@techreport{oai:ipsj.ixsq.nii.ac.jp:00186051,
 author = {Peng, Chen and Mohamed, Wahib and Shinichiro, Takizawa and Satoshi, Matsuoka and Peng, Chen and Mohamed, Wahib and Shinichiro, Takizawa and Satoshi, Matsuoka},
 issue = {22},
 month = {Feb},
 note = {The 2D convolution operator is the computational bottleneck in a variety of image processing and machine learning applications. We propose an algorithm to compute convolution by employing register files to cache image data (known as register cache), rather than using the user-managed scratch-pad memory. We take advantage of CUDA's warp shuffle functions to accelerate the intra-warp communication of partial results. Unlike the GEMM-based, FFT-based or Winograd method, our algorithm executes the convolution computation without using any GPU memory as a workspace, and is general to all filter shapes. Our algorithm performs better than state-of-the-art 2D convolution implementations. Using a single TitanXp GPU, it is in average 4.7x faster than NPP (Nvidia Performance Primitives), and 1.8x faster than the highly-optimized ArrayFire library., The 2D convolution operator is the computational bottleneck in a variety of image processing and machine learning applications. We propose an algorithm to compute convolution by employing register files to cache image data (known as register cache), rather than using the user-managed scratch-pad memory. We take advantage of CUDA's warp shuffle functions to accelerate the intra-warp communication of partial results. Unlike the GEMM-based, FFT-based or Winograd method, our algorithm executes the convolution computation without using any GPU memory as a workspace, and is general to all filter shapes. Our algorithm performs better than state-of-the-art 2D convolution implementations. Using a single TitanXp GPU, it is in average 4.7x faster than NPP (Nvidia Performance Primitives), and 1.8x faster than the highly-optimized ArrayFire library.},
 title = {Pushing the Limits for 2D Convolution Computation On CUDA-enabled GPUs},
 year = {2018}
}