[BUG] Reduce operations give partly wrong results with 3D tensors. about matx HOT 4 CLOSED

thanks for the report @ZJUGuoShuai. we recently moved most reductions to cub, so I will check for a regression

from matx.

BTW, it works great with 1D and 2D tensors :)

from matx.

Hi @ZJUGuoShuai, it looks like the order isn't quite what we expect, and we should have errored. The way our reductions work is the output tensor must have the same inner ranks as the input tensor. In other words,

{2,3,2} -> {2} // valid
{2,3,2} -> {6} // not valid

By going from a rank 3 to a rank one you are asking it to sum over the outer 2 dimensions, and the final output would have two values that are the sum of the two 3x2 matrices:

auto mat = vec.View({2, 3, 2});
auto mat_sum = matx::make_tensor<float, 1>({2});
matx::sum(mat_sum, mat);

000000: 21.0000
000001: 57.0000

It looks like instead you want to just sum over 6 rows of 2 columns each, which would be:

  auto vec = matx::make_tensor<float, 1>({12});
vec.SetVals({1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12});
auto mat = vec.View({6,2});
auto mat_sum = matx::make_tensor<float, 1>({6});
matx::sum(mat_sum, mat);
mat_sum.Print();

000000: 3.0000
000001: 7.0000
000002: 11.0000
000003: 15.0000
000004: 19.0000
000005: 23.0000

I also had to make a small change since we did have a regression when using CUB for row-wise reductions, so that's back to take a slower path for now and will be fixed next week. You should be able to get the correct answer now with the latest code and the example above.

from matx.

@cliffburdick Thanks, now I know how to use Reduce operations:

the output tensor must have the same inner ranks as the input tensor

from matx.