llrCUDA

[Ralf Recker]

Originally Posted by Ken_g6

V0.16, about 3% faster than 0.15.

5*2^1282755+1 is prime! Time : 911.813 sec.
And about 0.7 ms/bit!

Changes:

- Everted if statement with carry clause.
- Merged cuda_normalize2_kernel into the two cuda_normalize1_kernels.
- Fixed flag error check - or, not and.
- Used flag error check only on the error checking cases. This is how Jean Penne did it.
- "flag" is now returned as a special err value, as it's an exceptional case. This also saves bandwidth.

ralf@quadriga ~/tmp/llrcuda.0.16 $ time ./llrCUDA -q5*2^1282755+1 -d
Starting Proth prime test of 5*2^1282755+1, FFTLEN = 131072 ; a = 3
5*2^1282755+1 is prime! Time : 848.820 sec.

real 14m8.854s
user 5m31.889s
sys 6m15.963s

---

Linux 64 bit / GTX 460 / CUDA SDK 3.2 / 260.19.29 drivers

[msft]

Originally Posted by Jean Penné

I hope it can resolve your overflow problem...
Jean

Thank you everything,
karatsquareg is slow with conversions.
I modify addsignal & giant_to_double.
$ time ./llrCUDA -d -q7*2^3015762+1
Starting Proth prime test of 7*2^3015762+1, FFTLEN = 262144 ; a = 3
7*2^3015762+1 is prime! Time : 3994.725 sec.

real 66m35.339s
user 30m51.700s
sys 27m54.610s

[Ralf Recker]

$ time ./llrCUDA -d -q7*2^3015762+1 vs. $ time ./llr -q7*2^3015762+1 -d

GPU: ca. 1.167 ms/bit - llrCUDA v0.17
CPU: ca. 2.321 ms/bit - llr 3.8.4

Result:


ralf@quadriga ~/tmp/llrcuda.0.17 $ ./llrCUDA -q7*2^3015762+1 -d
Starting Proth prime test of 7*2^3015762+1, FFTLEN = 262144 ; a = 3
7*2^3015762+1 is prime! Time : 3529.997 sec.


Waiting for the CPU task to finish (near 50%)...

[Ralf Recker]

Starting Proth prime test of 7*2^3015762+1
Using all-complex Pentium4 type-3 FFT length 192K, Pass1=256, Pass2=768, a = 3
7*2^3015762+1 is prime! Time : 6947.736 sec.

[msft]

$ time ./llrCUDA -d -q3*2^5082306+1
Starting Proth prime test of 3*2^5082306+1, FFTLEN = 524288 ; a = 5
3*2^5082306+1 is prime! Time : 13317.977 sec.

real 221m58.636s
user 122m8.680s
sys 115m11.880s

[msft]

$ time ./llrCUDA -d -q7*2^3015762+1
Starting Proth prime test of 7*2^3015762+1, FFTLEN = 262144 ; a = 3
7*2^3015762+1 is prime! Time : 3760.825 sec.

real 62m42.125s
user 20m43.400s
sys 24m50.010s

Tuning Finish.
Please report bug(or prime).

[msft]

Fix bug.
Thank you Ken_g6,

[jasonp]

If anyone is using the giants library, then I recommend changing the lower-level transforms to use the arbitrary-precision arithmetic from an older version of msieve. Start here, and use

common/ap.c
common/mp.c
common/fastmult.c
include/ap.h
include/fastmult.h
include/mp.h
include/util.h

This has a much cleaner implementation of the FFT arithmetic and uses 32-bit integers throughout.

msft, the code you were wondering about does the carry propagation after the FFT multiply completes.

[msft]

Hi, jasonp
Good information.
thank you,

[Jean Penné]

Hi,

I just released a new llrpi version that can test all the numbers that LLR 3.8.4 can test ; if you are interested, see :

http://www.mersenneforum.org/showthr...238#post250238

for details.

Regards,
Jean

[msft]

Hi,
rdft() is familiar to me.
Thank you,