🚀 go-pugleaf

Just blueskying, but:

    The biggest advantage of vector instructions is that they convey
to the memory system the access pattern.  You have to add a lot of
logic to your memory buffers to detect patterns beyond the simple "the
next address is a stride of 64 past the previous" address
differencing.

    A secondary advantage of vector instructions is that they map to
multiple simple operations.

    Most vector implementations are pipelined, not element-by-element
parallel.  It isn't too hard to do superscalar/superpipelined
implementations of scalar instruction sets that obtain parallelism
comparable to most vector implementations. Multiple memory ports get
harder, but are doable. But the access pattern info isn't there.

    A disadvantage of architecting a vector instruction set is the
combinatoric explosion: first you provide element by element vector
ops, then vector reductions, then recurrences; then you have to decide
which vector ops are going to chain...

    So: why not combine vector memory access instructions that convey
access pattern, with scalar computational operations?



--

Andy Glew, glew@ichips.intel.com
Intel Corp., M/S JF1-19, 5200 NE Elam Young Parkway, 
Hillsboro, Oregon 97124-6497

This is a private posting; it does not indicate opinions or positions
of Intel Corp.