Aliya Clark
Reputation: 131
Reduction of data that resides in shared memory when size is not a power of 2
I am working on a CUDA project and I can't figure out how to reduce an array when its size isn't a power of 2.
There are many problems related to this on SO, however in my case the kernel has already been launched with a 2D block and 2D grid launch configuration and the array is in shared memory. I don't think padding is a option since the array has a size of 280 to 300 elements and I would have to pad up to 512 elements.
Is there any efficient algorithm for this?
Upvotes: 0
Views: 95
Answers (0)
Related Questions
- CUDA: Using grid-strided loop with reduction in shared memory
- What is the difference of dynamic shared memory as kernel attribute and kernel argument in CUDA
- Maximum threads per block vs shared memory size
- CUDA Reduction on Shared Memory with Multiple Arrays
- CUDA: Tiled matrix-matrix multiplication with shared memory and matrix size which is non-multiple of the block size
- What CUDA shared memory size means
- When is padding for shared memory really required?
- CUDA - Parallel Reduction over one axis
- Kernel lauch specifying the stream, but with default shared memory size