CUDALucas gives all-zero residues

[fivemack]

This is on a new shiny GTX1080, running cuda-8.0-prerelease and Nvidia driver 367.35; I did

svn checkout svn://svn.code.sf.net/p/cudalucas/code/trunk cudalucas-code
cd cudalucas-code
make

and then ...

pumpkin@pumpkin:~/cudalucas-build-cuda80/cudalucas-code$ ./CUDALucas -r 0

------- DEVICE 0 -------
name                GeForce GTX 1080
Compatibility       6.1
clockRate (MHz)     1733
memClockRate (MHz)  5005
totalGlobalMem      8504279040
totalConstMem       65536
l2CacheSize         2097152
sharedMemPerBlock   49152
regsPerBlock        65536
warpSize            32
memPitch            2147483647
maxThreadsPerBlock  1024
maxThreadsPerMP     2048
multiProcessorCount 20
maxThreadsDim[3]    1024,1024,64
maxGridSize[3]      2147483647,65535,65535
textureAlignment    512
deviceOverlap       1

Using threads: square 256, splice 128.
Starting self test M86243 fft length = 4K
Iteration 10000 / 86243, 0x0000000000000000, 4K, CUDALucas v2.05.1, error = 0.01408, real: 0:00, 0.0801 ms/iter
Expected residue [23992ccd735a03d9] does not match actual residue [0000000000000000]

Using threads: square 256, splice 128.
Starting self test M132049 fft length = 8K
Iteration 10000 / 132049, 0x0000000000000000, 8K, CUDALucas v2.05.1, error = 0.00079, real: 0:02, 0.1178 ms/iter
Expected residue [4c52a92b54635f9e] does not match actual residue [0000000000000000]

Using threads: square 256, splice 128.
Starting self test M216091 fft length = 14K
Iteration 10000 / 216091, 0x0000000000000000, 14K, CUDALucas v2.05.1, error = 0.00004, real: 0:01, 0.1285 ms/iter
Expected residue [30247786758b8792] does not match actual residue [0000000000000000]

Using threads: square 256, splice 128.
Starting self test M756839 fft length = 42K
Iteration 10000 / 756839, 0x0000000000000000, 42K, CUDALucas v2.05.1, error = 0.00000, real: 0:02, 0.1647 ms/iter
Expected residue [5d2cbe7cb24a109a] does not match actual residue [0000000000000000]

pumpkin@pumpkin:~/cudalucas-build-cuda80/cudalucas-code$ ldd ./CUDALucas
	linux-vdso.so.1 =>  (0x00007ffcd01d2000)
	libcufft.so.8.0 => /usr/local/cuda-8.0/lib64/libcufft.so.8.0 (0x00007f6f604d4000)
	libcudart.so.8.0 => /usr/local/cuda-8.0/lib64/libcudart.so.8.0 (0x00007f6f60273000)
	libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007f6f5ff59000)
	libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f6f5fb94000)
	libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007f6f5f990000)
	libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f6f5f771000)
	librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007f6f5f569000)
	libstdc++.so.6 => /usr/lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007f6f5f265000)
	libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007f6f5f04e000)
	/lib64/ld-linux-x86-64.so.2 (0x000056154fda2000)

Any ideas? Adding "-polite 1" to the command line makes it slower but otherwise no effect

[airsquirrels]

You need to rebuild with the compute version set correctly. 62 if I recall

This is in the Makefile

[fivemack]

Compute version 61 is apparently right; 62 gives all-zero residues again. Thanks.

[Batalov]

Originally Posted by fivemack

This is on a new shiny GTX1080, running cuda-8.0-prerelease and Nvidia driver 367.35; ...

Driver versions 364.xx .. 367.xx reportedly have bugs. Try to get the 368.22+ driver or roll back to 362.xx

[fivemack]

Originally Posted by Batalov

Driver versions 364.xx .. 367.xx reportedly have bugs. Try to get the 368.22+ driver or roll back to 362.xx

Older drivers don't seem to support GTX1080 properly, and the newest extant Linux driver is 367.35 (2016-07-15).

The first DC I ran was successful.