Dev cloud back end JIT failure

[paboyle]

Much more painless than CUDA support was so far. C++ friendliness is good,
though hitting some not trivially copy constructible complaints for Lambda captures.

First cut compile of entire application goes (almost) through, which is pretty awesome.
(Compile is rather slow, but mustn't grumble)

git clone https://www.github.com/paboyle/Grid
cd Grid
git checkout sycl
./bootstrap.sh
mkdir build
cd build
../configure --enable-simd=GEN --enable-gen-simd-width=64 --enable-precision=single --enable-sycl CXX=dpcpp
make -j 12 -C Grid
make -j 12 -C benchmarks
cd benchmarks
./Benchmark_su3_gpu

But running gives

Grid : Message : ================================================ 
Grid : Message : MPI is initialised and logging filters activated 
Grid : Message : ================================================ 
Grid : Message : Requested 1073741824 byte stencil comms buffers 
Grid : Message : 0.523132 s : Grid is setup to use 1 threads
Grid : Message : 0.523143 s : ============================================================
Grid : Message : 0.523156 s : = Benchmarking SU3xSU3  ext/ins z = x*y
============================================================
Grid : Message : 0.523179 s :   L  		bytes			GB/s		 GFlop/s
Grid : Message : 0.523190 s : -----------------------------------------------------
terminate called after throwing an instance of 'cl::sycl::compile_program_error'
  what():  The program was built for 1 devices
Build program log for 'Intel(R) Gen9 HD Graphics NEO':
error: undefined reference to `__mulsc3()'
error: backend compiler failed build.
 -17 (CL_LINK_PROGRAM_FAILURE)
Aborted

Advice welcome?

[bader]

@paboyle, did you tried to report this to https://software.intel.com/en-us/forums/intel-oneapi-forums? We have more engineers there helping with the questions related to DevCould and oneAPI toolchain.

What toolchain/compiler version do you use?

[paboyle]

I was using beta06 on the devcloud; I'm registering with the forum and will report there

[bader]

I just realized that we have a separate forum for DevCould issues: https://software.intel.com/en-us/forums/intel-devcloud, but your seems to be related to toolchain.
It looks like the app is using complex number in device code and I'm surprised it doesn't work as it should be supported in Beta06.
One reason it might not work is that app is using complex numbers with double precision FP, but fp64 is not supported by the GPU HW.

[paboyle]

../configure --enable-simd=GEN --enable-gen-simd-width=64 --enable-precision=single --enable-accelerator=sycl CXX=dpcpp

Updated the GitHub branch and configure option now as above.
This builds CUDA, HIP and SYCL as options, but warts continue to exist for both HIP and SYCL.

The complex arithmetic JIT bug persists. Beta06 on devcloud. Wasn't able to register for the forum - the confirmation email never came.

[bd4]

@paboyle I was able to get a simple std::complex<double> example working by linking to libsycl-cmath-fp64.o and libsycl-complex-fp64.o in /opt/intel/inteloneapi/compiler/latest/linux/lib, see https://github.com/intel/llvm/blob/sycl/sycl/test/devicelib/std_complex_math_fp64_test.cpp

[paboyle]

Wow Bryce - thanks a million. That is incredibly helpful. Will have a go.

[bd4]

Let me know if it works for you. This is not a good "standard" SYCL solution, but FWIW hipsycl seems to work out of the box with std::complex. ComputeCpp 2.0.0 warns that __muldc3 is not found at compile time, then fails in a similar way as dpcpp at runtime. I started a forum conversation at their forum here:
https://support.codeplay.com/t/std-complex-multiplication/355

Hacks are fine short term, but I'd also like to understand the direction of the SYCL standard on this. Seems to me like it's a gap in the standard - if the device compiler / runtime happens to support it, everything works great, but AFAIK it's not required or even clearly defined as an extension.

[paboyle]

Yeah - complex and GPU is a mess right now.
HIP and CUDA are sticking in the STL partially replacement namespace "thrust", but I had hoped the more standard C++ environment of SYCL could make it even better than such an ugly hack.

[bd4]

HIP seems to work out of the box with std::complex, at least for basic arithmetic, I haven't tried using transcendental functions yet.

[paboyle]

Hi Bryce

-- very interesting - single still fails, but double precision now passes simple arithmetic tests with complex arithmetic with these files linked.

Haven't check all transcendental complex maths yet, but double complex exponential works.

Of course all the thrust discussion is about getting that functionality. +,-,*,/ is easy - it's the library functionality that is the real benefit from reuse...

Assuming they eventually support all the STL mandated transcendental it puts SYCL in a much more civilised state than CUDA.

[paboyle]

I assume I can get single with the non-fp64 versions :

libsycl-cmath-fp64.o
libsycl-cmath.o
libsycl-complex-fp64.o
libsycl-complex.o

[bd4]

Yes I think so, I haven't tested single precision yet. Including both in the link line also seems to work, which is potentially convenient for a library install that might require both. For cmake, I added the .o files to target_sources on the gtensor INTERFACE library, to get the object files linked on executable that link gtensor.

[jeremyong]

I am trying to resolve this right now when compiling against the devcloud as well (targeting Intel Gen 9 integrated GPU)

Unfortunately, I get a number of math-related undefined references even when I link against the compiled objects mentioned in this thread.

terminate called after throwing an instance of 'cl::sycl::compile_program_error'
  what():  The program was built for 1 devices
Build program log for 'Intel(R) Gen9':

error: undefined reference to `__devicelib_scalbnf()'
undefined reference to `__devicelib_logf()'
undefined reference to `__devicelib_expf()'
undefined reference to `__devicelib_frexpf()'
undefined reference to `__devicelib_ldexpf()'
undefined reference to `__devicelib_log10f()'
undefined reference to `__devicelib_modff()'
undefined reference to `__devicelib_exp2f()'
undefined reference to `__devicelib_expm1f()'
undefined reference to `__devicelib_ilogbf()'
undefined reference to `__devicelib_log1pf()'
undefined reference to `__devicelib_log2f()'
undefined reference to `__devicelib_logbf()'
undefined reference to `__devicelib_sqrtf()'
undefined reference to `__devicelib_cbrtf()'
undefined reference to `__devicelib_hypotf()'
undefined reference to `__devicelib_erff()'
undefined reference to `__devicelib_erfcf()'
undefined reference to `__devicelib_tgammaf()'
undefined reference to `__devicelib_lgammaf()'
undefined reference to `__devicelib_fmodf()'
undefined reference to `__devicelib_remainderf()'
undefined reference to `__devicelib_remquof()'
undefined reference to `__devicelib_nextafterf()'
undefined reference to `__devicelib_fdimf()'
undefined reference to `__devicelib_fmaf()'
undefined reference to `__devicelib_sinf()'
undefined reference to `__devicelib_cosf()'
undefined reference to `__devicelib_tanf()'
undefined reference to `__devicelib_powf()'
undefined reference to `__devicelib_acosf()'
undefined reference to `__devicelib_asinf()'
undefined reference to `__devicelib_atanf()'
undefined reference to `__devicelib_atan2f()'
undefined reference to `__devicelib_coshf()'
undefined reference to `__devicelib_sinhf()'
undefined reference to `__devicelib_tanhf()'
undefined reference to `__devicelib_acoshf()'
undefined reference to `__devicelib_asinhf()'
undefined reference to `__devicelib_atanhf()'
undefined reference to `__devicelib_log()'
undefined reference to `__devicelib_exp()'
undefined reference to `__devicelib_frexp()'
undefined reference to `__devicelib_ldexp()'
undefined reference to `__devicelib_log10()'
undefined reference to `__devicelib_modf()'
undefined reference to `__devicelib_exp2()'
undefined reference to `__devicelib_expm1()'
undefined reference to `__devicelib_ilogb()'
undefined reference to `__devicelib_log1p()'
undefined reference to `__devicelib_log2()'
undefined reference to `__devicelib_logb()'
undefined reference to `__devicelib_sqrt()'
undefined reference to `__devicelib_cbrt()'
undefined reference to `__devicelib_hypot()'
undefined reference to `__devicelib_erf()'
undefined reference to `__devicelib_erfc()'
undefined reference to `__devicelib_tgamma()'
undefined reference to `__devicelib_lgamma()'
undefined reference to `__devicelib_fmod()'
undefined reference to `__devicelib_remainder()'
undefined reference to `__devicelib_remquo()'
undefined reference to `__devicelib_nextafter()'
undefined reference to `__devicelib_fdim()'
undefined reference to `__devicelib_fma()'
undefined reference to `__devicelib_sin()'
undefined reference to `__devicelib_cos()'
undefined reference to `__devicelib_tan()'
undefined reference to `__devicelib_pow()'
undefined reference to `__devicelib_acos()'
undefined reference to `__devicelib_asin()'
undefined reference to `__devicelib_atan()'
undefined reference to `__devicelib_atan2()'
undefined reference to `__devicelib_cosh()'
undefined reference to `__devicelib_sinh()'
undefined reference to `__devicelib_tanh()'
undefined reference to `__devicelib_acosh()'
undefined reference to `__devicelib_asinh()'
undefined reference to `__devicelib_atanh()'

error: backend compiler failed build.
 -11 (CL_BUILD_PROGRAM_FAILURE)

[bader]

@paboyle, this issue should be addressed by #2400. Could you confirm, please?

[paboyle]

Closing as this works on devcloud under beta10 with -fsycl-device-lib=all