1 --- !ditz.rubyforge.org,2008-03-06/issue
2 title: Parallelize alignment
3 desc: Alignment should be parallelized for OpenCL.
7 reporter: David Hilvert <dhilvert@auricle.dyndns.org>
10 creation_time: 2009-01-12 06:38:19.968512 Z
13 id: 091887e5b8deb698b1e390e6fdeaae0597334225
15 - - 2009-01-12 06:40:02.443275 Z
16 - David Hilvert <dhilvert@auricle.dyndns.org>
19 - - 2009-01-12 06:45:02.003891 Z
20 - David Hilvert <dhilvert@auricle.dyndns.org>
23 One reasonable approach to parallelization of alignment might be to perform
24 only measurement of error metrics on the compute device, but to perform many of
25 these measurements in parallel (with one measurement for each element having
26 remaining candidate transformations for measurement).
28 The host side would then perform the task of determining candidate
29 transformations for the next compute run and queueing these for execution.
30 - - 2009-01-12 07:32:23.000454 Z
31 - David Hilvert <dhilvert@auricle.dyndns.org>
34 As an alternative to processing a single match value (alignment error) at a
35 time on the compute device, a set of errors could be computed (e.g., from a
36 candidate set). Such a set could perhaps be processed for each work item,
37 where a work item would correspond to a particular point in the approximation
38 image space, or could occur across different work items, where each work item
39 would correspond to an (approximation point, candidate alignment) pair.