LLR Version 3.8.22 released

[mathwiz]

Originally Posted by Prime95

I suggest getting prime95 and running an AVX-512 torture test on the machine.

I also suggest building LLR with gwnum 29.7.

Is there a Changelog somewhere specifically for gwnum? For example, do 29.6 or 29.7 contain speedups for ECM and/or LLR?

[Prime95]

Originally Posted by mathwiz

Is there a Changelog somewhere specifically for gwnum? For example, do 29.6 or 29.7 contain speedups for ECM and/or LLR?

No speedups. There were some bug fixes -- mostly edge cases:

1) Fixed RNDVAL calculation error for zero-padded AVX-512 rational FFTs. Fixed error in calculating MULBYCONST values where const was negative. The HI values must be the same as positive consts because that value is used in avx512_prctab_index

2) Fixed case where gwsetup failed because it could not find a zero-padded or non-zero-padded implementation. Fixed by downgrading to FMA3 FFTs.

3) Fixed bug where gwtogiant wrote to 5th word in upper half of zero-padded FFT.
Test case was QA run of 10740401261265*2^25135-7

4) Fixed another bug where ZMM_RNDVAL was calculated using big_word rather than small_word. Test case was QA run of 117^3072-5.

5) Fixed bug where roundoff error exceeded 0.5. Zero-padded FFTs were using a different adjustment for b than the non-zero-padded code. Test case was QA run of 27904^53415-7.

6) Allow initialization of FMA3 FFTs above 600 million.

7) Changed FFT selection for rational FFTs. 1889^20480+1 was generating too large a roundoff error using a 10K AVX-512 FFT.

[pinhodecarlos]

Originally Posted by rebirther

I have recompiled 64bit only with the gwnum 29.7.


download

Thank you but looks like the download is blocked. Could you send a copy to Jean to be shared on his homepage. Much appreciated.

[rebirther]

Originally Posted by pinhodecarlos

Thank you but looks like the download is blocked. Could you send a copy to Jean to be shared on his homepage. Much appreciated.

I have sent the files but we are holding back this buggy version until it is fixed.

[pinhodecarlos]

Originally Posted by rebirther

I have sent the files but we are holding back this buggy version until it is fixed.

Ah, I see. Thank you.

[Prime95]

I've under the weather for 3 days and not investigated the problem.

Can someone with an AVX-512 Windows machine put this at the bottom of prime.txt:

[qa]
SPECIFIC_K=4344392810277
SPECIFIC_B=2
SPECIFIC_N=1290000
SPECIFIC_C=-1
CHECK_OFTEN=1

Then run prime95, Advanced/Time 9900. This will loop forever QAing the number. Stop the QA after at least one loop.

From the number of error messages you get on each test I should be able to narrow it down a bit.

Just post or email the results.txt output of just one of the QA loops.

[mackerel]

I didn't post earlier, but have replicated the varying residues on my 7800X/Win10 system. Starting it up now...

Edit:

Trying gwsetup on 4344392810277*2^1290000-1.
Starting irrational QA run on 4344392810277*2^1290000-1, kbits=40, cbits=5
Using zero-padded AVX-512 FFT length 128K, Pass1=128, Pass2=1K, clm=1, virtual_bits_per_word=19.794/20.258
Random seed is 1554412441
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Squares complete. MaxErr=0.1027945326
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test failed.
Test complete. MaxErr=0.1299960318

Is this the output required? The seed and MaxErr changes between each run.

Edit 2: forgot to say this is with 29.7b1

[Prime95]

Thanks, perfect.

[Prime95]

I've had little luck by just looking at the code. I know what gwnum routine is having trouble, but no luck in imagining the cause.

Grasping at straws, can a Windows user try this: https://www.dropbox.com/s/7btid8fwkz...ime95.exe?dl=0

I do not expect this to work. It's late. Tomorrow I'll try to whip up a debug prime95 version that can help narrow down the offending code.

[ATH]

Originally Posted by Prime95

I've had little luck by just looking at the code. I know what gwnum routine is having trouble, but no luck in imagining the cause.

Grasping at straws, can a Windows user try this: https://www.dropbox.com/s/7btid8fwkz...ime95.exe?dl=0

I do not expect this to work. It's late. Tomorrow I'll try to whip up a debug prime95 version that can help narrow down the offending code.

It still fails the Gerbicz error checks

[Prime95]

Next round of testing: https://www.dropbox.com/s/4l50ae4gee...ime95.exe?dl=0

As before, add this to prime.txt:

[qa]
SPECIFIC_K=4344392810277
SPECIFIC_B=2
SPECIFIC_N=1290000
SPECIFIC_C=-1
CHECK_OFTEN=1

Then run prime95, Advanced/Time 9900. This will loop forever QAing the number. Stop the QA after at least one loop.

Output will be a bit more than before, best to email it to me.

After that try it again with "SPECIFIC_SEED=1234" added to prime.txt.

Thanks!