That's a good idea to keep in mind. In hici() in particular we assume
the jumps will all predict correctly because we will be hashing elements
of the same length repeatedly.
Henry Rich
On 4/10/2022 4:42 AM, Elijah Stone wrote:
hici is a decent example. The branches at the end can be eliminated
by unconditionally hashing the last two words of the input (as by
v[k-2] and v[k-1], if v and k were never modified). In the event that
the number of words to be hashed is a multiple of 3 or one more than a
multiple of 3, some words will be hashed more than once, but this is
harmless. At a rough estimate, this costs 10 cycles for random
inputs. (For batch inputs, such as i. on skinny matrices, the branch
wins, but a transposed-ish traversal order will do better still, so it
doesn't matter much.)
On Sat, 9 Apr 2022, Henry Rich wrote:
If you show me an example I'd be obliged.
Henry Rich
On 4/9/2022 10:19 PM, Elijah Stone wrote:
Duff's device is a technique for loop unrolling which involves
jumping into the middle of the unrolled loop. Overlapping access is
an alternate technique also useful for loop unrolling (as well as
arranging for aligned access), which involves doing a small amount
of redundant work, but saves branches. Though it is slightly less
general, this tends to be more performant on modern hardware.
On Sat, 9 Apr 2022, Henry Rich wrote:
I don't understand the last sentence. 'duff'? 'overlapping access'?
Henry Rich
On 4/9/2022 9:37 PM, Elijah Stone wrote:
On Sat, 9 Apr 2022, Henry Rich wrote:
JE generally does few data-dependent branches, and I expect it
would not be a good idea to use two hyperthreads in one core; but
you'll have to make that decision.
Branch miss is tens of cycles; cache miss is hundreds.
Re branches: from what I've seen, there is too much duff and not
enough overlapping access. Though I guess those are
metadata-dependent branches, not data-dependent branches :)
----------------------------------------------------------------------
For information about J forums see
http://www.jsoftware.com/forums.htm
--
This email has been checked for viruses by AVG.
https://www.avg.com
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
