Cgminer Source

You can build and benchmark it. It's ready.

 

For a newbie like me: How to benchmark on windows? Can't find any executable file....

 

I built mine on cygwin, and it benchmarks pretty well.
Card 1: Radeon R9 200 Series, Pitcairn, 1280 shaders, 2048MB memory, GPU Clock 955, Memory Clock 1400
Card 2: Radeon R7 Graphics, Spectre, 512 shaders, 1024MB memory, GPU Clock 867, Memory Clock 1066

[2016-01-26 17:01:17] GPU0 | (5s):737.8M (avg):739.9Mh/s | A:0 R:0 HW:0 WU:0.0/m
[2016-01-26 17:01:17] GPU1 | (5s):258.2M (avg):257.1Mh/s | A:0 R:0 HW:0 WU:0.0/m

 

MinGW64 defines nanosleep() and sleep(), MinGW32 doesn't usually.

Download this version of adl https://drive.google.com/folderview?id=0ByitvRo_awTXb0VzR0hFb1FIRUE&usp=sharing

Copy them into ADL_SDK directory inside the cgminer-3.7 directory


Open compat.h in cgminer folder and find this section of the code, starting from line number 16

Code:

#ifndef HAVE_LIBWINPTHREAD
static inline int nanosleep(const struct timespec *req, struct timespec *rem)
{
    struct timeval tstart;
    DWORD msecs;

    cgtime(&tstart);
    msecs = (req->tv_sec * 1000) + ((999999 + req->tv_nsec) / 1000000);

    if (SleepEx(msecs, true) == WAIT_IO_COMPLETION) {
        if (rem) {
            struct timeval tdone, tnow, tleft;
            tdone.tv_sec = tstart.tv_sec + req->tv_sec;
            tdone.tv_usec = tstart.tv_usec + ((999 + req->tv_nsec) / 1000);
            if (tdone.tv_usec > 1000000) {
                tdone.tv_usec -= 1000000;
                ++tdone.tv_sec;
            }

            cgtime(&tnow);
            if (timercmp(&tnow, &tdone, >))
                return 0;
            timersub(&tdone, &tnow, &tleft);

            rem->tv_sec = tleft.tv_sec;
            rem->tv_nsec = tleft.tv_usec * 1000;
        }
        errno = EINTR;
        return -1;
    }
    return 0;
}
#endif

static inline int sleep(unsigned int secs)
{
    struct timespec req, rem;
    req.tv_sec = secs;
    req.tv_nsec = 0;
    if (!nanosleep(&req, &rem))
        return 0;
    return rem.tv_sec + (rem.tv_nsec ? 1 : 0);
}

Replace with this code

Code:

#if !(__MINGW32__)

#ifndef HAVE_LIBWINPTHREAD
static inline int nanosleep(const struct timespec *req, struct timespec *rem)
{
    struct timeval tstart;
    DWORD msecs;

    cgtime(&tstart);
    msecs = (req->tv_sec * 1000) + ((999999 + req->tv_nsec) / 1000000);

    if (SleepEx(msecs, true) == WAIT_IO_COMPLETION) {
        if (rem) {
            struct timeval tdone, tnow, tleft;
            tdone.tv_sec = tstart.tv_sec + req->tv_sec;
            tdone.tv_usec = tstart.tv_usec + ((999 + req->tv_nsec) / 1000);
            if (tdone.tv_usec > 1000000) {
                tdone.tv_usec -= 1000000;
                ++tdone.tv_sec;
            }

            cgtime(&tnow);
            if (timercmp(&tnow, &tdone, >))
                return 0;
            timersub(&tdone, &tnow, &tleft);

            rem->tv_sec = tleft.tv_sec;
            rem->tv_nsec = tleft.tv_usec * 1000;
        }
        errno = EINTR;
        return -1;
    }
    return 0;
}
#endif

static inline int sleep(unsigned int secs)
{
    struct timespec req, rem;
    req.tv_sec = secs;
    req.tv_nsec = 0;
    if (!nanosleep(&req, &rem))
        return 0;
    return rem.tv_sec + (rem.tv_nsec ? 1 : 0);
}

#endif

Haven't tried compiling it myself as i've been a little busy on some job assignment.
Please try this solution and let me know if it works.
Otherwise i'll provide you with another solution which involves editing unistd.h

Cheers

 

I have Compiled successfully with MingW32 on windows without any problems

AMD 6900 = 1.5Gh/s

 

Don't worry i've slightly modified unistd.h in "MinGW\include" and the link is provided below.
Rename "MinGW\include\unistd.h" to unistd.h.old
Download unistd.h below into your MinGW\include folder and try again.
https://drive.google.com/file/d/0ByitvRo_awTXQXJ6WFNZcWJTTGM/view?usp=sharing

Let me know if u encounter any other errors.

 

Points docked for having configure called in autogen.sh, and EVEN THEN it doesn't work right, because there's no --enable-opencl used when configure is called from autogen.sh - most people like to call configure themselves. The autogen.sh script isn't executable when cloned, either.

My real reason to bitch, however, is that the README doesn't say you must use -k blake256 to get it to mine Decred at all - those trying --benchmark are likely getting totally bogus values. It uses the poclbm SHA256d kernel by default.

Nitpicks: Could have cleaned out SHA256d code and kernels, plus did a few other cleanups since it doesn't seem to care much about putting it into a multi-algo miner - for a Decred-only miner, it could be WAY easier to use and lighter, but again, that's just a nitpick.

EDIT:

Just looked at the OpenCL - DEAR LORD, who wrote that? I mean, it's probably efficient, yes, but reading it makes you want to tear your eyes out. You do know macros exist for a reason, and so do loops... if you're worried about the branching with looping, just use "#pragma unroll" and use it so the code is readable. Also, the AMD compiler is probably too stupid to turn some of your rotates into vector swizzles... just saying.

EDIT2:

Looked at the END of the OpenCL - loose ends here. If - and yes, this is hella unlikely (depending on share difficulty) - two work items find a share at the same time, then it is very possible they both write to the same index in the same output array and/or increment the counter at the end of it at the same time, making the result undefined. For those that don't code, this means "probably total garbage." The OpenCL builtin function atomic_inc() is your friend.

EDIT3: I just found out you don't need -k blake256, you need --blake256. Documentation on this, again, would be REALLY helpful.