A CUDA tutorial to make people learn CUDA program from 0
Turing T4 GPU
- compile by hand
nvcc xxx.cu -o xxx
if that does not work, pls try:
nvcc xxx.cu --gpu-architecture=compute_yy -o xxx
xxx is file name, yy is GPU compute capability, ep.A100's compute capability is 86.
- one-click compile and run
please ensure:
1.cmake version >= 3.8
2.you have CUDA TOOLKIT installed in system root directory, downloaded link is https://developer.nvidia.com/cuda-downloads.
mkdir build
cd build
cmake .. && make -j8
cd bin
./xxx
- related performance data is attached at the top of code file.
- the performance data is diverse and diverse on different GPU platforms and NVCC compiler, so some counter-intuitive result is normal, we should only explore and debug the result.
- welcome all comments and pull requests.
- add cuda stream
- add quantize
- add fp32/fp16 gemv(vec * mat,mat is col major)
- add fp32/fp16 gemv(vec * mat,mat is row major)
- add some code explaination(WIP)
- add fp32 dropout