Rietveld Code Review Tool
Help | Bug tracker | Discussion group | Source code | Sign in
(5090)

Issue 1548: speexdsp: add initial check-asm for inner_product_*

Can't Edit
Can't Publish+Mail
Start Review
Created:
5 months, 3 weeks ago by tmatth
Modified:
5 months, 2 weeks ago
Reviewers:
jm
Visibility:
Public.

Patch Set 1 #

Unified diffs Side-by-side diffs Delta from patch set Stats Patch
M configure.ac View 1 chunk +5 lines, -0 lines 0 comments Download
M libspeexdsp/Makefile.am View 1 chunk +1 line, -1 line 0 comments Download
A libspeexdsp/resample_checkasm.h View 1 chunk +52 lines, -0 lines 0 comments Download
M libspeexdsp/resample_neon.h View 5 chunks +18 lines, -0 lines 0 comments Download
M libspeexdsp/resample_sse.h View 4 chunks +18 lines, -0 lines 0 comments Download

Messages

Total messages: 6
tmatth
5 months, 3 weeks ago #1
jm
Does this really work for the float/double version? Seems like you can't expect a perfect ...
5 months, 2 weeks ago #2
tmatth
On 2017/06/29 18:33:26, jm wrote: > Does this really work for the float/double version? Seems ...
5 months, 2 weeks ago #3
jm
On 2017/06/29 18:39:55, tmatth wrote: > I was wondering about that, it did in my ...
5 months, 2 weeks ago #4
tmatth
On 2017/06/29 19:02:26, jm wrote: > On 2017/06/29 18:39:55, tmatth wrote: > > I was ...
5 months, 2 weeks ago #5
tmatth
5 months, 2 weeks ago #6
On 2017/06/29 19:11:19, tmatth wrote:
> On 2017/06/29 19:02:26, jm wrote:
> > On 2017/06/29 18:39:55, tmatth wrote:
> > > I was wondering about that, it did in my limited testing, what would be a
> > > reasonable epsilon value?
> > 
> > That's the tough question. The only way I'm aware of to estimate that
involves
> > computing two more inner products :-)
> > 
> > That being, it's really suspicious that the test didn't trigger...
> 
> Ah needed to set_quality > 8 to test the double path, in which case:
> 
> 1500 0 1024 32 -> ASM mismatch: got:0.00014772263282588255,
> expected:0.00014772263282586864
> Fatal (internal) error in resample_sse.h, line 113: assertion failed: expected
> == ret

So as one example, testresample2 fails at quality 9 with epsilon == 1e-16 and
passes with 1e-15.
Sign in to reply to this message.

RSS Feeds Recent Issues | This issue
This is Rietveld