When he sat in his living room, meticulously organizing each kernel so as not to draw the same one twice or drop one on the floor, he was, consciously or not, conducting some strange artsy experiment by doing what any scientist worth his salt does: Examine the small to understand the big.
The third article, "Self-organizing kernel adaptive filtering," by S. Zhao, B. Chen, Z. Cao, P. Zhu, and J. C. Principe, deals with kernel adaptive filters and presents a model-selection strategy based on minimum description length that keeps the kernel least-mean-square model tuned to the complexity of the input data.
Table 1 presents a list of kernels organized according to this theoretical framework.
BLAS consists of a relatively small set of highly optimized kernels, organized in three levels (1, 2, and 3), corresponding to vector, matrix-vector, and matrix-matrix operations, respectively.
In the FreeBSD OS, the sysctl tool allows to query more than 500 kernel variables organized hierarchically.
In CUDA, the thread hierarchy within a kernel is organized by the thread, the block of threads and the grid of blocks as shown in Figure 1, where different thread hierarchies have different memory scopes.
The parallel imple- mentations utilize OpenMP and CUDA kernels, and are organized according to thread scope.
The kernel heteroassociative memories were organized into the modular network.
Using CUDA, a kernel is executed on threads organized in blocks; each thread is responsible for a portion of data, and each block of threads shares a local memory (called shared).
This is obtained by modification of the kernels that have to be organized in such a way to enable (1) the overlap of the CUDA computation with the GPU data transfer/MPI communication and (2) the asynchronous execution of different kernels.
PNN is an implementation of a statistical algorithm called Kernel discriminant analysis in which the operations are organized into a multi-layered feedforward network with four layers: input, pattern, summation, and output layers.
