gnmf crashes with --vec

[philipportner]

Running the scripts/algorithms/gnmf.daph script with --vec crashes.

Command to reproduce

bin/daphne --vec scripts/algorithms/gnmf.daph rank=2 n=100 e=500 W=\"outW.csv\" H=\"outH.csv\"

Stack trace

#0  __GI_raise (sig=sig@entry=6) at ../sysdeps/unix/sysv/linux/raise.c:50
#1  0x00007ffff75cb859 in __GI_abort () at abort.c:79
#2  0x00007ffff79b3ee6 in ?? () from /lib/x86_64-linux-gnu/libstdc++.so.6
#3  0x00007ffff79c5f8c in ?? () from /lib/x86_64-linux-gnu/libstdc++.so.6
#4  0x00007ffff79c5ff7 in std::terminate() () from /lib/x86_64-linux-gnu/libstdc++.so.6
#5  0x00007ffff79c6258 in __cxa_throw () from /lib/x86_64-linux-gnu/libstdc++.so.6
#6  0x00007fffcd4c9e02 in CompiledPipelineTask<DenseMatrix<double> >::accumulateOutputs (this=this@entry=0x555559d0c6b0, localResults=std::vector of length 2, capacity 2 = {...},
    localAddRes=std::vector of length 2, capacity 2 = {...}, rowStart=rowStart@entry=0, rowEnd=rowEnd@entry=1) at /usr/include/c++/11/bits/allocator.h:174
#7  0x00007fffcd4cbb4e in CompiledPipelineTask<DenseMatrix<double> >::execute (this=<optimized out>, fid=<optimized out>, batchSize=2595) at /home/philipportner/daphne/src/runtime/local/vectorized/Tasks.cpp:35
#8  0x00007fffcd4ae320 in WorkerCPU::run (this=0x55555a223e80) at /home/philipportner/daphne/src/runtime/local/vectorized/WorkerCPU.h:84
#9  0x00007ffff79f5793 in ?? () from /lib/x86_64-linux-gnu/libstdc++.so.6
#10 0x00007ffff7f99609 in start_thread (arg=<optimized out>) at pthread_create.c:477
#11 0x00007ffff76c8353 in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95

[sweetpellegrino]

Hey Philipp,

I think I can provide some direction for this issue.
I strongly believe that the current implementation of the vectorization pass produces broken pipelines.

Here is one pipeline of gnmf.daph

%21:2 = "daphne.vectorizedPipeline"(%20, %18, %5, %8, %19, %16, %18, %4, %0, %0, %2, %0) ({
    ^bb0(%arg0: !daphne.Matrix<?x2xf64:sp[1.000000e+00]>, %arg1: !daphne.Matrix<2x100xf64:sp[1.000000e+00]>, %arg2: i1, %arg3: f64, %arg4: !daphne.Matrix<?x200xf64:sp[1.000000e+00]>, %arg5: !daphne.Matrix<200x100xf64>, %arg6: !daphne.Matrix<?x100xf64:sp[1.000000e+00]>, %arg7: i1):
      %24 = "daphne.matMul"(%arg0, %arg1, %arg2, %arg2) : (!daphne.Matrix<?x2xf64:sp[1.000000e+00]>, !daphne.Matrix<2x100xf64:sp[1.000000e+00]>, i1, i1) -> !daphne.Matrix<?x?xf64:sp[1.000000e+00]>
      %25 = "daphne.ewAdd"(%24, %arg3) : (!daphne.Matrix<?x?xf64:sp[1.000000e+00]>, f64) -> !daphne.Matrix<?x?xf64>
      %26 = "daphne.matMul"(%arg4, %arg5, %arg2, %arg2) : (!daphne.Matrix<?x200xf64:sp[1.000000e+00]>, !daphne.Matrix<200x100xf64>, i1, i1) -> !daphne.Matrix<?x?xf64>
      %27 = "daphne.ewDiv"(%26, %25) : (!daphne.Matrix<?x?xf64>, !daphne.Matrix<?x?xf64>) -> !daphne.Matrix<?x?xf64>
      %28 = "daphne.ewMul"(%arg6, %27) : (!daphne.Matrix<?x100xf64:sp[1.000000e+00]>, !daphne.Matrix<?x?xf64>) -> !daphne.Matrix<?x?xf64:sp[1.000000e+00]>
      %29 = "daphne.matMul"(%28, %28, %arg2, %arg7) : (!daphne.Matrix<?x?xf64:sp[1.000000e+00]>, !daphne.Matrix<?x?xf64:sp[1.000000e+00]>, i1, i1) -> !daphne.Matrix<?x?xf64:sp[1.000000e+00]>
      "daphne.return"(%28, %29) : (!daphne.Matrix<?x?xf64:sp[1.000000e+00]>, !daphne.Matrix<?x?xf64:sp[1.000000e+00]>) -> ()
   } ,

We see that matMul (%29) is in one pipeline with ewMul (%28). However the current implementation for matMul only allows for a row split for lhs and broadcast for rhs. This does not work in case lhs and rhs are both %28.

The approach that we taken in the thesis will not produce such pipelines, but split it.