Yeah, I can't wait for m7m to be back. I'm anxious to finally see the performance difference with SHA optimizations.
cpuminer-opt says Yescrypt uses both AVX2 and SHA, but I don't see much difference in hashrate.
That isn't quite correct, there's no AVX in yescrypt (I'll fix that), but there is SHA. Does it say it's mining with SHA? If Linux do you have openssl 1.1? If Windows are you using SHA build? |
|
|
|
Sir, where did you find this info? Cache Block diagram looks exactly like the one from sandy bridge times, everything seems to be the same as inside kaby.
Nope, the structure is the same, but the amount increased. This leads to better mining performance in cache-bound algos, and better flexibility (w/o looking at core count). E.g. on i5 8600K you may now use 2 double threads for cryptonight (2x4Mb) at around 320+ H/s and you will be able to use the rest (1Mb) for video decoding (which uses cache as well). This is good, and affects power consumption and heating (in a good way) as well. It's the same 2Mb per core on i7 and 1.5Mb per core on i5, just more cores. So no real changes There's no chnage in architecture that I'm aware of, everything seems to be scaled 1.5 vs previous generation, no compile differences, no source code differences. It's not a SW issue, I'm doubtful it's a Coffeelake issue. It might be an issue with that specific CPU or system. More data is required including from other Coffeelake owners. OP has been quiet for a couple of days, maybe it's no longer an issue. |
|
|
|
is it just me or is this super difficult to install? i'm an unix/dba admin (not a programmer) and this is impossible to compile. trying to compile it with static libs so i can use it across multiple machines, but no matter what i do (yes i followed all tutorials online) nothin works
There's only one tutorial that matters, README.md See file RELEASE_NOTES for change log and compile instructions.
|
|
|
|
few benchmarks on yescryptr16
3.7.7 "v1" (gcc 4.8.3)
avx - ~930 h/s
sse2 - ~ 870 h/s
3.7.7 v2 (gcc 5.3.1)
avx - ~950 h/s
sse2 - ~970 h/s
3.7.7 4ward (gcc 6.2.1)
avx - ~970 h/s
sse2 - ~960 h/s
additional algos that show better performance in sse2 than avx (although very small):
yescrypt, poltimos and lbry
This shows gcc-5.3.1 might be the issue. Is that on a Coffeelake? If not it eliminates that as a Coffeelake issue and looks purely like a compiler version issue. |
|
|
|
If Coffeelake has a design quirk that can be worked around in software
such a workaround would probably have a negative effect on other models. If it's a coffee lake issue it needs a Coffeelake fix. I suppose that populating scrypt algo parameters to command line can help a lot, if yescript, like regular scrypt, can be calced with different algo presets (precache amount, link split size and so on), then coffeelake make take profit from better cache-fitting splitting. Anyway tomorrow i'll try to recompile miner with different presets in scrypt.c Again I speculate but maybe the
compiler isn't yet tweaked for Codffeelake. What version did you compile with? i currently use windows precompiled versions on both 5820 and 8700, tomorrow i'll try latest gcc with skylake opt flag, but i suppose that compiler won't make any changes inside asm instruction so opt flag won't help, at least a lot. You stated lower performance with fewer threads. That is mostly windows problem - with 6 threads it uses only 3 physical cores and 3 ht cores - clearly seen with cputemp - after start of 6 threads, 3 cores start to generate heat (70-75C on busy ones, 45C on spare) and shows 100% load, with 12 threads all cores are hot and busy Under ubuntu difference is within the margin of error I suggest you try other algos with 6 and 12 threads to get a more complete profile. If some algos are affected more than others
it may reveal a pattern. yep, i'll try If 6 threads aren't balanced use custom cpu affinity (--cpu-affinity 0x555) It hasn't been an issue on Intel before but if you say only 3 cores are heating up then maybe the mapping has changed. There is no ASM but there is hardcoded SSE2 code but none between SSE2 and AVX. The only differences between SSE2 and AVX compile are generated by the compiler. |
|
|
|
Looks like there is some performance problem with yescript16 implementation on coffee-lake cpus Intel i5 4440 (stock 3.3GHz) @4 threads generates ~600h/s Intel i7 5820k (no overclock, 3.6GHz) @12 threads generates ~1200h/s in pool and up to 1400 in solo mining (6 threads generate little less) with both cpuminer-opt (3.7.6 and 3.7.7v2) under windows64 and under ubuntu64 Intel i7 8700k (stock 4.3GHz) @12 threads generates only 950h/s (both pool and solo), overclocking to 5Ghz (50x100) with cache overclock to 4.6Ghz (stock is 4.2) gives no profit, even more usually performance degrade (power limit disabled, core temperatures are ~75C so no throttling involved), 1-2-3-4-5-6 threads gives less results, overclocking bus to 130Mhz (also with ram) gives no result - maximum is about 950h/s Even more funny - on stock frequency, switching from AVX to SSE2 gives some performance boost from 950 to 1000-1050h/s I understand that 8700k lacks quad-channel RAM and has little bit less L3 cache (12 vs 15Mb), compared to 5820, but bottleneck is obviously something different because ram overclock gives no result (so double channel is not a problem, we should see performance boost when overclocking bus and ram) and cache is also not a problem (25% cache is gone but we gain >30% frequency bonus (when overclocked) so our smaller cache works at higher speeds together with cpu cores - we can put less but more frequent and calc it in less time - ) also, compared to 4440, if cache was a bottleneck, we have twice more (12 vs 6Mb), taking in mind much higher speed and optimized pipelane, if cache only matters, we should have 2x gain, compared to 4440 I hope for a fix  That's quite a first post, you did your homework. Your results are concerning but not a software issue. If Coffeelake has a design quirk that can be worked around in software such a workaround would probably have a negative effect on other models. If it's a coffee lake issue it needs a Coffeelake fix. On the technical side it's difficult to compare the 8700K with either of the two other CPUs you tested. I have a 6700K @ 4 GHz and it gets 780 H/s. With a projected linear increase the 8700K should get around 1170. Clearly the 8700k has a problem. I haven't done a deep dive into the architecture to see if there is a design change that could have an effect. As a new CPU it could still have a few issues that need to be ironed out. I noticed in my brief test that reducing the thread count by half on my 6700K had no effect on total hash rate. This tells me the bottleneck is memory access (cache or main). You stated lower performance with fewer threads. That may be a clue. If your CPU is not I/O bound when mining an I/O bound algo there may be a problem on the compute side. Your observation that SSE2 build is faster than AVX is very interesting and deserves more testing. There is no AVX specific code in yescrypt so there should be no difference in hash speed. Yescrypt is also very self contained, ie it doesn't use any libraries. The only effect the AVX flag would have is on the compiler. It may compile code differently but there are no big gains between SSE2 and AVX, they are both mostly limited to 128 bit vectors. It's only with AVX2 that there is a quantum leap to 256 bit vectors. Again I speculate but maybe the compiler isn't yet tweaked for Codffeelake. What version did you compile with? I suggest you try other algos with 6 and 12 threads to get a more complete profile. If some algos are affected more than others it may reveal a pattern. It would also be interesting to see if other Coffelake owners see the same issues. |
|
|
|
it does not depend on algo, even if I trying to get help noting displayed. CPU Celeron G3930
C:\App\cpuminer-opt-3.7.7-windows-v2>cpuminer-avx2.exe --help
C:\App\cpuminer-opt-3.7.7-windows-v2>cpuminer-avx.exe --help
C:\App\cpuminer-opt-3.7.7-windows-v2>cpuminer-4way.exe --help
C:\App\cpuminer-opt-3.7.7-windows-v2>cpuminer-4way.exe
C:\App\cpuminer-opt-3.7.7-windows-v2>
First of all your CPU doesn't have AVX or AVX2 so stay away from those and 4way. You should be using aes-sse42. But there's a bigger problem if it won't display help. I've never seen that kind of a problem, it looks like it's your system. No one else is complaining so the problem is at your end. |
|
|
|
I have a problem running cpuminer-opt v3.7.x on a couple of machines with win 10 x64. when started, it silently waits several seconds, then exit, without writing symbol. any advice? do I miss some runtime libs ?
Advice? Yes, provide proper information. What CPU, algo, command line options? It's all displayed when the program starts, always provide that when reporting a problem. |
|
|
|
Damn! You are right Too many "Lyras"  If you're comparing AVX2 vs AVX performance lyra2v2 is better, less memory hard. |
|
|
|
tried this // #if (WINDOWS)
/* sizeof(unsigned long) = 4 for MinGW64 */
// typedef unsigned long long ulong;
// #else
typedef unsigned long ulong;
// #endif
typedef unsigned int uint; same problem, unless I'm doing it wrong ) It's not wrong, thanks. I'm at a loss to explain it. I could try to find out where it's crasshing but it may not be necessary. I noticed some optimization opportunities that I hadn't noticed before but it will require a complete rewrite. |
|
|
|
Making 2 packages is more work for me, I'm trying to reduce the work.
I'm leaning toward replacing avx2-sha with avx-sha unless someone shows a good
reason for avx2-sha.
Edit: a couple more points
AVX2 and SHA improve different algos and different parts of the same algos. AVX won't have
any effect on SHA code on Ryzen CPUs. There are no technical concerns with AVX-SHA.
The only question is performance on algos that have use AVX2. The best algo to test this is
lyra2v2, it is almost 100% AVX2 and not too hard on memory so it will expose any weaknesses
in AVX2 on Ryzen.
I'm getting around 1.6~2MH/s in Lyra2z. But I'm getting low difficult share errors on every share: [2017-12-19 17:45:42] Starting Stratum on stratum+tcp://us-east.lyra2z-hub.miningpoolhub.com:20581
[2017-12-19 17:45:42] 16 miner threads started, using 'lyra2rev2' algorithm.
[2017-12-19 17:45:43] Stratum difficulty set to 10
[2017-12-19 17:46:05] CPU #9: 2097.15 kH, 116.83 kH/s
[2017-12-19 17:46:05] CPU #11: 2097.15 kH, 116.65 kH/s
[2017-12-19 17:46:05] CPU #7: 2097.15 kH, 116.33 kH/s
[2017-12-19 17:46:06] CPU #15: 2097.15 kH, 114.68 kH/s
[2017-12-19 17:46:06] CPU #6: 2097.15 kH, 111.85 kH/s
[2017-12-19 17:46:06] CPU #3: 2097.15 kH, 111.58 kH/s
[2017-12-19 17:46:06] CPU #14: 2097.15 kH, 110.05 kH/s
[2017-12-19 17:46:07] CPU #10: 2097.15 kH, 108.93 kH/s
[2017-12-19 17:46:07] CPU #5: 2097.15 kH, 107.02 kH/s
[2017-12-19 17:46:07] CPU #8: 2097.15 kH, 106.52 kH/s
[2017-12-19 17:46:07] CPU #2: 2097.15 kH, 106.20 kH/s
[2017-12-19 17:46:09] CPU #1: 2097.15 kH, 98.64 kH/s
[2017-12-19 17:46:10] CPU #13: 2097.15 kH, 93.44 kH/s
[2017-12-19 17:46:16] CPU #4: 2097.15 kH, 73.31 kH/s
[2017-12-19 17:46:16] CPU #0: 2097.15 kH, 73.11 kH/s
[2017-12-19 17:46:20] CPU #12: 2097.15 kH, 64.31 kH/s
[2017-12-19 17:46:39] Stratum difficulty set to 7
[2017-12-19 17:46:42] CPU #13: 3174.84 kH, 98.57 kH/s
[2017-12-19 17:46:42] Rejected 1/1 (100.0%), 34.63 MH, 1634.59 kH/s
[2017-12-19 17:46:42] reject reason: low difficulty share of 8.935987400308036e-8
[2017-12-19 17:46:42] factor reduced to : 0.67 Is it miner-related or pool-related? User error, look carefullly at the algo. Another note about lyra2z, 4way is likely slower due to previously mentioned issues with it. |
|
|
|
One thing I couldn't resolve is that neoscrypt always fails to run (same in My9bot's version)
I can look into this now. -DNOASM didn't solve the issue, it still fails. But you are right, this algo is definitely not the best performer, so I don't know if its worth investing much time in it I found a potential problem that could cause data misalignment. In algo/neoscrypt.c line 56: #if (WINDOWS)
/* sizeof(unsigned long) = 4 for MinGW64 */
typedef unsigned long long ulong;
#else
typedef unsigned long ulong;
#endif
typedef unsigned int uint;
Further down in the code some data defined as uint is accessed as ulong which has stricter alignment requirements if defined as long long. If the condition is removed and ulong is the same as uint it shouldn't crash. I found no other compilation divergence and alignment bugs can slip through unnoticed for a time. The trigger seems to be the different compile environment. Can you test to confirm? |
|
|
|
|
Making 2 packages is more work for me, I'm trying to reduce the work.
I'm leaning toward replacing avx2-sha with avx-sha unless someone shows a good
reason for avx2-sha.
Edit: a couple more points
AVX2 and SHA improve different algos and different parts of the same algos. AVX won't have
any effect on SHA code on Ryzen CPUs. There are no technical concerns with AVX-SHA.
The only question is performance on algos that have use AVX2. The best algo to test this is
lyra2v2, it is almost 100% AVX2 and not too hard on memory so it will expose any weaknesses
in AVX2 on Ryzen.
|
|
|
|
One thing I couldn't resolve is that neoscrypt always fails to run (same in My9bot's version)
I can look into this now. I suspect the neoscrypt problem is related to its use of asm. It is controlled by a NOASM compile flag. I didn't find any hooks for Windows so it should compile the same as Linux. It works with the old mingw environment used to build previous binaries. The only obvious difference is the mingw host, The old way was built on a Windows host, the new ones cross compiled on a Linux host. I don't know if that makes any difference. Digging deeper will be very time consuming and considering neoscrypt performance on CPU it's hardly worth the effort, especially if the result is to disable ASM and performance drops. An easy test would be to recompile with -DNOASM and see if it works and how the performance compares with the old mingw build. I'll leave it up to you if you're interested. The results will help determine how to proceed. |
|
|
|
I've uploaded a new windows binaries package of v3.7.7 to git with support for SHA. I also trimmed some of the file names to remove redundancy. 4way includes avx2 avx2 includes avx avx includes aes and sse4.2 It's avaiable on the releases page or this direct link: https://github.com/JayDDee/cpuminer-opt/files/1569739/cpuminer-opt-3.7.7-windows-v2.zipBe careful with sha, only the AMD Ryzen family supports it at this time. Consider this a beta for the new Windows build system. Great! SHA works great! Your miner is blazing fast! I will do the benchmarks like I did with 4ward build. Apologize me, but can I ask one more build? SHA AVX (not AVX2) version. I explain: Ryzen doesn't have a "fine" AVX2 implementation. In some algos, AVX is faster than AVX2. We are trying to find the best setup with SHA, so maybe AVX setup can give one more option to try. What do you think? That's a very reasonable request and doable. However I'm trying to reduce the number of binaries I build, 8 is too many. Also I'm hesitant to "downgrade" the technology, it just doesn't feel right. Let me think about it and I'll do something for next release. I'm considering eliminating the sse42 build. There are no specific optimizations targetting sse42 so there should no performance loss when using the sse2 build. If there are no reports showing a benefit to the sse42 build it will be removed. Edit: I could replace the avx2-sha build with avx-sha. It won't affect any Intel users until Cannonlake next year and Ryzen users can choose the 4way build if they prefer AVX2. Thoughts? |
|
|
|
One thing I couldn't resolve is that neoscrypt always fails to run (same in My9bot's version)
I can look into this now. |
|
|
|
I've uploaded a new windows binaries package of v3.7.7 to git with support for SHA. I also trimmed some of the file names to remove redundancy. 4way includes avx2 avx2 includes avx avx includes aes and sse4.2 It's avaiable on the releases page or this direct link: https://github.com/JayDDee/cpuminer-opt/files/1569739/cpuminer-opt-3.7.7-windows-v2.zipBe careful with sha, only the AMD Ryzen family supports it at this time. Consider this a beta for the new Windows build system. |
|
|
|
since then i compiled it all to a script: #!/bin/bash
OPENSSL_VER='1.1.0g'
CURL_VER='7.56.1'
GMP_VER='6.1.2'
DIR=$(pwd)
CORES=$(grep -c ^processor /proc/cpuinfo)
sudo apt-get install build-essential libz-mingw-w64-dev libz-mingw-w64 git autoconf -y
[[ -f openssl-${OPENSSL_VER}.tar.gz ]] || wget https://www.openssl.org/source/openssl-${OPENSSL_VER}.tar.gz
[[ -d openssl-${OPENSSL_VER} ]] || tar xvfz openssl-${OPENSSL_VER}.tar.gz
cd openssl-${OPENSSL_VER}
./Configure mingw64 shared --cross-compile-prefix=x86_64-w64-mingw32- || exit 1
make -j ${CORES} || exit 1
cd ${DIR}
[[ -f curl-${CURL_VER}.tar.bz2 ]] || wget https://curl.haxx.se/download/curl-${CURL_VER}.tar.bz2
[[ -d curl-${CURL_VER} ]] || tar xvfj curl-${CURL_VER}.tar.bz2
cd curl-${CURL_VER}
./configure --with-winssl --with-winidn --host=x86_64-w64-mingw32 || exit 1
make -j ${CORES} || exit 1
cd ${DIR}
[[ -f gmp-${GMP_VER}.tar.bz2 ]] || wget https://gmplib.org/download/gmp/gmp-${GMP_VER}.tar.bz2
[[ -d gmp-${GMP_VER} ]] || tar xvfj gmp-${GMP_VER}.tar.bz2
cd gmp-${GMP_VER}
./configure --host=x86_64-w64-mingw32 || exit 1
make -j ${CORES} || exit 1
cd ${DIR}
[[ -d cpuminer-opt ]] || git clone https://github.com/JayDDee/cpuminer-opt.git
cd cpuminer-opt
ln -s ../gmp-${GMP_VER}/gmp.h
cat > winbuild-cross.sh << EOF
#!/bin/bash
export LDFLAGS="-L${DIR}/curl-${CURL_VER}/lib/.libs -L${DIR}/gmp-${GMP_VER}/.libs -L${DIR}/openssl-${OPENSSL_VER}"
F="--with-curl=${DIR}/curl-${CURL_VER} --with-crypto=${DIR}/openssl-${OPENSSL_VER} --host=x86_64-w64-mingw32"
sed -i 's/"-lpthread"/"-lpthreadGC2"/g' configure.ac
mkdir release
cp /usr/x86_64-w64-mingw32/lib/zlib1.dll release/
cp /usr/x86_64-w64-mingw32/lib/libwinpthread-1.dll release/
cp /usr/lib/gcc/x86_64-w64-mingw32/6.2-win32/libstdc++-6.dll release/
cp /usr/lib/gcc/x86_64-w64-mingw32/6.2-win32/libgcc_s_seh-1.dll release/
cp ${DIR}/openssl-${OPENSSL_VER}/libcrypto-1_1-x64.dll release/
cp ${DIR}/curl-${CURL_VER}/lib/.libs/libcurl-4.dll release/
make distclean || echo clean
rm -f config.status
./autogen.sh || echo done
CFLAGS="-O3 -march=core-avx2 -Wall -DFOUR_WAY" ./configure \$F
make -j ${CORES}
strip -s cpuminer.exe
mv cpuminer.exe release/cpuminer-4way.exe
make clean || echo clean
rm -f config.status
./autogen.sh || echo done
CFLAGS="-O3 -march=core-avx2 -Wall" ./configure \$F
make -j ${CORES}
strip -s cpuminer.exe
mv cpuminer.exe release/cpuminer-aes-avx2.exe
make clean || echo clean
rm -f config.status
./autogen.sh || echo done
CFLAGS="-O3 -march=znver1 -Wall" ./configure \$F
make -j ${CORES}
strip -s cpuminer.exe
mv cpuminer.exe release/cpuminer-aes-sha.exe
make clean || echo clean
rm -f config.status
./autogen.sh || echo done
CFLAGS="-O3 -march=corei7-avx -Wall" ./configure \$F
make -j ${CORES}
strip -s cpuminer.exe
mv cpuminer.exe release/cpuminer-aes-avx.exe
make clean || echo clean
rm -f config.status
./autogen.sh || echo done
CFLAGS="-O3 -maes -msse4.2 -Wall" ./configure \$F
make -j ${CORES}
strip -s cpuminer.exe
mv cpuminer.exe release/cpuminer-aes-sse42.exe
make clean || echo clean
rm -f config.status
./autogen.sh || echo done
CFLAGS="-O3 -march=corei7 -Wall" ./configure \$F
make -j ${CORES}
strip -s cpuminer.exe
mv cpuminer.exe release/cpuminer-sse42.exe
make clean || echo clean
rm -f config.status
./autogen.sh || echo done
CFLAGS="-O3 -march=core2 -Wall" ./configure \$F
make -j ${CORES}
strip -s cpuminer.exe
mv cpuminer.exe release/cpuminer-sse2.exe
EOF
chmod a+x winbuild-cross.sh
./winbuild-cross.sh
This is perfect. It's exactly what I needed. Since 4way is somewhat stalled right now I tried it out in a Ubuntu-16.04 VM. I ran the script manually, had to make adjustments for compiler version, and lack or "znver1" arch but I built a 4way with sha. I just can't run it anywhere. I just need to tweak a few things and the next release should have the full suite of binaries, with the addition of AVX2-SHA and 4WAY-SHA. |
|
|
|
Is there a possibility to make a SHA version with AES-AVX/AES-AVX2/4WAY to improve the tests?
There should be no problem adding 4way to the Ryzen build. 4way only needs AVX2 and all Ryzens have it. Itcould become an issue in the future if some low-budget CPUs are built with SHA but not AVX2. Edit: At this time the only algo that can use both 4way and SHA is skein. |
|
|
|
I have i7-7800X CPU @ 3.50GHz
Trying to see which CPU miner is best for lyra2z... anyone know...? I tried 4way and AES-avx2. they seem to get the same KH and KH/s.
For aes-sse42, it seems i get lower.
From personal test, it seems like aes-avx2 is way to go. can anyone confirm? also, what would be safe CPU temperature for CPU mining?
AVX2 is the way to go, 4 way does nothing for lyra2. You can also try reducing the number of threads, it might improve cache performance. The cooler the CPU the better. Make sure the fan is at 100%, it's cheaper to replace a worn out fan than a burned out CPU. I see. Do I have to modify my bat file to run less threads? sorry this is my first time using cpuminer so wanted to see what I need to try to see if I can get the maximum performance. The option needs to be on the cpuminer command line or in the config file if you use one. If you run cpuminer from a bat file then yes, you need to add the option to the comand line in the bat file. |
|
|
|
|
Show Posts
