Done:

https://bitcointalk.org/index.php?topic=1326803.msg13542056#msg13542056

This is the home of cpuminer-opt, the optimized CPU miner.

This thread is locked. all discussion of cpuminer-opt should now take place
in the new thread started by my alter ego JayDDee.

https://bitcointalk.org/index.php?topic=5226770.msg53865575#msg53865575

----------------------------------------------------------

Supporting over 90 agorithms with many optimized for CPUs with the latest technologies:

AMD Ryzen: AVX2, SHA
Intel Skylake X: AVX512
Intel Icelake: AVX512, SHA, and VAES  (source code only at this time)

Older CPUs with SSE2 are also supported.

Unique features

Block and share time estimates.

True hashrate reporting with comparison with actual share submission rate.

Download

Source code and Windows binaries can be downloaded from git.

https://github.com/JayDDee/cpuminer-opt/releases

Download only from the official JayDDee github repository. There is malware masquerading
as cpuminer-opt even with fake ANN posts that look just like this one.

New in v3.12.3

Issue #238: Fixed skunk AVX2.

Issue #239: Faster AVX2 & AVX512 for skein +44%, skein2 +30%, plus marginal
increases for skunk, x16r, x16rv2, x16rt, x16rt-veil, x16s, x21s.

Faster anime VAES +57%, AVX512 +21%, AVX2 +3%.

Redesigned code reponsible for issue #236.


New in v3.12.2

Fixed xevan, skein, skein2 AVX2, #238.

Reversed polarity of AVX2 vector bit test utilities, and all users, to be
logically and semantically correct. Follow up to issue #236.

New in v3.12.1

Fixed anime AVX2 low difficulty shares, git issue #236.

Periodic summary now reports lost hash rate due to rejected and stale shares,
displayed only when non-zero.

New in v3.12.0.1

Fixed hodl rejects, git issue #237.

Fixed debug code added in v3.12.0 to work with AVX2 to be enabled only
after low difficulty share have been seen to avoid unnecessarily excessive
log outout.

Added more digits of precision to diff in log output to help diagnose
low difficulty shares.

New in v3.12.0

Faster phi2 AVX2 +62%, AVX512 +150% on Intel CPUs. AMD Ryzen AVX2 is
YMMV due to its inferiour AVX2 implementation.

Fixed Hodl stats, rejects are still an issue since v3.9.5, git issue #237.

API can now be enabled with "-b port" or "--api-bind port".
It will use the default address 127.0.0.1.

Editorial: Short form options should only be used on the command line to save
typing. Configuration files and scripts should always use the long form
"--api-bind addr:port" without relying on any defaults. This is a general
recommendation that applies to all options for any application.

Removed obsolete cryptonight, all variants, and supporting code for more
size reduction and faster compiling.

Tweaked the timing of the CPU temperature and frequency log (Linux only).

Added some debug code to collect more info about low difficulty rejects,
git issue #236.

Requirements:

1. A x86_64 architecture CPU with a minimum of SSE2 support. This includes Intel Core2 and newer and AMD equivalents.
In order to take advantage of AES_NI optimizations a CPU with AES_NI is required. This includes Intel Westmere and newer
and AMD equivalents. Further optimizations are available on some algos for CPUs with AVX and AVX2, Sandybridge and
Haswell respectively.

32 bit CPUs and ARM CPUs are not supported.

2. 64 bit Linux OS. Ubuntu and Fedora based distributions, including Mint and Centos are known to work and have all
dependencies in their repositories. Others may work but may require more effort.

64 bit Windows OS is supported using the pre-compiled binaries package or may be compiled with mingw_w64 and msys.

Apple and Android operating systems are not supported. FreeBSD YMMV.

Older CPUs, other architectures and operating systems may be supported by TPruvot's cpuminer-multi.

Security warning

Miner programs are often flagged as malware by antivirus programs. This is usually a false positive, they are
flagged simply because they are cryptocurrency miners. However, some malware has been spread using the
cover that miners are known to be subject to false positives. Always be on alert. The source code of cpuminer-opt
is open for anyone to inspect. If you don't trust the software don't download it.

Some cryptographic code has been taken from trusted sources but has been modified for speed at the expense of
accepted security practices. This code should not be imported into applications where secure cryptography is
required.

Errata:

Old algorithms that are rarely used anymore will not get the latest optimizations.

Cryptonight and variants are no longer supported.

Neoscrypt crashes on Windows, use legacy version.

Hodl requires a CPU with AES, use legacy version if without.

Benchmark testing does not work on several algos.

Donations

cpuminer-opt has no fees of any kind but donations are accepted.

BTC: 12tdvfF7KmAsihBXQXynT6E6th2c2pByTT

First release will only support CPUs with AES-NI. I intend to put in a capabilities check to let users know if their
CPU won't cut it. Eventually, if I stay motivated,  there will be three architectures supported AES_NI, SSE2,
and then everything else.

Here are some numbers on a i7-4790k, 4 cores, 8 threads, stock_clock, compared with the closest competition:
 
                          quark        qubit        x11         x13        x15   
cpuminer-jdd      1080kh     1045kh   707kh     320kh    280kh   
multi                                     395kh                   168        160
minerd                                                680
cp3u                    905

Neoscrypt is unchaged from cpuminer-multi. Most of the other algos are unstable.
I don't expect the numbers to change before release. I still need to try compiling on
Windows (yuk vcxproj) but if I have problems I won't hold up the test release.
I might have something for you by the weekend.

This is fun!

Fix the source code, LOL. If I a binary only miner consistenly registers a lower hash rate I'd suspect some of the
hash was being diverted, not that any of the green team would do that.
The locally reported hash rate could also be fudged but monitoring network activity would coinfirm one way
or another.

I'll keep that in mind. Things are progressing well. I'm not sure which CPU miners you are familiar with but I'm working
with darkcoin cpu miner 1.3 (elmad), cp3u (palmd) & cpuminer-multi (tpruvot )and taking the best from each. I've even
managed to improve on the current best in each algo (all X, quark, qubit). The other algos are either unstable or too
difficult for the moment. If you know of any other interesting CPU miners point them out to me. I'll take a look and see
if I can make use of them. Also if you (or anyone else) knows of other contributors to these programs not yet credited
let me know.

It has been quite a challenge. I've never done any c++ development before and no development at all for 10 years. I had a
hell of a time messing with the Makefile sources and #includes. I'm trying to tweak each algo some more. Some work better
inlined, others not, simple stuff for now. Everything focussed on AES for now.

By my own ad-hoc versioning I'm at beta2 (that goal reached since the first post). Maybe one more beta before I'm comfortable
sharing it privately. Then I'll start work on support for CPUs limited to sse2, then...

Should this be moved to a new thread?

Sorry, no follow.

Let's do some math. I don't know the performance of a 750 i'll use a 750ti in my example.
A 750ti produces about 2100 kh. X15 is currently paying around 0.3 BTC/GH/day.

A 750ti will produce: (0.3 * 2100) / 1000000 = 0.00063 BTC/day

If you get a 10% increase for .01 BTC your ROI will be:

0.01 / (0.1 * 0.00063 ) = 158.7 days

This is for a 750ti, a 750 will be a little longer, and without factoring power cost. That's a long time to
invest in an algo that could fade back into the shadows at any time.

Now if you had 100 or 1000 cards, as SP suggested, the ROI would drop to a 1.57 days, and, 3.8 hours
respectively. 0.01 BTC is pretty cheap when you can ROI that fast.

Putting a minimum donation amount to get a private optimized miner is unfair to small miners
while allowing farms to get it dirt cheap.

Coincidentally I'm working on a faster CPU miner for X algos that nearly doubles the X15 performance
but noone wants to pay for a CPU miner. I'm not doing it to make money, I'm doing it because it's fun
and I'm learning something along the way. And if I ever get it finished I'll probably give it away.

That's an interesting approach to slow the arms race, choose an algo developpers don't like to work on.

A new opportunity may have come. Is anyone working on an argon2 miner? It just showed up on hashpower
at the top of their profitability.

If I could code cuda I'd be all over it.

Sure there is, well it can be estimated based on expected mining revenue.

scryptr paid for 500 kh of quark. His ROI is how long it would take to recover the cost of the miner with 500 kh
after subtracting the cost of electricity to power that much hash. The falacy is that all the hash produced by the
miner is going toward ROI.

The problem is optimization is an incremental process that produces diminishing returns and requires a lot more
effort to produce small improvements. These kinds of improvements only appeal to large operations that with scale
can ROI in a much shorter time. It is complicated by the migration to cuda 7.5 where a lot of effort is needed just to
recover lost hash compared to 6.5, not something most people are prepared to pay for.

I expect donations to be pretty thin until something new happens. It just the nature of the business.

I have a couple of questions.

While mining some algos I get very few accepts and in some case none for over 30 minutes. Also the hash
rate reported at the pool is very low or zero, much lower than the miner's display.

I presume that no blocks are being found because there are no validation errors and no rejects. Is this a
difficulty issue and is there a way around it?

I also have a question about lyra2v2. I was looking forward to it due to the exceptional performance of cpuminer-multi
on lyra2 (v1). IIRC you were stuck because one of the sub-algos was non standard. Is there any progress or have you
dropped it for other projects?

I'm also working on something you might be interested in, more on that at a later time.

Edit: I sorted out the low hash rate issue, the problem was at my end.

I'm guessing that message is triggered when a coin balance gets too low regardless of the reason.
They don't mine BTC so it must be the automatic exchanges that are falling behind.

Re ffpool, they just posted a message that it was shutting down and that was it. Curiously the web
site is still up.

Darkcoin CPU Miner 1.3 is the fasted x11 miner I've found, but I agree cpu mining x11 is a waste.

Again I question your math, You can only mine one algo at a time so a miner only gets the increase of one algo.
Also most miners are only interested in a couple of algos and don't care about improvements in the others.

If you increase an algo by 10% you should expect only 10% of what the design of the original algo is worth.

I'm not so sure about your math. Only the delivered increase in performance counts toward ROI, the existing performance
has already been figuratively paid for.

You deliver marginal increases in performance. These improvements are simply not worth it for smalll miners, it requires
scaling before the increases become substantial. If you released less often and with bigger gains it would look better
from a sales perspective

To appeal to small miners you need a market edge like being the first to market with a new algo, as you did with spread
and DJM did with neoscrypt, or the early days of Maxwell optimization when larger gains were realized. This is the kind of work that
will motivate small miners to contribute. I think DJM's approach is just that, hold back until market conditions improve.
I don't think the razor blade marketing approach will work in this case.

Razor blade marketing, for those who don't know, is a technique where the big money is made by frequent replacement
of wearable items. Rasors are cheap, blades are expensive. Printer ink use the same technique, printers are cheap, ink
is expensive.

I'm not saying I agree or disagree with anyone's chosen approach, just saying if it was me and I was trying to make
money writing mining software.

Three points that all can deliver a linux binary while keeping the source private.

1. Static builds can be compiled, as previously mentioned. Depending on the amount of extra work this may be viable.

2. The dev could use a VM with your OS to build a standard dynamically linked version, extra work for the dev but pure linux.

3. Run the windows build using Wine with no performance loss, a bastard but easiest to implement.

In Canada the price of a 980ti is about 5x 750ti, needs to drop some.

Did you look back in the thread for the old 6th card problem discussion?

@crysx

I have a vague memory  of a similar issue being discussed a while back in this thread, specifically the part about
adding a 6th card. I also recall some mention of pagefile/swap space. Might be worth a look.

Works on my 750ti with driver 352.63.

It works on Linux. You might also want to set persistence mode: nvidia-smi -pm 1 -i 0