Re: Dynamic mutidimensional arrays

[Michael Riedl]

Hallo,

if you carefully desing that (esacially the loop structure around with (i>k>j) you will not have too much cache misses.

This under the assumtion that A and B represent one block of memory each, not that e.g. A^[i] is pointing to somewere and e.g. A^[i+1] to somewhere totally different. Then surely any cache pre-fetch will fail. If A and B are exteemly large a blocking implementation would help out anyway, but that is off-topic here.

Michael

Am 05.04.23 um 11:00 schrieb Benjamin Kowarsch:

Hi

On Wed, 5 Apr 2023 at 17:46, Fischlin Andreas <andreas.fischlin@env.ethz.ch> wrote:

Have you actually looked at LgMatrices? There is no such run time penalty from calling a procedure such as value(M, i,j). You simply write e.g.

     val := A^[i]^[k] * B^[k]^[j];

But that will lead to unnecessary cache misses because of multiple levels of indirection.

If you have such an _expression_ deeply nested in a loop, these cache misses will add up and significantly impact performance.

It is better to allocate all metadata (such as length of a vector) and payload data together in a single memory block. This is called data locality.

regards

benjamin