My mistake, it shouldn't be there - at the moment -O1 for ld only turns on some optimizations for shared libraries, not the program binary.
-O3 takes whopping 5Khs/ out of mine super fast GT460
Topic: [ANN] cudaMiner & ccMiner CUDA based mining applications [Windows/Linux/MacOSX] - page 1113. (Read 3426918 times)
April 22, 2013, 04:37:49 PM
-O3 takes whopping 5Khs/ out of mine super fast GT460
April 22, 2013, 04:36:08 PM
That doesn't qualify as almost!
April 22, 2013, 04:29:08 PM
That doesn't qualify as almost!
April 22, 2013, 04:26:29 PM
Posted an April 22nd release.
Please let me know how it compiles on Linux 64 bit, and how it performs on Titan now.
The patch posted earlier wasn't really doing things right. CUDA textures should have stayed ulong2 and ulong4 type, but the uint32_t type needed to be moved over to unsigned long (from unsigned int previously) because otherwise there would be a mismatch with the texture types.
Please let me know how it compiles on Linux 64 bit, and how it performs on Titan now.
The patch posted earlier wasn't really doing things right. CUDA textures should have stayed ulong2 and ulong4 type, but the uint32_t type needed to be moved over to unsigned long (from unsigned int previously) because otherwise there would be a mismatch with the texture types.
Christian,
configure works fine if i run:
./configure -with-cuda=/usr/local/cuda
instead of ./configure.sh
And it almost compiles - http://pastebin.com/raw.php?i=JZb62Jtd
April 22, 2013, 04:26:04 PM
My mistake, it shouldn't be there - at the moment -O1 for ld only turns on some optimizations for shared libraries, not the program binary.
April 22, 2013, 04:20:09 PM
hmm, the patch posted earlier suggests the following configure line for 64 bits
./configure "CFLAGS=-O3" "CXXFLAGS=-O3" "LDFLAGS=-Wl,-O1" --with-cuda=/usr/local/cuda
not sure what the -Wl,-O1 linker flag is supposed to do.
April 22, 2013, 03:55:22 PM
Posted an April 22nd release.
Please let me know how it compiles on Linux 64 bit, and how it performs on Titan now.
The patch posted earlier wasn't really doing things right. CUDA textures should have stayed ulong2 and ulong4 type, but the uint32_t type needed to be moved over to unsigned long (from unsigned int previously) because otherwise there would be a mismatch with the texture types.
Please let me know how it compiles on Linux 64 bit, and how it performs on Titan now.
The patch posted earlier wasn't really doing things right. CUDA textures should have stayed ulong2 and ulong4 type, but the uint32_t type needed to be moved over to unsigned long (from unsigned int previously) because otherwise there would be a mismatch with the texture types.
April 22, 2013, 10:50:06 AM
autoadjust does not find the best values for my titan, had to find the best values 
it works with 70x4 280khash/s
it works with 70x4 280khash/s
it's autotune (TM) (R).
how's 35x8 ?
Christian
April 22, 2013, 10:22:47 AM
autoadjust does not find the best values for my titan, had to find the best values
e: now i checked -D option
it works with 70x4 280khash/s
e: now i checked -D option
it works with 70x4 280khash/s
April 22, 2013, 09:28:48 AM
_shared__ uint32_t X[WARPS_PER_BLOCK][WU_PER_WARP][16+4];
Thanks! This helped. I did not know about newly added alignment restrictions in shared memory targeting SM 2.0 and higher. I guess that's because they're now having a unified pointer and addressing scheme. So if there's an alignment requirement, it applies to everything.
Finally the Titan kernel will get my large memory transaction fixes, which should boost performance notably.
Christian
April 22, 2013, 08:42:05 AM
I've just run into the same compiler issue that borked the Titan kernels when I tried to compile salsa_kernel.cu for sm_30. The kernel will just crash.
Maybe using the NSight debugger I can figure out why this occurs.
Maybe using the NSight debugger I can figure out why this occurs.
Does the crash produce: CUDA_EXCEPTION_6, Warp Misaligned Address ?
I've been able to compile & run salsa_kernel for sm_21, without tex-cache, when accesses to X variable are 128-bit aligned,
ie. when it's declared like this:
Code:
_shared__ uint32_t X[WARPS_PER_BLOCK][WU_PER_WARP][16+4];
April 22, 2013, 08:34:44 AM
how much are you guys getting with a 580?
240KH/s give or take 10KH/s
(slightly OC)as soon ive mined some coins il send a donation for sure
April 22, 2013, 08:13:30 AM
how much are you guys getting with a 580?
240KH/s give or take 10KH/s
April 22, 2013, 07:47:10 AM
how much are you guys getting with a 580?
April 22, 2013, 07:10:28 AM
Seen this.
The challenges with the scrypt hashing are a bit greater than just using the funnel shifter for rotation. One issue is the speed and efficiency of memory access, the other issue is getting enough occupancy on Kepler's SMX (multiprocessor) units - shared memory and register limits are an issue. This mainly affects the GTX 660Ti, GTX 670, 680 and Titan devices which currently perform rather poor in comparison to the 5xx series.
April 22, 2013, 07:07:46 AM
I've seen reports of a single overclocked Titan doing 290 kHash/s, using a somewhat earlier code version.
April 22, 2013, 04:49:34 AM
I've just run into the same compiler issue that borked the Titan kernels when I tried to compile salsa_kernel.cu for sm_30. The kernel will just crash.
Maybe using the NSight debugger I can figure out why this occurs.
Maybe using the NSight debugger I can figure out why this occurs.
April 22, 2013, 03:54:06 AM
April 22, 2013, 01:34:53 AM
Christian,
Success,
copied from settings above but seems to be only 260kH/s per TITAN.
Success,
copied from settings above but seems to be only 260kH/s per TITAN.
April 22, 2013, 01:31:37 AM
this is my 670gtx (GIGABYTE GV-N670OC-2GD) doing over 200khash/s
JESUS CHRIST that is a great OC! Code:
cudaminer.exe --url http://notroll.in:6332/ --userpass jasonharty24.4:12345 -i 0 -m 1 -C 2 -l 70x4
Jump to: