Hi,
>clFFT problems?
Yes.
>What does OOURA code? Only checks rounding errors?
Translation from Complex FFT to HalfComplex FFT.

I have calculated my first LLR test on ATI gpu:

M( 64847711 )C, 0xffffffff80000000, n = 131072, clLucas v1.02

My card is Gigabyte R9 290.

Test was 5h 30m long.

[QUOTE=AK76;386081]M( 64847711 )C, [COLOR=Red]0xffffffff80000000[/COLOR], n = 131072, clLucas v1.02
[/QUOTE]
This is not a valid result, though. I may want to tune down OC on your card, check for dust etc etc. The usual things.

More likely, the FFT size is way too small. Let clLucas autoselect the FFT size, then use the next bigger power of two.

It is a fast card, but 5 hours can't be real (unfortunately).

The fft is too small. For an exponent of that size, you should be using n=4194304.

clLucas still only works with 2^k ffts, right?

[QUOTE=owftheevil;386087]The fft is too small. For an exponent of that size, you should be using n=4194304.

clLucas still only works with 2^k ffts, right?[/QUOTE]

Well, yes and no. Non 2^k FFT sizes work, but are slow.

EDIT: I'm surprised clLucas didn't stop right away...

I do not set -f parameter myself. clLucas set this number. I do the same test one more time. My ATI is not overclocked.

For me, it looks like this (HD7950 OC'd to the max (1100/1400) for this test):

[code]
C:\projects\clLucas>clLucas_x64.exe -aggressive 64847711
Platform :Advanced Micro Devices, Inc.
Device 0 : Tahiti

Build Options are : -D KHR_DP_EXTENSION

start M64847711 fft length = 3276800
err = 0.5, increasing n from 3276800

start M64847711 fft length = 3538944
Iteration 10000 0xf063d79e2bcf83d7, n = 3538944 err = 0.2813 (2:50 real, 16.9916 ms/iter, ETA 305:59:27)
^C caught. Writing checkpoint.

C:\projects\clLucas>rm c64847711

C:\projects\clLucas>clLucas_x64.exe -f 4194304 -aggressive 64847711
Platform :Advanced Micro Devices, Inc.
Device 0 : Tahiti

Build Options are : -D KHR_DP_EXTENSION

start M64847711 fft length = 4194304
Iteration 10000 0xf063d79e2bcf83d7, n = 4194304 err = 0.005371 (1:04 real, 6.3979 ms/iter, ETA 115:12:58)
Iteration 20000 0x929ed99e32ae2e48, n = 4194304 err = 0.005371 (1:04 real, 6.3386 ms/iter, ETA 114:07:48)
Iteration 30000 0x20c4df779543306b, n = 4194304 err = 0.005371 (1:03 real, 6.3383 ms/iter, ETA 114:06:26)
^C caught. Writing checkpoint.
[/code]Without "-aggressive" it is a bit slower, but the FFT selection is the same. AK76, did you maybe experiment with the -f parameter and this exponent at some time so that there were checkpoint files containing the bad FFT size?

I set FFT manually at 4194304 ande result is:


clLucas_x64 -f 4194304 64847711

Iteration 100000 0x8bc61577dc9b0813, n = 4194304 err = 0.004395 (1:06 real, 6.6892 ms/iter, ETA 120:17:37)

====================================

When run on automatic settings:


clLucas_x64 64847711

Iteration 20000 0x929ed99e32ae2e48, n = 3538944 err = 0.2344 (4:05 real, 24.5069 ms/iter, ETA 441:15:37)


The different is biiiig.

[QUOTE=AK76;386141]I do not set -f parameter myself. clLucas set this number. I do the same test one more time. My ATI is not overclocked.[/QUOTE]

Very strange... :loco:

[QUOTE=AK76;386147]I set FFT manually at 4194304 ande result is:


clLucas_x64 -f 4194304 64847711

Iteration 100000 0x8bc61577dc9b0813, n = 4194304 err = 0.004395 (1:06 real, 6.6892 ms/iter, ETA 120:17:37)

====================================

When run on automatic settings:


clLucas_x64 64847711

Iteration 20000 0x929ed99e32ae2e48, n = 3538944 err = 0.2344 (4:05 real, 24.5069 ms/iter, ETA 441:15:37)


The different is biiiig.[/QUOTE]

Yes, the difference is very big, but it compares well to my test results. You can try to set "-aggressive" to improve the timings, but apart from that, "-f 4194304" is the best thing to do. With this FFT size, you can even run tests with slightly bigger exponents as [URL="http://www.mersenneforum.org/showpost.php?p=366726&postcount=294"]LaurV explained[/URL] very nicely.