[QUOTE=msft;283156]Hi ,
Ver 1.4
Support non power of two fft length.:lol:
n = 786432 5.4 msec/Iter,n = 1310720 8.9 msec/Iter,n = 1835008 12.5 msec/Iter,n = 2359296 16 msec/Iter,n = 3670016 25.7 msec/Iter,n = 5242880 37.7 msec/Iter with GTX-550Ti
Thank you,[/QUOTE]

Is it safe to exchange versions (1.3 -> 1.4) during an LL test, or is better to finish the test and then exchange versions?

Luigi

Hi ,ET_
[QUOTE=ET_;283160]Is it safe to exchange versions (1.3 -> 1.4) during an LL test, or is better to finish the test and then exchange versions?

Luigi[/QUOTE]
Anyway cannot exchange fft length.

[QUOTE=msft;283165]Hi ,ET_

Anyway cannot exchange fft length.[/QUOTE]

Yep :redface: Thank you.

Luigi

2 wishes
 

Firstly, could anybody post pre-compiled 1.4 binaries for both linux and win(64)? This would help a lot and I could integrate the urls into the GPU guide.

Secondly, could any forum admin rename or add "CUDALucas" into the thread's title. This would help a lot, too...

Thanks!

Hi ,Brain
[QUOTE=Brain;283198]Firstly, could anybody post pre-compiled 1.4 binaries for both linux and win(64)?[/QUOTE]
1.4 binarie file for Linux64.

[QUOTE=msft;283250]Hi ,Brain

1.4 binarie file for Linux64.[/QUOTE]

Is it runable on Cuda 4.0?

Hi ,moebius
[QUOTE=moebius;283275]Is it runable on Cuda 4.0?[/QUOTE]
I have no information.
Do you have any comments?

[QUOTE=moebius;283275]Is it runable on Cuda 4.0?[/QUOTE]

According to CUDA_C_Programming_Guide.pdf the CUDA runtime version must be the same version used for compiling. The CUDA driver must support the same or a newer version of CUDA.

Oliver

[QUOTE=msft;283250]Hi ,Brain

1.4 binarie file for Linux64.[/QUOTE]
Thanks a lot. I thought of self-compiling the win version but I haven't MS Visual Studio installed. Last time I did, I almost smashed my win installation. So I'm still looking for a volunteer (kjaget, apsen, ethan_eo?)... Somewhen, I will set up a vm for MS VS and CUDALucas compilation.

Makefile.win says 9.0 is required. Is this correct? If anybody could post a short how-to-compile for Windows this would be nice. nvcc.exe <compile me>

Makefile.win:
[CODE]NVIDIA_SDK = "C:/ProgramData/NVIDIA Corporation/NVIDIA GPU Computing SDK"
LIBS = c:\\cuda\\lib64\\cudart.lib c:\\cuda\\lib64\\cufft.lib

CUFLAGS = -m64 --ptxas-options=-v -ccbin="C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\bin" -DWIN64 -Xcompiler /EHsc,/W3,/nologo,/Ox,/Oy,/GL -arch=sm_13 -DMERS_PACKAGE -DBIT_SIEVE -DTESTING_SMALL_EXPONENTS -DSIEVE_SIZE_IN_BYTES=32 -DNUM_SMALL_PRIMES=32768 -DDO_NOT_USE_LONG_DOUBLE -IC:/cuda/include -Ic:/cuda/include/cudart -I$(NVIDIA_SDK)/C/common/inc -D__x86_64__ -O3

LINK = link
LFLAGS = /nologo /LTCG #/ltcg:pgo

CUSRC = CUDALucas.cu setup.cu rw.cu balance.cu zero.cu

CUOBJS = $(CUSRC:.cu=.obj)

CUDALucas.exe: $(CUOBJS)
$(LINK) $(LFLAGS) $^ $(LIBS) /out:$@

%.obj: %.cu
nvcc -c $< -o $@ $(CUFLAGS)
[/CODE]

Using CUDALucas 1.4, there's an interesting speed regression going from CUDA 3.2 to CUDA 4.1RC2 on my GTX 480 using 64-bit linux:

CUDA 3.2:
Iteration 20000 6.306 msec/Iter M( 45002497 )C, 0x6827d44a3efced00, n = 2621440, CUDALucas v1.4

CUDA 4.1RC2:
Iteration 20000 7.126 msec/Iter M( 45002497 )C, 0x6827d44a3efced00, n = 2621440, CUDALucas v1.4

Note that both of these runs used the 4.1RC2 driver. Only the compiler & libraries differ.

Hi ,frmky
[QUOTE=frmky;283635]CUDA 3.2:
Iteration 20000 6.306 msec/Iter M( 45002497 )C, 0x6827d44a3efced00, n = 2621440, CUDALucas v1.4

CUDA 4.1RC2:
Iteration 20000 7.126 msec/Iter M( 45002497 )C, 0x6827d44a3efced00, n = 2621440, CUDALucas v1.4[/QUOTE]
This is CUFFT issue.
Add information, CUDALucas 1.4 slower than 1.3 with same fft length.