NVidias Thrust library is used to interact with the GPU CLion integration works fine (code highlighting, code completion, debugging, ...)
The used system:
> nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.63.01 Driver Version: 470.63.01 CUDA Version: 11.4 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA GeForce ... Off | 00000000:0B:00.0 On | N/A |
| 0% 57C P0 39W / 151W | 694MiB / 8116MiB | 7% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
This is a GTX 1070 (Pascal)
- don't use -pedantic warnings! This creates a lot of warnings which are of no use in this case.
- at least on arch linux the Thrust and CUB cmake targets are not found. Set them manually via Thrust_DIR and CUB_DIR
- All Thurst code has to go into files with the .cu extension! If not the nvidia compiler will not be used.
CMAKE example file:
project(gpuapp LANGUAGES CXX CUDA)
set(SOURCES main.cu )
set( Thrust_DIR /opt/cuda/include/thrust/cmake)
set( CUB_DIR /opt/cuda/include/cub/cmake/)
find_package(Thrust REQUIRED CONFIG)
thrust_create_target(Thrust)
add_executable(${PROJECT_NAME} ${SOURCES})
target_link_libraries(${PROJECT_NAME} PUBLIC Thrust )
| directory | descritption |
|---|---|
| gpuapp | a minimal gpu example (with Thrust) |
| imageproc | a possible interop between OpenCV and Thrust (shared object on the GPU). This will only compile if you have a OpenCV version compiled with CUDA! |