here is my feeble attempt at learning OpenCL, please don't make fun of me too much 🍔
This code uses OpenCL 1.1 on a NVIDIA GPU.
(Only tested on Ubuntu). For NVIDIA GPUs, I've installed the following packages: nvidia-346 nvidia-346-dev nvidia-346-uvm nvidia-libopencl1-346 nvidia-modprobe nvidia-opencl-icd-346 nvidia-settings. Since the opencl-headers package in the main repository is for OpenCL 1.2, you can get the OpenCL 1.1 header files from here.
Then to compile:
g++ -std=c++0x main.cpp -o main.out -lOpenCL
OpenCL is installed on OS X by default, but since this code uses the C++ bindings, you'll need to get that too. Get the official C++ bindings from the OpenCL registr and copy it to the OpenCL framework directory, or do the following:
wget https://www.khronos.org/registry/cl/api/1.1/cl.hpp
sudo cp cl.hpp /System/Library/Frameworks/OpenCL.framework/Headers/
To compile:
clang++ -std=c++0x -framework OpenCL main.cpp -o main.out
this example is based off of this example (example-ception), but it goes a bit further. In the blogspot example, two 10-element vectors are created and a thread is used for each pair of elements. In this example, 10 threads are spawned but two 100-element vectors are used, and it is shown how to split up a specific number of elements per thread.
Measures the duration of adding two vectors. See the README in the folder for more details.
Demonstrates that one array can be modified several times without having to re-read and re-write data to and from the GPU.
- figure out how OpenCL manages memory (when are buffers cleared on the GPU?)
- figure out how to view the OpenCL assembly code if possible (is warp divergence happening?)
From the guide on programming OpenCL for NVIDIA:
- CUDA streaming multiprocessor corresponds to an OpenCL compute unit
- CUDA thread corresponds to an OpenCL work-item
- CUDA thread block corresponds to an OpenCL work-group