Indirect Deconvolution Algorithm

Neural network frameworks today commonly implement Deconvolution and closely related Convolution operator via a combination of GEMM (dense matrix-matrix multiplication) and a memory transformation. The recently proposed Indirect Convolution algorithm suggests a more efficient implementation of Convolution via the Indirect GEMM primitive - a modification of GEMM where pointers to rows are loaded from a buffer rather than being computed assuming constant stride. However, the algorithm is inefficient for Deconvolution with non-unit stride, which is typical in computer vision models. We describe a novel Indirect Deconvolution algorithm for efficient evaluation of the Deconvolution operator with nonunit stride by splitting Deconvolution with a large kernel into multiple subconvolutions with smaller, variable-size kernels, which can be efficiently implemented on top of the Indirect GEMM primitive.