Re: CUDA Wrapper err:module:import_dl

[James McKenzie]

Seth Shelnutt wrote:
> |Alright, I'm stumped and everyone else is too. Few years ago we did 
> everything we are doing now and it all worked fine. Now using wine 
> 1.2rc2 we can't get a wrapper I compiled to work. What this wrapper 
> does is it translates/wraps the calls made to the cuda runtime dll 
> (cudart.dll or nvcuda.dll, same file different names) and sends it to 
> the linux library. The linux library and windows are identical except 
> for some missing dx functions in the linux library. My sources can be 
> found here. http://shelnutt2.host56.com/cudart/ and 
> http://shelnutt2.host56.com/cufft/ (However we aren't even getting to 
> the second dll yet). Anytime we try to run folding at home in wine, we 
> are getting the following. FAH checks for nvcuda.dll first because 
> that comes with the driver installs on windows, if it doesn't find it, 
> it defaults to cudart.dll which is packaged with fah, or can be 
> installed separately. We have replaced all the cuda dlls with the 
> compiled wrapper or sum linked to the compiled wrapper, just like we 
> use to with the original cuda wrapper that we came up with.
>
> err:module:import_dll Loading library nvcuda.dll (which is needed by 
> L"Z:\home\sidicas\.FAHGPU3\FahCore_11.exe") failed (error c0000020).
> err:module:import_dll Loading library cudart.dll (which is needed by 
> L"Z:\home\sidicas\.FAHGPU3\FahCore_11.exe") failed (error c0000020).
>
> |
> |sidicas@Belldandy:~/.FAHGPU3$ ls -al *.dll
> -rw-r--r-- 1 sidicas sidicas 1617920 2010-06-03 01:48 amdcalcl.dll
> -rw-r--r-- 1 sidicas sidicas 2486272 2010-06-03 01:48 amdcalrt.dll
> -rw-r--r-- 1 sidicas sidicas  124540 2010-06-02 07:50
> cudart32_30_14.dll
> -rw-r--r-- 1 sidicas sidicas  124540 2010-06-02 07:50 cudart.dll
> -rw-r--r-- 1 sidicas sidicas   82699 2010-06-01 21:39
> cufft32_30_14.dll
> -rw-r--r-- 1 sidicas sidicas   82699 2010-06-01 21:39 cufft.dll
> lrwxrwxrwx 1 sidicas sidicas      10 2010-06-03 02:12 nvcuda.dll
> -> cudart.dll
> sidicas@Belldandy:~/.FAHGPU3$                              |
>
>
> |sidicas@Belldandy:~/.FAHGPU3$ uname -a
> Linux Belldandy 2.6.26-2-amd64 #1 SMP Sun Jun 21 04:47:08 UTC 2009
> x86_64 GNU/Linux
> sidicas@Belldandy:~/.FAHGPU3$ nvcc --version
> nvcc: NVIDIA (R) Cuda compiler driver
> Copyright (c) 2005-2009 NVIDIA Corporation
> Built on Fri_Feb_19_18:18:31_PST_2010
> Cuda compilation tools, release 3.0, V0.2.1221
> sidicas@Belldandy:~/.FAHGPU3$ ldd ./cudart.dll
>         linux-gate.so.1 =>  (0xf7ee3000)
>         libcudart.so.3 => /usr/local/cuda/lib/libcudart.so.3
> (0xf7e83000)
>         libwine.so.1 => /usr/lib/libwine.so.1 (0xf7d2d000)
>         libm.so.6 => /lib32/libm.so.6 (0xf7d09000)
>         libc.so.6 => /lib32/libc.so.6 (0xf7bb6000)
>         libdl.so.2 => /lib32/libdl.so.2 (0xf7bb2000)
>         libpthread.so.0 => /lib32/libpthread.so.0 (0xf7b9b000)
>         librt.so.1 => /lib32/librt.so.1 (0xf7b92000)
>         libstdc++.so.6 => /usr/lib32/libstdc++.so.6 (0xf7aa4000)
>         libgcc_s.so.1 => /usr/lib32/libgcc_s.so.1 (0xf7a97000)
>         /lib/ld-linux.so.2 (0xf7ee4000)
> sidicas@Belldandy:~/.FAHGPU3$
> sidicas@Belldandy:~/.FAHGPU3$ ldd ./cufft.dll
>         linux-gate.so.1 =>  (0xf7f29000)
>         libcufft.so.3 => /usr/local/cuda/lib/libcufft.so.3
> (0xf7308000)
>         libwine.so.1 => /usr/lib/libwine.so.1 (0xf71b2000)
>         libm.so.6 => /lib32/libm.so.6 (0xf718e000)
>         libc.so.6 => /lib32/libc.so.6 (0xf703b000)
>         libdl.so.2 => /lib32/libdl.so.2 (0xf7037000)
>         libpthread.so.0 => /lib32/libpthread.so.0 (0xf7020000)
>         libcudart.so.3 => /usr/local/cuda/lib/libcudart.so.3
> (0xf6fdf000)
>         libstdc++.so.6 => /usr/lib32/libstdc++.so.6 (0xf6ef1000)
>         libgcc_s.so.1 => /usr/lib32/libgcc_s.so.1 (0xf6ee4000)
>         /lib/ld-linux.so.2 (0xf7f2a000)
>         librt.so.1 => /lib32/librt.so.1 (0xf6eda000)
> sidicas@Belldandy:~/.FAHGPU3$
> sidicas@Belldandy:~/.FAHGPU3$        
> |
>
>
> More thorugh testing has been done, 
> http://foldingforum.org/viewtopic.php?f=54&t=14673&start=105#p146177 
> http://foldingforum.org/viewtopic.php?f=54&t=14673&start=105#p146177 
> . When we makes reference to gpu2 wrapper, that is our original cuda 
> wrapper designed for cuda 2.0-2.2 built couple years ago. The GPU3 
> wrapper is what we are having problems with. It is our cuda 3.0 
> wrapper. The .dll files are in the folder (we even tried in system32), 
> ldd shows they are properly linked, yet it does not work. If we use 
> the old wrapper compile before it works fine with cuda 2.x or with 
> cuda 3.0 we just have to sym link the cuda lib to the old name it's 
> expecting.
>
> I've tried compiling the old wrapper and linking it to cuda 2.3 
> however this fails. So I am thinking that if when I compile the old 
> wrapper which is shown to work (when an old compiled version is used) 
> but does not work when I compile on my funtoo box it must be some 
> weird compiling issue. I have wine installed (without 64bit support), 
> and I installed cuda runtime from nvidia's site as the one in portage 
> installs to /opt/ and I wanted to make sure there wasn't any weird 
> issues linking with that.
>
> I do get a few warnings when I compile but they aren't anything that 
> would effect this.
>
> s.shelnutt@funtoo-64 ~/gpu3/wrapper/cudart $ make
> winegcc -c  -mno-cygwin -m32 -I/usr/local/cuda/include/   -o cudart.o 
> cudart.c
> In file included from /usr/local/cuda/include/cuda_runtime_api.h:53,
>                  from /usr/local/cuda/include/crt/host_runtime.h:108,
>                  from cudart.c:7:
> /usr/local/cuda/include/host_defines.h:57:1: warning: "__cdecl" redefined
> <command-line>: warning: this is the location of the previous definition
> cudart.c: In function 'wine_cudaGetTextureReference':
> cudart.c:430: warning: passing argument 1 of 'cudaGetTextureReference' 
> from incompatible pointer type
> /usr/local/cuda/include/cuda_runtime_api.h:189: note: expected 'const 
> struct textureReference **' but argument is of type 'struct 
> textureReference **'
> cudart.c: In function 'wine_cudaRegisterFatBinary':
> cudart.c:800: warning: return makes pointer from integer without a cast
> cudart.c: In function 'wine_cudaUnregisterFatBinary':
> cudart.c:805: warning: 'return' with a value, in function returning void
> cudart.c: In function 'wine_cudaRegisterVar':
> cudart.c:810: warning: 'return' with a value, in function returning void
> cudart.c: In function 'wine_cudaRegisterTexture':
> cudart.c:815: warning: 'return' with a value, in function returning void
> cudart.c: In function 'wine_cudaRegisterShared':
> cudart.c:820: warning: 'return' with a value, in function returning void
> cudart.c: In function 'wine_cudaRegisterSharedVar':
> cudart.c:825: warning: 'return' with a value, in function returning void
> cudart.c: In function 'wine_cudaRegisterFunction':
> cudart.c:831: warning: 'return' with a value, in function returning void
> winegcc -shared cudart.dll.spec -mno-cygwin -o cudart.dll.so 
> http://cudart.dll.so cudart.o    -L/usr/local/cuda/lib/ -lcudart   
> -lodbc32 -lole32 -loleaut32 -lwinspool -lodbccp32 -luuid
>
> If anyone has any ideas about why this doesn't work, that would be 
> great. I've exhausted everything I could think of.

Seth:

Does this work with Wine-1.0.1?  If so, we have a regression...

You could run a mostly lengthy regression test and find the culprit if 
it is.

James McKenzie