1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
|
New signature #1:
SpImpl:SpMXSV(...., vector< pair<IT,IT> > vec_boundaries) // default is vec_boundaries = empty vector
- This hints the parallel logic down to the sequential operation
- Advantages:
a) We no longer need to repartition in the parallel function
b) Potentially smaller (but multiple) working sets. Any help?
The original prototype was:
void SpImpl::SpMXSpV(const Dcsc<IT,NT1> & Adcsc,
IT mA,
IT nA,
const IT * indx,
const NT2 * numx,
IT veclen,
vector<IT> & indy, // becomes a vector of vectors, size = boundaries.size() = p_c
vector< typename promote_trait<NT1,NT2>::T_promote > & numy) // same change here
New signature #2:
SpImpl:SpMXSV(...., vector< pair<IT,IT> > mat_boundaries) // default is mat_boundaries = empty vector
mat_boundaries.size() is ideally equal to the number of threads available.
If not, each thread handles multiple mat_boundaries.
How we do combine signatures #1 and #2?
|
