Measure the performance of computing the lane id #2
Description
Currently the lane id is read (here) by accessing the %laneid register. An alternative is to compute it by i % WARPSIZE
According to this and this, reading from the %laneid register is more costly than i % WARPSIZE. It would be good to have some benchmark results that show the difference between the two versions.
First create a benchamarks/laneid/ directory where you will be placing all your files. You can start with some simple kernels, for instance, each thread i reading its lane id and writing it to the i-th index of an array. You may subsequently move to more involved kernels. For each benchmark include appropriate plot(s) and information about your GPU device and CUDA version.