Well, I rechecked the code several times (and took a walk several times
too :))
Originaly, i was using vecLib's "dotpr" function as a reference point,
and had a bit-to-bit match with dotpr's results.
Now, as i de-serialized the Sum, i am getting different results
Here is a test output showing the mismatches then relative error is
larger then 1%
Result mismatch, orig -7.40401e-08 new -7.49715e-08 diff
-9.31323e-10 rel 0.0125786 VecLib -7.40401e-08
Result mismatch, orig 7.91624e-09 new 8.84756e-09 diff 9.31323e-10
rel 0.117647 VecLib 7.91624e-09
Result mismatch, orig 5.40167e-08 new 5.21541e-08 diff
-1.86265e-09 rel 0.0344828 VecLib 5.40167e-08
Result mismatch, orig 7.82311e-08 new 7.91624e-08 diff 9.31323e-10
rel 0.0119048 VecLib 7.82311e-08
Result mismatch, orig 3.44589e-08 new 3.35276e-08 diff
-9.31323e-10 rel 0.027027 VecLib 3.44589e-08
Result mismatch, orig 6.0536e-09 new 5.58794e-09 diff -4.65661e-10
rel 0.0769231 VecLib 6.0536e-09
Result mismatch, orig 4.88944e-08 new 4.98258e-08 diff 9.31323e-10
rel 0.0190476 VecLib 4.88944e-08
Result mismatch, orig -4.19095e-09 new -4.65661e-09 diff
-4.65661e-10 rel 0.111111 VecLib -4.19095e-09
Result mismatch, orig -1.90921e-08 new -1.81608e-08 diff
9.31323e-10 rel 0.0487805 VecLib -1.90921e-08
Result mismatch, orig 5.30854e-08 new 5.21541e-08 diff
-9.31323e-10 rel 0.0175439 VecLib 5.30854e-08
"orig" is my original altivec-accelerated function, "new" is the
function with deserialized Sum. "VecLib" is the result obtained from
dotpr.
diff is the difference between two results (original and new), and rel
is relative error.
The largest relative error i am getting is with the smallest values.
Absolute error is rather small, but because the numbers are small to,
relative error appear to be too large.
On Nov 7, 2005, at 7:41 PM, Holger Bettag wrote:
Sometimes the relative error is rather high, like 11% or so.
That does sound like way too much.
I think i did something wrong with shifting to partial sums, but i can
not guess what.
When that happens to me, it is usually a stupid mistake like
forgetting to
rename a variable in the tail loop, when its meaning has slightly
changed
in the main loop. Or something like that. Take a walk, and look at the
code again in an hour with a fresh mind. :-)
