Forcing a Tensor to shared memory in a SM

[vishpat]

I have implemented an image processing program using Candle that converts a colored PNG image into a greyscale image. This is done using a simple matrix multiplication operation. This program takes about 13 seconds to run on CPU vs 5 seconds when it runs on GPU for a 7345 x 2832 image. I wondered if it is possible to optimize the program further by replicating the weight tensor into the shared memory of all of the SMs on the GPU.