Inefficient memory access patterns in CUDA kernels #12
Labels
enhancement
New feature or request
Comments
dimitrivlachos
added a commit
that referenced
this issue
Nov 5, 2024
Implement extended dispersion spotfinding Implement a GPU-based version of the extended dispersion spotfinding algorithm. This builds on regular dispersion by making two passes. This allows for the detection of fainter spots by using the first pass to detect candidate spots and exclude them from the background calculation in the second pass. Extended dispersion spotfinding is unavoidably slower than regular dispersion by the fact that it requires two passes. However, the performance gained through massively parallel processing on the GPU should make this a viable option, when needed, even for fast feedback. Create several CUDA kernels to perform the extended dispersion spotfinding algorithm (`threshold.cu`, `erosion.cu`). Refactor the dispersion kernel to share code with extended dispersion. Move common code to `cuda_common.hpp`. Create basic test script for extended dispersion spotfinding. Add an `--algorithm` argument to `spotfinder.cc` along with the necessary code to parse it, allowing for algorithm selection at runtime. Add new files to the CMakeLists.txt file to include them in the build. See also: #12, #13, #14
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Currently, all CUDA kernels (including new kernels from #1) use global memory accesses when accessing neighbouring pixels in their convolution operations. This approach is inefficient and slows down performance.
We should modify these kernels to use more spatially optimised memory for accessing neighbouring pixels, thereby reducing global memory latency and enhancing overall performance.
This could include the use of local memory, shared memory, texture memory or surface memory, to name a few.
The text was updated successfully, but these errors were encountered: