Currently, the GPU implementation of matrix summaries in SCAMP generates patchy output when the pooling factor on the matrix is smaller than 1024x, i.e. if we have an input of length one million and an output matrix size of 4000 x 4000, we will get patchy output on the GPU because this is a pooling factor of ~250.
Currently, the GPU implementation of matrix summaries in SCAMP generates patchy output when the pooling factor on the matrix is smaller than 1024x, i.e. if we have an input of length one million and an output matrix size of 4000 x 4000, we will get patchy output on the GPU because this is a pooling factor of ~250.