Re: [wsjt-devel] wsprd_exp improvements

2015-07-03 Thread Steven Franke
Joe,
> On Jul 3, 2015, at 10:43 AM, Joe Taylor  wrote:
> 
> Hi Steve,
> 
>> Hmmm. That’s interesting and surprising. I tested it separately on 40m and 
>> 20m files and saw 35% and 17% reduction in execution time, respectively. So 
>> when I got 25% reduction for the large batch containing files from all 
>> bands, I was convinced that I had a robust result.
>> 
>> I wonder if this uncovers a difference between compiler optimizations on the 
>> Mac and Windows.
> 
> Not likely, I think.  Maybe it's a difference between our sets of 
> example files?  Or maybe I've made a mistake with my test.

I’m collecting a new batch of files. In the meantime, I wonder if you happened 
to save the wspr_timer files. It should be instructive to compare them. I have 
a hunch that the fact that I am running on a laptop with a solid-state disk 
drive can explain part of the difference. If you are running with a spinning 
disk, maybe the readwavfile times are vastly different and are swamping the 
other improvements?

I’ll include my times, below. Note that wsprd_exp spends 1/4 as much time in 
sync_and_demod(0)! I deliberately used some of this savings to increase the 
time spent in the decoder (by removing quickmode on the second pass). Note that 
these results are for 214 20m files, so the average time per file is almost 
exactly 1 second for wsprd_exp.

wsprd:

Code segmentSeconds   Frac
---
readwavfile  36.440.14
Coarse DT f0 f1  24.950.09
sync_and_demod(0)74.420.28
sync_and_demod(1)30.100.11
sync_and_demod(2) 9.390.04
Fano decoder 66.380.25
---
Total   267.171.00

wsprd_exp:

Code segmentSeconds   Frac
---
readwavfile  34.420.16
Coarse DT f0 f1  24.160.11
sync_and_demod(0)18.670.09
sync_and_demod(1)18.780.09
sync_and_demod(2) 8.980.04
Fano decoder 83.690.39
---
Total   212.351.00

Steve k9an

> 
>   -- Joe
> 
> --
> Don't Limit Your Business. Reach for the Cloud.
> GigeNET's Cloud Solutions provide you with the tools and support that
> you need to offload your IT needs and focus on growing your business.
> Configured For All Businesses. Start Your Cloud Today.
> https://www.gigenetcloud.com/
> ___
> wsjt-devel mailing list
> wsjt-devel@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/wsjt-devel


--
Don't Limit Your Business. Reach for the Cloud.
GigeNET's Cloud Solutions provide you with the tools and support that
you need to offload your IT needs and focus on growing your business.
Configured For All Businesses. Start Your Cloud Today.
https://www.gigenetcloud.com/
___
wsjt-devel mailing list
wsjt-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/wsjt-devel


Re: [wsjt-devel] wsprd_exp improvements

2015-07-03 Thread Joe Taylor
Hi Steve,

> Hmmm. That’s interesting and surprising. I tested it separately on 40m and 
> 20m files and saw 35% and 17% reduction in execution time, respectively. So 
> when I got 25% reduction for the large batch containing files from all bands, 
> I was convinced that I had a robust result.
>
> I wonder if this uncovers a difference between compiler optimizations on the 
> Mac and Windows.

Not likely, I think.  Maybe it's a difference between our sets of 
example files?  Or maybe I've made a mistake with my test.

-- Joe

--
Don't Limit Your Business. Reach for the Cloud.
GigeNET's Cloud Solutions provide you with the tools and support that
you need to offload your IT needs and focus on growing your business.
Configured For All Businesses. Start Your Cloud Today.
https://www.gigenetcloud.com/
___
wsjt-devel mailing list
wsjt-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/wsjt-devel


Re: [wsjt-devel] wsprd_exp improvements

2015-07-03 Thread Steven Franke
Hi Joe,
> On Jul 3, 2015, at 10:18 AM, Joe Taylor  wrote:
> 
> Hi Steve,
> 
> On 7/3/2015 9:29 AM, Steven Franke wrote:
>> Joe and wspr in wsjtx testers:
>> 
>> I’ve significantly improved the efficiency of the sync-search in the current 
>> version (r5668) of wsprd_exp.c. On a large (1953 files) batch of files 
>> representing all bands and the entire diurnal cycle the r5668 wsprd and 
>> wsprd_exp compare as follows:
>> 
>> wsprd:
>> 5083 decodes in 1294s (29% in sync0 and 17% in Fano)
>> 
>> wsprd_exp:
>> 5132 decodes in 970s (9% sync0 and 24% in Fano)
>> 
>> Approximately 25% decrease in average execution time with no decrease (maybe 
>> a slight increase) in performance.
>> 
>> If you get a chance and can confirm that you also see a significant time 
>> savings, then I propose that we make this version of wsprd_exp the new 
>> default wsprd. At some point we should be able to cash in some of the time 
>> saved for more decodes.
> 
> I ran the same test as previously, this time using wsprd_exe built from 
> your code in r5668.  Here's my table from before, with the new results 
> added as line 7:
> 
> Decodes Time1  AvgTimeImprovement   Decoder
>   (s)(s)Decodes  Speed
> 
> 1.  1451   2111 5.5 1.001.00   baseline
> 2.  1693   1599 4.1 1.171.32   wspr4
> 3.  2208335 0.9 1.526.30   WSJT-X v1.6.0 r5636
> 4.  2464413 1.1 1.705.11   partial coherence
> 5.  2567431 1.1 1.774.90   full coherence
> 6.  2839   2136 5.5 1.960.99   more candidates
> 7.  2570428 1.1 1.774.93   r5668
> 
> On this batch of files (all 20 meters, many decodable signals per file) 
> the results are nearly the same as those of my Run#5 ("subtraction with 
> full coherence and test for local maxima").  Slightly faster, slightly 
> more decodes.

Hmmm. That’s interesting and surprising. I tested it separately on 40m and 20m 
files and saw 35% and 17% reduction in execution time, respectively. So when I 
got 25% reduction for the large batch containing files from all bands, I was 
convinced that I had a robust result. 

I wonder if this uncovers a difference between compiler optimizations on the 
Mac and Windows.

Steve

> 
>   -- Joe
> 
> --
> Don't Limit Your Business. Reach for the Cloud.
> GigeNET's Cloud Solutions provide you with the tools and support that
> you need to offload your IT needs and focus on growing your business.
> Configured For All Businesses. Start Your Cloud Today.
> https://www.gigenetcloud.com/
> ___
> wsjt-devel mailing list
> wsjt-devel@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/wsjt-devel


--
Don't Limit Your Business. Reach for the Cloud.
GigeNET's Cloud Solutions provide you with the tools and support that
you need to offload your IT needs and focus on growing your business.
Configured For All Businesses. Start Your Cloud Today.
https://www.gigenetcloud.com/
___
wsjt-devel mailing list
wsjt-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/wsjt-devel


Re: [wsjt-devel] wsprd_exp improvements

2015-07-03 Thread Joe Taylor
Hi Steve,

On 7/3/2015 9:29 AM, Steven Franke wrote:
> Joe and wspr in wsjtx testers:
>
> I’ve significantly improved the efficiency of the sync-search in the current 
> version (r5668) of wsprd_exp.c. On a large (1953 files) batch of files 
> representing all bands and the entire diurnal cycle the r5668 wsprd and 
> wsprd_exp compare as follows:
>
> wsprd:
> 5083 decodes in 1294s (29% in sync0 and 17% in Fano)
>
> wsprd_exp:
> 5132 decodes in 970s (9% sync0 and 24% in Fano)
>
> Approximately 25% decrease in average execution time with no decrease (maybe 
> a slight increase) in performance.
>
> If you get a chance and can confirm that you also see a significant time 
> savings, then I propose that we make this version of wsprd_exp the new 
> default wsprd. At some point we should be able to cash in some of the time 
> saved for more decodes.

I ran the same test as previously, this time using wsprd_exe built from 
your code in r5668.  Here's my table from before, with the new results 
added as line 7:

 Decodes Time1  AvgTimeImprovement   Decoder
   (s)(s)Decodes  Speed

1.  1451   2111 5.5 1.001.00   baseline
2.  1693   1599 4.1 1.171.32   wspr4
3.  2208335 0.9 1.526.30   WSJT-X v1.6.0 r5636
4.  2464413 1.1 1.705.11   partial coherence
5.  2567431 1.1 1.774.90   full coherence
6.  2839   2136 5.5 1.960.99   more candidates
7.  2570428 1.1 1.774.93   r5668

On this batch of files (all 20 meters, many decodable signals per file) 
the results are nearly the same as those of my Run#5 ("subtraction with 
full coherence and test for local maxima").  Slightly faster, slightly 
more decodes.

-- Joe

--
Don't Limit Your Business. Reach for the Cloud.
GigeNET's Cloud Solutions provide you with the tools and support that
you need to offload your IT needs and focus on growing your business.
Configured For All Businesses. Start Your Cloud Today.
https://www.gigenetcloud.com/
___
wsjt-devel mailing list
wsjt-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/wsjt-devel