Closed phobeus81 closed 10 years ago
Use cuda 5.0 toolkit - nvidia still offer it for download
----- Ursprüngliche Nachricht ----- Von: "phobeus81" notifications@github.com Gesendet: 05.12.2013 20:40 An: "cbuchner1/CudaMiner" CudaMiner@noreply.github.com Betreff: [CudaMiner] Unable to compile spinlock_kernel.o (#42)
When trying to compile the most recent version from the repository, nvcc is raising an exception and fails to compile spinlock_kernel.o. make all-recursive make[1]: Entering directory /home/phobeus/src/cudaminer' Making all in compat make[2]: Entering directory/home/phobeus/src/cudaminer/compat' make[3]: Entering directory /home/phobeus/src/cudaminer/compat' make[3]: Nothing to be done forall-am'. make[3]: Leaving directory /home/phobeus/src/cudaminer/compat' make[2]: Leaving directory/home/phobeus/src/cudaminer/compat' make[2]: Entering directory /home/phobeus/src/cudaminer' nvcc -g -O2 -Xptxas "-abi=no -v" -arch=compute_30 --maxrregcount=63 -o spinlock_kernel.o -c spinlock_kernel.cu UNREACHABLE executed! Stack dump:
Okay, if it is not working with Cuda 5.5, I am having an issue as the Cuda 5.0 does not support any GCC higher than 4.7 that is delivered with most recent linux distributions. Just curious about it as it seems to be working fine with other versions from november.
I now use some compiler flags that cause CUDA 5.5 to barf, but give higher performance on 5.0
Okay, confirming. When going back to commit 595a0a901d998fe574f3b299af04d42446c52222 on the Makefiles in the root and the compat, I am able to compile a version running fine with 5.5 as well. Maybe adding some switch for handling this instead? Like said, most recent distributions might have issues compiling against Cuda 5.0. Having an flag in there might be an option. Else configure shall check for version and giving an appropriate error message as the spin compile failure might not be indicating an incompatibility with cuda version.
Same issue on Ubuntu 12.04.3 LTS x86_64 with cuda 5.5:
gcc -std=gnu99 -DHAVE_CONFIG_H -I. -msse2 -fopenmp -pthread -fno-strict-aliasing -I./compat/jansson -g -O2 -MT cudaminer-cpu-miner.o -MD -MP -MF .deps/cudaminer-cpu-miner.Tpo -c -o cudaminer-cpu-miner.o test -f 'cpu-miner.c' || echo './'
cpu-miner.c
mv -f .deps/cudaminer-cpu-miner.Tpo .deps/cudaminer-cpu-miner.Po
gcc -std=gnu99 -DHAVE_CONFIG_H -I. -msse2 -fopenmp -pthread -fno-strict-aliasing -I./compat/jansson -g -O2 -MT cudaminer-util.o -MD -MP -MF .deps/cudaminer-util.Tpo -c -o cudaminer-util.o test -f 'util.c' || echo './'
util.c
mv -f .deps/cudaminer-util.Tpo .deps/cudaminer-util.Po
gcc -std=gnu99 -DHAVE_CONFIG_H -I. -msse2 -fopenmp -pthread -fno-strict-aliasing -I./compat/jansson -g -O2 -MT cudaminer-sha2.o -MD -MP -MF .deps/cudaminer-sha2.Tpo -c -o cudaminer-sha2.o test -f 'sha2.c' || echo './'
sha2.c
mv -f .deps/cudaminer-sha2.Tpo .deps/cudaminer-sha2.Po
g++ -DHAVE_CONFIG_H -I. -msse2 -fopenmp -pthread -fno-strict-aliasing -I./compat/jansson -g -O2 -MT cudaminer-scrypt.o -MD -MP -MF .deps/cudaminer-scrypt.Tpo -c -o cudaminer-scrypt.o test -f 'scrypt.cpp' || echo './'
scrypt.cpp
mv -f .deps/cudaminer-scrypt.Tpo .deps/cudaminer-scrypt.Po
nvcc -g -O2 -arch=compute_10 --maxrregcount=124 --ptxas-options=-v -I./compat/jansson -o salsa_kernel.o -c salsa_kernel.cu
nvcc -g -O2 -Xptxas "-abi=no -v" -arch=compute_30 --maxrregcount=63 -I./compat/jansson -o spinlock_kernel.o -c spinlock_kernel.cu
UNREACHABLE executed!
Stack dump:
Will try to revert to Cuda 5.0 and report back.
I got this error. nvcc -g -O2 -Xptxas "-abi=no -v" -arch=compute_30 --maxrregcount=63 -o spinlock_kernel.o -c spinlock_kernel.cu UNREACHABLE executed! Stack dump:
using CUDA 5.0 instead of CUDA 5.5 would be the solution.
2013/12/13 Armando notifications@github.com
I got this error.
nvcc -g -O2 -Xptxas "-abi=no -v" -arch=compute_30 --maxrregcount=63 -o spinlock_kernel.o -c spinlock_kernel.cu UNREACHABLE executed! Stack dump:
- Running pass 'NVPTX DAG->DAG Pattern Instruction Selection' on function '@_Z28spinlock_scrypt_core_kernelBILi3EEvPj' Aborted
— Reply to this email directly or view it on GitHubhttps://github.com/cbuchner1/CudaMiner/issues/42#issuecomment-30550815 .
not CUDA 5.0 any more
https://developer.nvidia.com/cuda-toolkit-50-archive
2013/12/14 snakkewang notifications@github.com
not CUDA 5.0 any more
— Reply to this email directly or view it on GitHubhttps://github.com/cbuchner1/CudaMiner/issues/42#issuecomment-30562227 .
Also getting that error.
CUDA 5.0 is outdated (two years, since 2012) and is not an option for some cars like mine - GeForce 640 which is not detected as CUDA supporting. CUDA 5.5 works fine and detects my card.
Please make the miner working with CUDA 5.5.
best, Artem
I am mining on three GT 640, two GK108 and one GK208 chips running cudaminer binaries built with CUDA 5.0 (using fairly recent nVidia drivers)
also have a look at this https://github.com/cbuchner1/CudaMiner/blob/master/README
2013/12/14 Artiom Neganov notifications@github.com
Also getting that error.
CUDA 5.0 is outdated (two years, since 2012) and is not an option for some cars like mine - GeForce 640 which is not detected as CUDA supporting. CUDA 5.5 works fine and detects my card.
Please make the miner working with CUDA 5.5.
best, Artem
— Reply to this email directly or view it on GitHubhttps://github.com/cbuchner1/CudaMiner/issues/42#issuecomment-30583511 .
its not a recommend to do cuda 5.0, My OS break becouse need to downgrade to and old versio of gcc
Hi there, those wanting to use cuda 5.5 can simply change the Makefile. Where it says:
spinlock_kernel.o: spinlock_kernel.cu $(NVCC) -Xptxas "-abi=no -v" -arch=compute_11 --maxrregcount=124 $(JANSSON_INCLUDES) -o $@ -c $<
revert to old flags:
$(NVCC) -Xptxas "-abi=no -v" -arch=compute_35 --maxrregcount=64 $(JANSSON_INCLUDES) -o $@ -c $<
et voila'.
Apologies for the previous comment, it is confusing. I meant that old compile flags "-arch=compute_11 --maxrregcount=124" will compile the spinlock_kernel using cuda 5.5 nvcc. Anyway this will produce a cudaminer that is slightly slower than the older for my testing hardware (GT 610).
Hi. I'm trying to compile for Cuda 5.5 also and am confused.. I tried the Makefile change in your earlier post and now get:
spinlock_kernel.cu(340): error: identifier "atomicCAS" is undefined
spinlock_kernel.cu(345): error: identifier "atomicExch" is undefined
2 errors detected in the compilation of "/tmp/tmpxft_00002640_00000000-6_spinlock_kernel.cpp1.ii".
well it compile and i got this meesage runnind.
GPU #0: maximum warps: 217 GPU #0: 3.97 khash/s with configuration L4x3 using launch configuration L4x3 GeForce 9400 GT, 4224 hashes, 1.04 khash/s
its slow.
Hey Guys, I dont know how to compile, and I dont know what a nvcc is. I would love to learn more, any help or direction would be greatly appreciated. Happy Holidays!!! TY!
There is a chance that I will migrate the project to CUDA 5.5, because I am seeing a 20-30% gain with CUDA 5.5 only while it is limited to 5-10% using CUDA 5.0. This is with Dave Andersen's improved Keplerminer code, integrated into cudaminer.
First what plataform are you using (windows, mac or linux) ?
nvcc its the compiler for CUDA.
I am using Win 7 64 bit. Oh nvcc is the compiler. Awesome! I have to now learn how to compile. :)
Hi, got here the same issue until last 57dd2bf08fd592eb2025355f0abe598438ee1b9b commit. Now, it builds almost fine but with warnings under linux Fedora 19, x86_64 and with CUDA 5.5. I'm having 15.2kH/s now with my GeForce 9600 GT, x7.5 against 2kH/s with cgminer. Thank you for that good work.
Yep! now compiles on CUDA 5.5 (just tested on stock Ubuntu 12.04 x86_64)
auto-tuning choose kernel F44x5 with a 14.43 khash/s which makes it now as fast as the old spinlock version
cheers Christian!
this is an outdated ticket. closing.
When trying to compile the most recent version from the repository, nvcc is raising an exception and fails to compile spinlock_kernel.o.
make all-recursive make[1]: Entering directory
/home/phobeus/src/cudaminer' Making all in compat make[2]: Entering directory
/home/phobeus/src/cudaminer/compat' make[3]: Entering directory/home/phobeus/src/cudaminer/compat' make[3]: Nothing to be done for
all-am'. make[3]: Leaving directory/home/phobeus/src/cudaminer/compat' make[2]: Leaving directory
/home/phobeus/src/cudaminer/compat' make[2]: Entering directory `/home/phobeus/src/cudaminer' nvcc -g -O2 -Xptxas "-abi=no -v" -arch=compute_30 --maxrregcount=63 -o spinlock_kernel.o -c spinlock_kernel.cu UNREACHABLE executed! Stack dump:make: *\ [all] Error 2
When using a older version with commit "88c6da6d5c2b798d1de7031e8dbcc2678f635e4b" it seems to be compiling fine.