Output Tensor Generated by TACO might be Incorrect for Certain Tensor Computation

[jiawen11]

Hi,

I used TACO for tensor contraction following the "Library Example" shown on the official page https://github.com/tensor-compiler/taco. I found that the final results generated by TACO are incorrect.

Does it mean TACO does not support tensor contraction or are there issues within TACO for tensor contraction? In addition, does TACO support sorting the output tensor?

Any comments are welcome!

Thanks!

---

Code:

include

include "taco.h"

include <sys/time.h>

include

include

include

using namespace taco;

int main(int argc, char argv[]) { IndexVar i, j, k, l, m; Tensor X = read("x.tns", Format({Sparse, Sparse, Sparse})); Tensor Y = read("y.tns", Format({Sparse, Sparse, Sparse})); Tensor Z("Z", {4, 3, 4, 3}, Format({Sparse, Sparse, Sparse, Sparse})); X.pack(); Y.pack(); Z(j, k, l, m) = X(i,j,k) Y(i,l,m); Z.compile(); Z.assemble(); Z.compute(); write("output.tns", Z); }

Input tensor (x.tns is the same as y.tns): 0 0 0 1.000000 0 1 0 2.000000 0 3 0 10.00000 1 0 1 7.000000 1 1 0 8.000000 1 2 0 6.000000 2 0 2 3.000000 2 2 0 4.000000 3 0 1 5.000000

Incorrect output tensor (output.tns): 1 1 1 1 1 1 1 2 1 2 1 1 4 1 10 2 1 1 1 2 2 1 2 1 4 (should be accumulated with "2 1 2 1 64") 2 1 4 1 20 4 1 1 1 10 4 1 2 1 20 4 1 4 1 100 1 2 1 2 49 (should be accumulated with "1 2 1 2 25") 1 2 2 1 56 1 2 3 1 42 2 1 1 2 56 2 1 2 1 64 2 1 3 1 48 3 1 1 2 42 3 1 2 1 48 3 1 3 1 36 (should be accumulated with "3 1 3 1 16") 1 3 1 3 9 1 3 3 1 12 3 1 1 3 12 3 1 3 1 16 1 2 1 2 25

[Infinoid]

Hi,

Taco does support tensor contraction. But there are issues with sparse outputs at the moment. See issues #184, #297. I think that's what's happening here.