This project is a GPU-accelerated calculator designed for high-performance computations. It supports basic arithmetic operations, trigonometric functions, and (eventually) statistical calculations such as mean, median, and standard deviation. The calculator is built using CUDA and is optimized for NVIDIA GPUs.
Perfect if you need to add two absurdly large columns of numbers from the cli.
Definitely work-in-progress. I eventually want to do more complex calculations you might find in AI
- NVIDIA GPU with CUDA support
- CUDA toolkit
- Ensure you have a CUDA-compatible GPU and the CUDA toolkit installed.
- Clone the repository:
git clone https://github.com/maya-undefined/gpu-desktop-calculator - Navigate to the project directory:
cd gpu-desktop-calculator - Compile the project:
nvcc ./gdc.cu kernels.cu FH.cpp verb.cu -o gdc
sudo apt install nvidia-cuda-toolkit nvidia-container-toolkit docker.io
nvcc FH.cpp gdc.cu kernels.cu verb.cu -o gdc
Generally we use gdc <verb> file1 file2 out.file.
./gdc add f1.dat f2.dat out.dat
./gdc exp f1.dat out.dat
./gdc div f1.dat f2.dat out.dat
Files should be organized in columns of numbers. float is supported for now. The code will typically operate across (axis=0)
1 4.0
2.0 5.0
3.0 6.0
Files can be single or multiple column where the operation makes sense
A hypothetical usage to calculate a logistic regresion given a pre-trained w.dat might be
./gpu mul w.dat x.dat - | ./gpu add bias.dat - | ./gpu mul -1 - | ./gpu exp - | ./gpu add 1 - | ./gpu div 1 yhat.dat
y^=1+e−(w⋅x+b)1
Test data can be generated by datagen.sh. These will create 1.6 GB of test data.
complexPhysicsCalculation addArrays
cpu 75.9 67.3
gpu 61.7 61.8
implementation vs average time in seconds for generated data
Would be nice.
The File Handler (FH) class has some ways to io data rapidly. It needs to be faster. For now we mostly use parsing by lines and avoiding buffer flushes, which cause high IOWait.
Use a mothership process (MP) to handle the scheduling and data input of data.
It'd work something like this
./gdc mul w.dat x.dat | ./gpu add bias.dat | ./gpu mul -1 | ./gpu exp | \
./gpu add 1 | ./gpu end
This will tell the MP to allocate memory for data and keep it there, on the gpu to prevent constant data IO.
The ./gpu end command will tell the MP to start work. If desired, the
resulting data can be output to stdout or a file.
This way we can run the entire pipeline at once, or in stages at our leisure.
Be nice if this could store things on Google Drive, or elsewhere.
We could call a webhook every n-th calculations for observability, loss monitoring, etc.