peak memory usage of TRT CPU/GPU memory allocators: CPU 0 MiB, GPU 5 MiB
[04/08/2024-18:12:20] [I] [TRT] [MemUsageStats] Peak memory usage during Engine building and serialization: CPU: 3036 MiB
[04/08/2024-18:12:20] [I] [TRT] Loaded engine size: 0 MiB
[04/08/2024-18:12:21] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +0, now: CPU 0, GPU 0 (MiB)
[04/08/2024-18:12:21] [I] Input:
[04/08/2024-18:12:21] [I] @@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@=   ++++#++=*@@@@@
@@@@@@@@#.            *@@@@@
@@@@@@@@=             *@@@@@
@@@@@@@@.   .. ...****%@@@@@
@@@@@@@@: .%@@#@@@@@@@@@@@@@
@@@@@@@%  -@@@@@@@@@@@@@@@@@
@@@@@@@%  -@@*@@@*@@@@@@@@@@
@@@@@@@#  :#- ::. ::=@@@@@@@
@@@@@@@-             -@@@@@@
@@@@@@%.              *@@@@@
@@@@@@#     :==*+==   *@@@@@
@@@@@@%---%%@@@@@@@.  *@@@@@
@@@@@@@@@@@@@@@@@@@+  *@@@@@
@@@@@@@@@@@@@@@@@@@=  *@@@@@
@@@@@@@@@@@@@@@@@@*   *@@@@@
@@@@@%+%@@@@@@@@%.   .%@@@@@
@@@@@*  .******=    -@@@@@@@
@@@@@*             .#@@@@@@@
@@@@@*            =%@@@@@@@@
@@@@@@%#+++=     =@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@

[04/08/2024-18:12:21] [I] Output:
[04/08/2024-18:12:21] [I]  Prob 0  0.0000 Class 0:
[04/08/2024-18:12:21] [I]  Prob 1  0.0000 Class 1:
[04/08/2024-18:12:21] [I]  Prob 2  0.0000 Class 2:
[04/08/2024-18:12:21] [I]  Prob 3  0.0000 Class 3:
[04/08/2024-18:12:21] [I]  Prob 4  0.0000 Class 4:
[04/08/2024-18:12:21] [I]  Prob 5  1.0000 Class 5: **********
[04/08/2024-18:12:21] [I]  Prob 6  0.0000 Class 6:
[04/08/2024-18:12:21] [I]  Prob 7  0.0000 Class 7:
[04/08/2024-18:12:21] [I]  Prob 8  0.0000 Class 8:
[04/08/2024-18:12:21] [I]  Prob 9  0.0000 Class 9:
[04/08/2024-18:12:21] [I]
&&&& PASSED TensorRT.sample_onnx_mnist [TensorRT v100000] # .\sample_onnx_mnist.exe

If you are using TensorRT Python API, make sure CUDA-Python is installed in your system or your virtual environment.

mamba create -n tensorrt_env pip wheel
mamba activate tensorrt_env
  • Installing from PyPI, TensorRT runtime wheels
pip install cuda-python
pip install --pre --upgrade tensorrt
pip install --pre --upgrade tensorrt_lean tensorrt_dispatch

To verify the installation is working, use python code.

import tensorrt
print(tensorrt.__version__)
assert(tensorrt.Builder(tensorrt.Logger()))

import tensorrt_lean as trtl
print(trtl.__version__)
assert trtl.Runtime(trt.Logger())

import tensorrt_dispatch as trtd
print(trtd.__version__)
assert trtd.Runtime(trt.Logger())

The tensorrt/samples directory also provides the samples that utilize with the TensorRT, you can follow instructions from README.md to run samples.

Conclusion

By following these steps and installing the required software, you will have an CUDA-ready environment in Windows 11 system for further machine learnin/deep learning applications. This environment will provide the necessary tools and libraries for GPU-accelerated computing and Python package management.

Reference

]]> Setting Up a Personal Blog With Hugo and GitHub Pages http://visionbike.github.io/posts/setting-up-personal-blog-with-hugo-and-gh-pages/ Tue, 04 Jul 2023 01:53:57 +0800 http://visionbike.github.io/posts/setting-up-personal-blog-with-hugo-and-gh-pages/ Creating a personal blog with technical content is a excellent way to enhance the writting skill, keep memorial notes and share personal experience with others. Ideally, these goals need to be achieved when creating and mantaining a blog: Low-cost - Free or as close to free as posisble. Productive - Easy to write and maintain. Cloud Native - Utilizes public cloud services for hosting, allowing for infinite scaling. After researching, I found that using Markdown, Hugo and GitHub Pages is indeed a powerful combination for creating and maintaining a cost-effective, productive, and cloud-native blog. Creating a personal blog with technical content is a excellent way to enhance the writting skill, keep memorial notes and share personal experience with others. Ideally, these goals need to be achieved when creating and mantaining a blog:

  1. Low-cost - Free or as close to free as posisble.
  2. Productive - Easy to write and maintain.
  3. Cloud Native - Utilizes public cloud services for hosting, allowing for infinite scaling.

After researching, I found that using Markdown, Hugo and GitHub Pages is indeed a powerful combination for creating and maintaining a cost-effective, productive, and cloud-native blog.

  • Markdown is markup language that is extremely easy to read, write natively and can be converted into HTML.
  • Hugo is a static site generator written in the Go language that allows for content written in Markdown to be rendered into HTML webpages.
  • GitHub Pages is the GitHub service that hosts web contentstored in a GitHub repository.

In this post, I will show how to create an simple personal blog for FREE using above technologies. The blog was developed in Window Subsystem for Linux (WSL2).

Here’s an outline of the steps you can follow to create the personal blog using these technologies:

1. Setting up GitHub Account

If you don’t have one already, creating a GitHub account. GitHub Pages allows you to host your blog for free using a GitHub repository.

2. Installing Hugo

Before starting, make sure git is installed in the local machine.

sudo apt install -y git

Configure Git with your username and email address.

git config --global user.name "Your Name"

git config --global user.email "your.email@example.com"

To verify that Git has been installed successfully, you can check the version using git --version. This command will display the installed version of Git.

For Ubuntu user, you can install Hugo on your host by this command:

sudo apt install -y hugo

Run the folloing command for verification:

hugo version

The Hugo version should be shown if the installation is successfull.

hugo v0.92.2+extended linux/amd64 BuildDate=2023-01-31T11:11:57Z VendorInfo=ubuntu:0.92.2-1ubuntu0.1

3. Creating a new Hugo site

You can run hugo new site command to create a new Hugo site:

hugo new site &lt;USERNAME&gt;-Hugo -f yml

This command will set up the basic directory structure and configuration file in *.yml format for your blog.

What is `<USERNAME>`?

For convenient management and organization, you should name the your blog project as above format with <USERNAME> as the your GitHub’s username, i.e., Visionbike-Hugo. It’s helpful to keep track your project and ensuring clarity when managing multiple repositories.

The site will be associated with a GitHub repository where you can store the source code of your blog. Hence, you need to initialize git in the local project for further use.

cd &lt;USERNAME&gt;-Hugo
git init

You also need to create a new repository on GitHub storage for your blog’s source code.

Create source code repository

Creating Github Repository Without README File

By creating a repository without a README file, you can avoid accidental history conflicts when pushing your local project to a fresh repository. You can always add a README file later if needed.

Now, you link the local project to the GitHub repository by the git remote command:

git remote add origin https://github.com/&lt;USERNAME&gt;/&lt;USERNAME&gt;.github.io.git
git banrch -M master

By completing these steps, you have linked your local Hugo site to the GitHub repository. Now you can continue working on your site locally, commit any changes, and push them to the remote repository when ready.

4. Installing Hugo Theme

Installing a Hugo theme is a fantastic way to personalize your blog and enhance its visual appeal. You can access free Hugo themes via this website.

For my blog, I chose the PaperModX theme because of fonding its style and awesome features. I added its source code by the git submodule command.

git submodule add --depth 1 https://github.com/reorx/hugo-PaperModX themes/PaperModX

The command will add the PaperModX theme repository as a submodule in the themes/PaperModX directory of your Hugo site.

Updating submodules

If you have already added the submodule before, you can run the following command to reclone it.

git submodule update --init --recursive

For updating the theme, run this command.

git submodule update --remote --merge

5. Modify Hugo Configuration

Once you have added the theme, you can configure it in your Hugo site’s configuration file (config.yml). Refer to the theme’s documentation for specific instructions on customization and configuration options. You will most likely want to modify the following fields:

  • baseURL: This should be set into the URL GitHub Pages for hosting your blog. If the GitHub repository is named <USERNAME>.github.io, then the value of baseURL will be https://<USERNAME>.github.io/. If the GitHub repository has any other name, then the value will be https://<USERNAME>.github.io/<REPOSITORY_NAME>/. For instance, my GitHub username is visionbike, then:

    • If the GitHub repository is named visionbike.github.io, then the baseURL will be https://visionbike.github.io/.
    • If the GitHub repository is named visionbike-hugo, then the baseURL will be https://visionbike.github.io/visionbike-hugo/.

  • title: This will be the title of your blog site as it appears at the top of a visitor’s web browser when your site is open. It will also appear underneath your avatar, if one is present.

  • theme: The name of the theme Hugo should use to render your site. In my example, this will be set to PaperModX, since that is the name of the theme I am using.

Example contents of the config.yml file can be found below.

config.yml 
# base URL
baseURL: "http://visionbike.github.io/"

# site title
title: "Visionbike - Personal Blog of CV | DSP | ML notes"

# paginate
paginate: 5

# theme config
theme: "PaperModX"
themesdir: "themes"

# global config
enableInlineShortcodes: true
enableRobotsTXT: true
buildDrafts: false
buildFuture: false
buildExpired: false
enableEmoji: true

# css minify for speeding up site
minify:
  disableXML: true
  minifyOutput: true

# site param config
params:
  # environment
  env: "production"
  description: "Visionbike - Personal Blog of CV | DSP | ML notes"

  # color scheme: auto, dark, light
  defaultTheme: "dark"
  disableThemeToggle: true

  # header logo
  logo:
    text: "Visionbike"
    icon: "images/apple-touch-icon.png"
    iconHeight: 35
    iconWidth: 35

  # profile-mode
  profileMode:
    enabled: true
    title: "Phuc Thanh-Thien Nguyen"
    subtitle: "AI Researcher - Personal Blog of CV | DSP | ML notes"
    imageUrl: "images/avatar-real.png"
    imageTitle: "avatar-real"
    imageWidth: 180
    imageHeight: 180
    buttons:
      - name: Archives
        url: /archives/
      - name: Tags
        url: /tags/
  # home-info-mode
  homeInfoParams:
    title: "Phuc Thanh-Thien Nguyen"
    content: "AI Researcher - Personal Blog of CV | DSP | ML notes"

  # home social icons
  socialIcons:
    - name: github
      url: "https://github.com/visionbike/"
    - name: linkedIn
      url: "https://linkedin.com/in/nttphuc/"

  # post config
  author: "Visionbike"
  showCodeCopyButtons: true
  displayFullLangName: true
  showReadingTime: true
  showWordCount: true
  showPostNavLinks: true
  showBreadCrumbs: true
  enableImageZoom: true

  # assets images
  assets:
    # disableFingerprinting: true
    favicon: "images/favicon.ico"
    favicon16x16: "images/favicon-16x16.png"
    favicon32x32: "images/favicon-32x32.png"
    appleTouchIcon: "images/apple-touch-icon.png"
    safarPinnedTab: "images/safari-pinned-tab.svg"

  # search page
  fuseOpts:
    isCaseSensitive: false
    shouldSort: true
    location: 0
    distance: 1000
    threshold: 0.4
    minMatchCharLength: 0
    keys: ["title", "permalink", "summary", "content"]

# laguage config
languages:
  en:
    weight: 1
    # language code
    languageCode: "en"
    # determines default content language: "en", "zh-cn", "fr", "vi", ...
    defaultContentLanguage: "en"
    # menu language
    menu:
      main:
        - name: "Posts"
          weight: 1
          url: "/posts/"
        - name: "Tags"
          url: "/tags/"
          weight: 3
        - name: "Archive"
          url: "/archives/"
          weight: 4
        - name: "Publish"
          url: "/publish/"
          weight: 5
        - name: "About"
          weight: 10
          url: "/about/"
        - name: "Search"
          weight: 100
          url: "/search/"

# for search page
outputs:
  home:
    - HTML
    - RSS
    - JSON

# syntax highlight
pygmentsUseClasses: true
pygmentsCodeFences: true
markup:
  goldmark:
    renderer:
      unsafe: true
    highlight:
      lineNos: false
      codeFences: true
      noClasses: false

After modifying the configuration file accordingly, you can commit and push changes from your local repository to GitHub.

git add config.yml
git commit -m "modify configuration file"
git push -u origin master

6. Creating New Hugo Post

You are almost done finishing your personal blog!

To create the first post, you execute the hugo new command in the terminal.

hugo new posts/first-post/index.md

The command will create a new folder named first-post with new Markdown file index.md, inside the content/posts directory. Creating a new directory for each single post helps you manage your resource better when images, media sources can be store directly in this directory. The Markdown file will contain the template for your first blog post and you can start writing your content using Markdown syntax.

The contents of the first-post.md file will look like as:

first-post.md

---
title: "First Post"
date: 2023-07-04T01:53:57+08:00
draft: true
---

You need to add content to the Markdown file and update the metadata header. In the metadata header, you will find information such as the post’s title, publishing date and draft status. Change the value of the draft field from true to false to indicate that the post is ready to be published on your blog site. Your can also add other features supported by the installed theme for your post, i.e., comments, share buttons, navigation, etc.

The, you add the desired content to the body of the post. For instance, I added the line “This is my first post! Hello world!” at the bottom of the file. Feel free to customize the content to reflect your own thoughts and ideas in Markdown syntax.

After modifying the first blog post, you can use git commit to commit and push the changes from your local repository to GitHub.

git add content/posts/first-post.md
git commit -m "add the first post"
git push -u origin master

7. Testing the Hugo Configuration

Before hosting your blog to GitHub pages, make ensure Hugo can parse the configuration file and build our new blog post successfully.

In the local machine, you can run hugo server command to serve your site locally.

hugo server --disableFastRender

It will watch for any changes in your files and automatically rebuild your site whenever there are updates. Hugo will provide a local development server address, i.e., http://localhost:1313, where you can access your site locally.

To parse the configuration and build your site, you simply run hugo command.

hugo

If Hugo encounters any errors, they will be reported here. If the site is successfully built, then you will see output similar to the following.

Start building sites … 
hugo v0.92.2+extended linux/amd64 BuildDate=2023-01-31T11:11:57Z VendorInfo=ubuntu:0.92.2-1ubuntu0.1

                   | EN  
-------------------+-----
  Pages            | 13  
  Paginator pages  |  0  
  Non-page files   |  0  
  Static files     | 18  
  Processed images |  0  
  Aliases          |  0  
  Sitemaps         |  1  
  Cleaned          |  0  

Total in 38 ms

8. Setting Up GitHub Actions Workflow

Lastly, the GitHub Actions workflow has to be prepared for automatically building and deploying your blog to GitHub Pages. This workflow is defined by a YAML file in the .github/workflows directory structure at the root of the project.

First, create the workflows folder.

mkdir -p .github/workflows

The command will create the .github/workflows directory if it doesn’t already exist. The -p option ensures that the parent directories are created if needed.

Then, we create a new file within the created folder directory, named deploy_gh_pages.yaml with the following contents.

deploy_gh_pages.yaml

---
---
name: Deploy Hugo site via GitHub Pages

on:
  push:
    branches:
      - master # Set a branch to deploy
  pull_request:

jobs:
  deploy:
    runs-on: ubuntu-22.04
    permissions:
      contents: write
    concurrency:
      group: ${{ github.workflow }}-${{ github.ref }}
    steps:
      - uses: actions/checkout@v3
        with:
          submodules: true # Fetch Hugo themes (true OR recursive)
          fetch-depth: 0 # Fetch all history for .GitInfo and .Lastmod

      - name: Setup Hugo
        uses: peaceiris/actions-hugo@v2
        with:
          hugo-version: "latest"
          extended: true

      - name: Build
        run: hugo --minify

      - name: Deploy
        uses: peaceiris/actions-gh-pages@v3
        # If you're changing the branch from main,
        # also change the `master` in `refs/heads/master`
        # below accordingly.
        if: github.ref == 'refs/heads/master'
        with:
          github_token: ${{ secrets.GITHUB_TOKEN }}
          publish_dir: ./public

The YAML file sets up the deployment process using Hugo and GitHub Pages. The workflow is triggered on a push to the master branch, and it uses the specified actions to build and deploy your blog. You can find more in this.

Finally, we use git commands to commit and push the changes from your local repository to GitHub.

git add .github/workflows/deploy_gh_pages.yaml
git commit -m "Add GitHub Actions workflow"
git push -u origin master

9. Configuring GitHub Pages

The GitHub Pages will allow GitHub to build automatically and serve our website whenever changes are made to the underlying repository.

First, we create new branch named gh-page. This branch will be used by GitHub Pages to build and serve your website. You can create the branch using the branch creation feature in your GitHub repository.

Create gh-pages

Then, go to the Settings tab near the top of your repository.

Repository setting

In the left hand pane, locate and click on the Pages category.

GitHub pages setting

By default, GitHub Pages will be disabled for your repository. To enable it, we need to select a branch for GitHub Pages to build and serve our website from. Under the Source section in the middle pane, you will see a dropdown menu labeled None. Click on the dropdown menu and select the gh-pages branch. This tells GitHub Pages to build and serve your website from the gh-pages branch.

Select deployment branch

After selecting the deployment branch, you will see a notification indicating where your site will be published. It will provide you with a URL where your website can be accessed.

Published URL

Wait for a few minutes to allow GitHub Pages to build and deploy your website. When the deployment completes, you can click on the URL provided in the notification to view your website. It may take some time for the changes to propagate and for your website to become accessible.

Site demo

Conclusion

Congratulations on setting up your blog using Hugo, Markdown, and GitHub Pages! This free and accessible solution enables you to create and share your technical knowledge with a wide audience. Happy blogging!

Reference

]]>