FAQ Database Discussion Community

purposely causing bank conflicts for shared memory on CUDA device

It is a mystery for me how shared memory on cuda devices work. I was curious to count threads having access to the same shared memory. For this I wrote a simple program #include <cuda_runtime.h> #include <stdio.h> #define nblc 13 #define nthr 1024 //[email protected] __device__ int inwarpD[nblc]; __global__ void kernel(){...

CUDA shared memory bank conflicts report higher

I've been working on optimizing some code and ran into an issue with the shared memory bank conflict report with the CUDA Nsight performance analysis. I was able to reduce it to a very simple piece of code that Nsight reports as having a bank conflict, when it doesn't seem...

mobile OpenCL local memory bank conflict. Why using local memory is slower than does global memory in kernel?

I'm developing face detection app in android platform using OpenCL. Face detection algorithm is based on Viola Jones algorithm. I tried to make Cascade classification step kernel code. and I set classifier data of cascade stage 1 among cascade stages to local memory(__local) because classifier data are used for all...