blas_magma.cpp does not compile

[maddyscientist]

On the develop branch after the recent merge of the mixed-precision eigCG support (62868e52cd701ba86b9c27684a0ddbc77fcede16). There are two issues that I can see:

• This is a C++ code yet it uses C99 complex numbers. These are not recognised by strict C++ compilers (e.g., clang/LLVM) and these fail to compile.
• There are many parts of the code that are not protected from compilation if MAGMA_LIB is not defined.

There are dozens of errors, here's a sample :)

blas_magma.cpp:523:46: error: use of undeclared identifier 'magmaDoubleComplex'
    cudaHostRegister(pQR, ldqr*nrefls*sizeof(magmaDoubleComplex), cudaHostRegisterMapped);
                                             ^
blas_magma.cpp:1047:17: error: use of undeclared identifier 'creal'
  lwork = (int)(creal(qwork));

@alexstrel Please fix this as soon as possible inserting the appropriate macros to prevent compilation and replace the C99 complex numbers with C++ complex numbers (use quda::complex or std::complex).

[maddyscientist]

Quoting from here: http://stackoverflow.com/questions/10540228/c-complex-numbers-in-c

A C++ compiler could choose to support the _Complex keyword as an extension (and a few do), but that isn't portable. If you want to have a portable C++ solution, you need to use the C++ std::complex templates, unfortunately.

The good news is that C++ std::complex numbers are guaranteed to be compatible with C99 complex numbers (in the sense that a pointer to one can always be converted to a pointer to the other, and the right thing will happen), which means that if you need to interoperate with a C library that expects C99 complex values, you shouldn't have any trouble.

So you should be able to work with C++ complex numbers on the QUDA side of things, and it should work ok.

[alexstrel]

ok, thanks.

[alexstrel]

Compilation errors fixed, I cleaned magma interface because it contained stuff required by GMRESDR solver only (I missed this point completly). There are now runtime errors, namely ERROR: CopyGenericColorSpinor has not been built for Nspin=4 fields this i did not observe with yesterday's feature branch.. Comments?

[maddyscientist]

Do you have a branch I can test this on to work out the problem?

[maddyscientist]

Ok, I see the problem. Can you swap this on line 413 of copy_color_spinor.cu

#if defined(GPU_STAGGERED_DIRAC)

for this

#if defined(GPU_WILSON_DIRAC) || defined(GPU_DOMAIN_WALL_DIRAC)

[alexstrel]

sure , this was the problem, thanks!

[alexstrel]

hmm, incremental stage was successful, but deflted CG collapsed, still need some time to resolve this issue before the pull request, CG: 0 iterations, <r,r> = 4.611374e+06, |r|/|b| = 9.999999e-01, heavy-quark residual = 1.730952e+00 CG: 1 iterations, <r,r> = 1.840267e+06, |r|/|b| = 6.317208e-01, heavy-quark residual = 6.319742e-01 CG: 2 iterations, <r,r> = 7.712731e+05, |r|/|b| = 4.089676e-01, heavy-quark residual = 6.319742e-01 CG: 3 iterations, <r,r> = 3.902850e+05, |r|/|b| = 2.909214e-01, heavy-quark residual = 6.319742e-01 CG: 4 iterations, <r,r> = 1.940245e+05, |r|/|b| = 2.051224e-01, heavy-quark residual = 6.319742e-01 CG: 5 iterations, <r,r> = 9.574212e+04, |r|/|b| = 1.440908e-01, heavy-quark residual = 6.319742e-01 CG: 6 iterations, <r,r> = 4.828334e+04, |r|/|b| = 1.023254e-01, heavy-quark residual = 6.319742e-01 CG: 7 iterations, <r,r> = 2.421310e+04, |r|/|b| = 7.246195e-02, heavy-quark residual = 6.319742e-01 CG: 8 iterations, <r,r> = 1.212623e+04, |r|/|b| = 5.127997e-02, heavy-quark residual = 6.319742e-01 CG: 9 iterations, <r,r> = 6.092978e+03, |r|/|b| = 3.634960e-02, heavy-quark residual = 6.319742e-01 CG: 10 iterations, <r,r> = 3.063799e+03, |r|/|b| = 2.577597e-02, heavy-quark residual = 6.319742e-01 CG: 11 iterations, <r,r> = 1.539253e+03, |r|/|b| = 1.827006e-02, heavy-quark residual = 9.515757e-03 CG: 12 iterations, <r,r> = 7.756649e+02, |r|/|b| = 1.296946e-02, heavy-quark residual = 9.515757e-03 WARNING: CG: new reliable residual norm 4.472954e+03 is greater than previous reliable residual norm 2.147411e+03 (total #inc 1) WARNING: CG: Restarting without reliable updates for heavy-quark residual CG: 13 iterations, <r,r> = 2.000731e+07, |r|/|b| = 2.082952e+00, heavy-quark residual = 1.041821e+00 CG: 14 iterations, <r,r> = 2.973748e+07, |r|/|b| = 2.539434e+00, heavy-quark residual = 1.041821e+00 CG: 15 iterations, <r,r> = 1.263694e+07, |r|/|b| = 1.655411e+00, heavy-quark residual = 1.041821e+00 CG: 16 iterations, <r,r> = 6.113315e+06, |r|/|b| = 1.151392e+00, heavy-quark residual = 1.041821e+00 CG: 17 iterations, <r,r> = 2.975311e+06, |r|/|b| = 8.032504e-01, heavy-quark residual = 1.041821e+00 CG: 18 iterations, <r,r> = 1.513198e+06, |r|/|b| = 5.728391e-01, heavy-quark residual = 1.041821e+00 CG: 19 iterations, <r,r> = 7.568782e+05, |r|/|b| = 4.051332e-01, heavy-quark residual = 1.041821e+00 CG: 20 iterations, <r,r> = 3.787794e+05, |r|/|b| = 2.866012e-01, heavy-quark residual = 1.041821e+00 CG: 21 iterations, <r,r> = 1.921011e+05, |r|/|b| = 2.041031e-01, heavy-quark residual = 2.492663e-02 CG: 22 iterations, <r,r> = 9.647581e+04, |r|/|b| = 1.446419e-01, heavy-quark residual = 2.492663e-02 CG: 23 iterations, <r,r> = 4.843184e+04, |r|/|b| = 1.024826e-01, heavy-quark residual = 2.492663e-02 CG: 24 iterations, <r,r> = 2.439292e+04, |r|/|b| = 7.273051e-02, heavy-quark residual = 2.492663e-02 CG: 25 iterations, <r,r> = 1.228554e+04, |r|/|b| = 5.161571e-02, heavy-quark residual = 2.492663e-02 CG: 26 iterations, <r,r> = 6.184767e+03, |r|/|b| = 3.662237e-02, heavy-quark residual = 2.492663e-02 CG: 27 iterations, <r,r> = 3.118669e+03, |r|/|b| = 2.600576e-02, heavy-quark residual = 2.492663e-02 CG: 28 iterations, <r,r> = 1.573619e+03, |r|/|b| = 1.847288e-02, heavy-quark residual = 2.492663e-02 CG: 29 iterations, <r,r> = 7.930396e+02, |r|/|b| = 1.311391e-02, heavy-quark residual = 2.492663e-02 CG: 30 iterations, <r,r> = 4.000733e+02, |r|/|b| = 9.314394e-03, heavy-quark residual = 2.492663e-02 CG: 31 iterations, <r,r> = 2.020004e+02, |r|/|b| = 6.618520e-03, heavy-quark residual = 7.989245e-04 WARNING: CG: new reliable residual norm 2.784615e+04 is greater than previous reliable residual norm 4.472954e+03 (total #inc 2) WARNING: CG: Restarting without reliable updates for heavy-quark residual WARNING: CG: new reliable HQ residual norm 1.471889e+00 is greater than previous reliable residual norm 1.041821e+00 CG: 32 iterations, <r,r> = 7.754078e+08, |r|/|b| = 1.296731e+01, heavy-quark residual = 1.471889e+00 CG: 33 iterations, <r,r> = 1.657868e+09, |r|/|b| = 1.896093e+01, heavy-quark residual = 1.471889e+00 CG: 34 iterations, <r,r> = 6.752485e+08, |r|/|b| = 1.210087e+01, heavy-quark residual = 1.471889e+00

[mathiaswagner]

I would prefer to have a pull request now which fixes the compilation. Having a develop branch that does not compile is highly undesirable. Fixing that is the highest priority.

We can keep the branch alive after that first pull and you can fix the other issues. Once these are fixed we can do a second pull.

[alexstrel]

since the observed run time problem is unrelated to the original compilation issue, I suggest to close this item.

[mathiaswagner]

I would like to keep that open until we confirmed that compilation works now and #336 is merged in develop.

[nmrcardoso]

tested here and blas_magma.cpp compiles without errors (using option BUILD_MAGMA = no).

On Fri, Jul 31, 2015 at 9:47 AM, Mathias Wagner notifications@github.com wrote:

Reopened #335 https://github.com/lattice/quda/issues/335.

— Reply to this email directly or view it on GitHub https://github.com/lattice/quda/issues/335#event-370574737.

[mathiaswagner]

Thanks, @nmrcardoso. I just would like to wait if it is also is fine with @mikeaclark and her clang compilation.

[maddyscientist]

Confirmed, the bug is fixed.

[alexstrel]

So, we can now close the issue?

[mathiaswagner]

closed with #336