YaConv: Convolution with Low Cache Footprint

This article introduces YaConv , a new algorithm to compute convolution using GEMM microkernels from Basic Linear Algebra Subprograms library that is efficient for multiple CPU architectures. Previous approaches either create copy of each image element filter or reload these elements into cache call, leading redundant instances the in cache. Instead, loads once and maximizes reuse elements. The output computed by scattering results microkernel calls correct locations image. main advantage this algorithm—which leads better performance comparison existing im2col approach on several architectures—is more use memory hierarchy. experimental evaluation convolutional layers PyTorch, along with parameterized study, indicates an average 24% speedup over convolution. Increased comes as result 3× reduction L3 accesses 2× fewer branch instructions.