The code is written for my rtx 4070s so there is no guarantee it will work on others gpus
but it probably will work on every rtx 4000, i will take a closer look at portability in the future
you can select your architecture passing the -DCMAKE_CUDA_ARCHITECTURES=XX flag where XX
is your cuda architecture.
| GPU Series | Compute Capability |
|---|---|
| RTX 40xx (Ada) | 89 |
| RTX 30xx (Ampere) | 86 |
| RTX 20xx (Turing) | 75 |
| GTX 16xx (Turing) | 75 |
| GTX 10xx (Pascal) | 61, 62 |
| GTX 900 (Maxwell) | 50, 52 |
otherwise you can use nvcc -ls-arch changed to nvcc --list-gpu-arch in newer cuda sdk release to get a detailed list.
mkdir -p build && cd build
cmake \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_CUDA_ARCHITECTURES=89 \
..
make -j`nproc --all`
./cuda
feh test.ppmnvidia has also recently dropped the support for various older gpu like gtx 950 etc.
If you type cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_ARCHITECTURES=52
You will get some random errors using latest versions of cuda sdk (es. on archlinux).
(these errors are similar to this and this)
And if you type something like this:
cmake .. -DCMAKE_BUILD_TYPE=Release \
-DCMAKE_CUDA_ARCHITECTURES=52 \
-DCMAKE_CUDA_COMPILER=/opt/cuda/bin/nvcc \
-DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++-15
to please the compiler you will finally get an unsupported architecture error.
Learning resources
- https://developer.nvidia.com/blog/cuda-pro-tip-write-flexible-kernels-grid-stride-loops/
- https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#thread-hierarchy
- https://codingbyexample.com/2018/12/17/cuda-threads-memory/
- https://www.youtube.com/playlist?list=PLC6u37oFvF40BAm7gwVP7uDdzmW83yHPe
- https://cuda-tutorial.readthedocs.io/en/latest/
- https://tschmidt23.github.io/cse599i/CSE%20599%20I%20Accelerated%20Computing%20-%20Programming%20GPUs%20Lecture%2017.pdf
profiling, nvprof is a sort of compatibility layer to use old nvprof syntax, but most of nvprof flags are simply ignored by nsys
gpu profiling
nsys --help profile
nsys --help nvprof
# es.
nsys profile -o report.qdrep ./cuda
nsys nvprof ./cuda
nsys profile --stats=true ./cuda
nsys nvprof --print-gpu-trace ./cuda
cpu profiling
perf stat -e task-clock,cycles,instructions,r1b1,r10e,stalled-cycles-frontend,stalled-cycles-backend,L1-dcache-load-misses,cache-misses ./cuda
perf stat -r 10 valgrind --tool=callgrind ./cuda
valgrind --tool=callgrind ./cuda
valgrind --tool=callgrind ./cuda | kcachegrinddynamic analysis
valgrind --undef-value-errors=no --tool=memcheck --leak-check=yes --show-reachable=yes --num-callers=20 --track-fds=yes ./cuda
asm
objdump -S -M intel cuda | gedit - &