Can every index be sparse?

[Llauset]

I tried to generate in the online generation the simple example having every index sparse and it fails.

Am I misunderstanding something?

It seems that some index in the outcome tensor should be Dense, is it true?

[rohany]

What schedule did you try and use with this expression and format pair? If you tried to parallelize the i loop, then the scheduling primitives available on the website will not work (such a thing is supported through the use of Assemble in the C++ API).

[Llauset]

Thank you for your answer Rohany.
I am new using TACO so for sure I am doing something wrong.
I presented the example on my first post trying to understand what it is happening if I set all the dimensions of the tensors to be sparse.
I do not understand what you mean by "shedule" or "format pair". I show you here the code I would like to run. Please, tell me if you have any sugestion about why it is not working.

  #include "taco.h"
  #include <iostream>
  
  using namespace taco;
  using namespace std;
  
  int main(int argc, char* argv[])
  {
      Format b_and_t_format({ Sparse, Sparse, Sparse });
      Format A_format({ Sparse, Sparse, Sparse, Sparse, Sparse, Sparse, Sparse });
 
      Tensor<double> B({ 3, 3, 3 }, b_and_t_format);
      Tensor<double> T({ 3, 3, 3 }, b_and_t_format);
      Tensor<double> A({ 3, 3, 3, 3, 3, 3, 3 }, A_format);
  
      B.insert({ 0, 1, 0 }, 0.5);
      B.insert({ 1, 0, 0 }, 0.5);
      B.insert({ 2, 0, 1 }, 0.5);
      B.insert({ 0, 2, 1 }, 0.5);
  
      T.insert({ 0, 0, 0 }, 1.0);
      T.insert({ 2, 1, 0 }, 1.0 / sqrt(6));
      T.insert({ 2, 1, 0 }, -1.0 / sqrt(6));
      T.insert({ 2, 0, 1 }, -1.0 / sqrt(6));
      T.insert({ 0, 2, 1 }, 1.0 / sqrt(6));
      T.insert({ 1, 0, 2 }, 1.0 / sqrt(6));
      T.insert({ 0, 1, 2 }, -1.0 / sqrt(6));
  
      // Pack data as described by the formats
      B.pack();
      T.pack();
      // Form a tensor contraction expresion
      IndexVar a, b, c, d, e, f, g, h, i, j, k;
      A(a, b, c, d, e, f, g) = B(a, h, e) * B(k, j, f) * B(d, i, g) * T(h, j, i) * T(k, b, c);
      // Compile the expression
      A.compile();
      // Assemble A's indices and numerically compute the result
      A.assemble();
      A.compute();
  }

[rohany]

If you do not attempt to parallelize your code, then every index can indeed be sparse. What error do you get when running that program?

[Llauset]

Segmentation fault.
Running it with gdb I get this information about the error:
Program received signal SIGSEGV, Segmentation fault.
0x00002aaaaaf1ecbf in std::vector<taco::IndexVar, std::allocatortaco::IndexVar >::_M_erase(__gnu_cxx::__normal_iterator<taco::IndexVar*, std::vector<taco::IndexVar, std::allocatortaco::IndexVar > >) () from /tmpdir/sparse_tensors_3D/taco/build/lib/libtaco.so
Missing separate debuginfos, use: debuginfo-install glibc-2.17-326.el7_9.x86_64

Have you tried to run the code ? Are you suggesting that it has to be a problem with my environment ?
Thank you !

[rohany]

Can you get a backtrace with gdb? It will be more fruitful if you compile TACO in debug mode (-DCMAKE_BUILD_TYPE=Debug) and get a backtrace.

[Llauset]

Here it is:

terminate called after throwing an instance of 'taco::TacoException'
what(): Compiler bug at /tmpdir/sparse_tensors_3D/taco/src/index_notation/transformations.cpp:1678 in topologicallySort
Please report it to developers
Condition failed: freeVarPos != std::numeric_limits<size_t>::max()
Cycles in iteration graphs must be resolved, through transpose, before the expression is passed to the topological sorting routine.

Backtrace:
#0 0x00002aaaac15a387 in raise () from /lib64/libc.so.6
#1 0x00002aaaac15ba78 in abort () from /lib64/libc.so.6
#2 0x00002aaaab8c9677 in __gnu_cxx::__verbose_terminate_handler () at ../../.././libstdc++-v3/libsupc++/vterminate.cc:95
#3 0x00002aaaab8d5116 in __cxxabiv1::__terminate(void (*)()) () at ../../.././libstdc++-v3/libsupc++/eh_terminate.cc:47
#4 0x00002aaaab8d5161 in std::terminate () at ../../.././libstdc++-v3/libsupc++/eh_terminate.cc:57
#5 0x00002aaaab8d5395 in __cxxabiv1::__cxa_throw (obj=, tinfo=0x2aaaab7f19c0 ,
dest=0x2aaaab1b66e6 taco::TacoException::~TacoException()) at ../../.././libstdc++-v3/libsupc++/eh_throw.cc:95
#6 0x00002aaaab1b65bf in taco::ErrorReport::explodeWithException (this=0x7fffffffbe30) at /tmpdir/sparse_tensors_3D/taco/src/error.cpp:65
#7 0x00002aaaab1b60b7 in taco::ErrorReport::~ErrorReport (this=0x7fffffffbe30, __in_chrg=)
at /tmpdir/sparse_tensors_3D/taco/include/taco/error.h:60
#8 0x00002aaaab336338 in taco::topologicallySort (hardDeps=..., softDeps=..., originalOrder=...)
at /tmpdir/sparse_tensors_3D/taco/src/index_notation/transformations.cpp:1678
#9 0x00002aaaab337a43 in taco::reorderLoopsTopologically (stmt=...) at /tmpdir/sparse_tensors_3D/taco/src/index_notation/transformations.cpp:1799
#10 0x00002aaaab1cfed5 in taco::TensorBase::compile (this=0x7fffffffc500) at /tmpdir/sparse_tensors_3D/taco/src/tensor.cpp:638
#11 0x0000000000404d9a in main ()

[rohany]

So the first part of the problem is that TACO is getting confused by you using the same variable in multiple places in the expression (which is a bug).

Changing this

      Tensor<double> B({ 3, 3, 3 }, b_and_t_format);
      Tensor<double> T({ 3, 3, 3 }, b_and_t_format);
      Tensor<double> A({ 3, 3, 3, 3, 3, 3, 3 }, A_format);

to

      Tensor<double> B1({ 3, 3, 3 }, b_and_t_format);
      Tensor<double> B2({ 3, 3, 3 }, b_and_t_format);
      Tensor<double> B3({ 3, 3, 3 }, b_and_t_format);
      Tensor<double> T1({ 3, 3, 3 }, b_and_t_format);
      Tensor<double> T2({ 3, 3, 3 }, b_and_t_format);
      Tensor<double> A({ 3, 3, 3, 3, 3, 3, 3 }, A_format);

and this

      A(a, b, c, d, e, f, g) = B(a, h, e) * B(k, j, f) * B(d, i, g) * T(h, j, i) * T(k, b, c);

to

      A(a, b, c, d, e, f, g) = B1(a, h, e) * B2(k, j, f) * B3(d, i, g) * T1(h, j, i) * T2(k, b, c);

makes that crash go away. However, the resulting problem is more serious -- TACO spits out some invalid code:

/var/folders/b9/rl15b0d13_d4j4h9qg4tx4rr0000gn/T/taco_tmp_IcsYuV/12ef400tr7dt.c:327:31: error: use of undeclared identifier 'pA7_begin'
                          if (pA7_begin < gA) {
                              ^
/var/folders/b9/rl15b0d13_d4j4h9qg4tx4rr0000gn/T/taco_tmp_IcsYuV/12ef400tr7dt.c:342:25: error: use of undeclared identifier 'pA6_begin'
                    if (pA6_begin < fA) {
                        ^
/var/folders/b9/rl15b0d13_d4j4h9qg4tx4rr0000gn/T/taco_tmp_IcsYuV/12ef400tr7dt.c:394:9: error: use of undeclared identifier 'pA2_begin'
    if (pA2_begin < bA) {
        ^
3 errors generated.
Error at /Users/rohany/Documents/research/taco-master/src/codegen/module.cpp:156 in compile:
 Compilation command failed:

This is a harder bug to fix, and unfortunately I don't have much time to track down and fix this (maintaining TACO is not my primary obligation right now). If there's anyone else who has time cc @tensor-compiler/core-developers they could take it. Theoretically, TACO should be able to generate code for this kernel.

[Llauset]

Ok, thank you very much for your help.
I will stay tuned and if the problems are being solved I will give a try to TACO again.