Re: Motion P1788/M0009.01_ExactDotProduct
Arnold Neumaier wrote:
> The Zhu/Hayes paper quoted by Demmel shows that one can do the computations
> with a single pass through the data in an online fashion, with a storage
> that only in the worst case takes the length of the accumulator needed for
> the exact dot product.
In other words, it uses a software implementation of a complete format,
albeit one that is adaptive, and avoids the cost of a fixed allocation
for the worst case. (I'm in the middle of writing comments on Motion 9,
where I'm mentioning various possibilities for internal formats.)
(I'm guessing; I have not yet looked up the reference.)
Michel.
---Sent: 2009-10-30 20:21:41 UTC