Any ideas on v26 FFT selection algorithm? - Page 2

Rhyled · 2010-06-23, 15:14

Running the 64-bit benchmark on ver 26 only took a couple of minutes (not hours). The numbers look very similar to what I've seen from version 25. There are lots more "helper thread" messages in the new version:
x64 results.txt

Now, returning to my regular prime search

Rhyled · 2010-06-23, 15:23

Quote:

Originally Posted by Prime95

I'm working on the code that selects which FFT implementation to run for a given FFT size. Kinda stuck.

V25 uses a hardwired selection process based on the L2+L3 cache size.

Preliminary benchmarks on my Core 2 and Core i7 boxes indicate that this isn't optimal.

Does looking at L1/L2 cache only give you a better estimate? The problem with L3 cache is that it's shared between all cores, so you can't really predict how much of it will be useful/available for a single core.

axn · 2010-06-23, 15:59

Quote:

Originally Posted by Rhyled

Running the 64-bit benchmark on ver 26 only took a couple of minutes (not hours). The numbers look very similar to what I've seen from version 25. There are lots more "helper thread" messages in the new version:
Attachment 5381

Now, returning to my regular prime search

Quote:

Originally Posted by Prime95

Add this to prime.txt:

StressTester=1
MinBenchFFT=4
MaxBenchFFT=32768
OnlyBench5678=0
BenchAllComplex=1
AllBench=1

Then run Options/Benchmark. Post your results.txt file in this thread.

Ahem.

NBtarheel_33 · 2010-06-23, 16:24

Quote:

Originally Posted by Prime95

P.S. I really wouldn't recommend using this version for production work, though it will probably work.

I couldn't help it. I'm trying about 10 TF-LMHs in the 151M range. Would it be worth double-checking the results with v25?

Trying an LL with v26 is probably a no-no, right? Would a doublecheck be worthwhile, to make sure everything is running OK?

Prime95 · 2010-06-23, 16:32

Quote:

Originally Posted by NBtarheel_33

I couldn't help it. I'm trying about 10 TF-LMHs in the 151M range. Would it be worth double-checking the results with v25?

Trying an LL with v26 is probably a no-no, right? Would a doublecheck be worthwhile, to make sure everything is running OK?

The TF code is unchanged. Feel free to run a double-check if you like. The code is not optimized for a P4. It is optimized for Core 2.

NBtarheel_33 · 2010-06-23, 18:17

Quote:

Originally Posted by Prime95

The TF code is unchanged. Feel free to run a double-check if you like. The code is not optimized for a P4. It is optimized for Core 2.

OK, I have 39 cores running, including three flavors of Core 2 - 1.86 GHz, 2 GHz, and 2.66 GHz. Also have a Q6600 @ 2.4 GHz. And of course, a whole bunch of P4s and Pentium Ds, some of which are dual core. So I should be able to get you some useful data.

starrynte · 2010-06-23, 20:22

I still went ahead and benchmarked my i5, is attached

Rhyled · 2010-06-23, 23:27

Quote:

Originally Posted by axn

Ahem.

Ah. I see the problem. I did indeed add the specified lines, AFTER prime95 created a default .txt file with contradictory benchmark options. Just drop the correct version of prime.txt (also with qa) into that zipped file and you'll make it easier on the testers.

Here's the correct 64-bit benchmark results - took about 6 hours to run on my Core i7. Note that the last 3 cores on my i7 only averaged about 20% loading each while in use by prime95. For the last series, with all 4 cores, my total cpu load was around 50%, not the 100% I'm used to seeing with prime95.

x64 results.zip

ATH · 2010-06-25, 13:32

Core2Quad (Yorksfield) Q9450 2.66Ghz: Benchmark QA (no mismatch, MaxErr=0.3125 (only 1 above 0.3))

Core2Duo (Conroe) E6750 2.66Ghz: Benchmark QA (no mismatch, MaxErr=0.29296875)

Laptop: Mobile Core2Duo T7300 (Merom) 2.00Ghz: Benchmark QA (no mismatch, MaxErr=0.34375 (10 above 0.3))

Prime95 · 2010-06-25, 23:35

Thanks for all the benchmarks. No more please. Y'all have uncovered an optimization problem with the length 2304 pass 2. Once I get it fixed, we'll need to rerun the benchmarks. I've started working on some Pentium-4 optimizations and one of them might apply to Core 2 and Core i5/7/9 as well.

2010-06-23, 15:14	#12
Rhyled May 2010 111111₂ Posts	Benchmark on i7 looks very similar to ver 25 Running the 64-bit benchmark on ver 26 only took a couple of minutes (not hours). The numbers look very similar to what I've seen from version 25. There are lots more "helper thread" messages in the new version: x64 results.txt Now, returning to my regular prime search

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Some ideas regarding NFS...	paul0	Factoring	3	2015-03-14 19:55
two ideas for NPLB	TimSorbet	No Prime Left Behind	16	2008-03-01 23:32
GROUP IDEAS	TTn	15k Search	15	2003-09-23 16:28
Domain name ideas...	Xyzzy	Lounge	17	2003-03-24 16:20
Couple of ideas/things to do	Stormblade	Lounge	12	2002-08-20 02:21

2010-06-25, 13:32	#20
ATH Einyen Dec 2003 Denmark 2·3²·191 Posts	Core2Quad (Yorksfield) Q9450 2.66Ghz: Benchmark QA (no mismatch, MaxErr=0.3125 (only 1 above 0.3)) Core2Duo (Conroe) E6750 2.66Ghz: Benchmark QA (no mismatch, MaxErr=0.29296875) Laptop: Mobile Core2Duo T7300 (Merom) 2.00Ghz: Benchmark QA (no mismatch, MaxErr=0.34375 (10 above 0.3))

2010-06-25, 23:35	#21
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 17743₈ Posts	Thanks for all the benchmarks. No more please. Y'all have uncovered an optimization problem with the length 2304 pass 2. Once I get it fixed, we'll need to rerun the benchmarks. I've started working on some Pentium-4 optimizations and one of them might apply to Core 2 and Core i5/7/9 as well.