Re:Fortran is still a thing
by Const-me on Tuesday November 27, 2018 @09:46PM
(#18548875)
> A good tensor implementation accounts for strides that are SIMD compatible (eg; each dimension is a multiple of the SIMD register width).
Never implemented any tensors, but in my experience, sometimes you better do what GPUs do with some texture formats: switch from linear layout into dense blocks layout. E.g. for dense float32 matrix and SSE code, a good choice is a 2D array of 4x4 blocks: exactly 1 cache line, and only consumes 4 out of 16 registers while being processed i.e. you can do a lot within a block while not hitting any RAM latency.
Never implemented any tensors, but in my experience, sometimes you better do what GPUs do with some texture formats: switch from linear layout into dense blocks layout. E.g. for dense float32 matrix and SSE code, a good choice is a 2D array of 4x4 blocks: exactly 1 cache line, and only consumes 4 out of 16 registers while being processed i.e. you can do a lot within a block while not hitting any RAM latency.