[BUG] Unexpected behaviour on Nvidia Jetson Orin

[SamDSchofield]

Describe the Bug
MatX code that appears to work correctly on my desktop machine works incorrectly (operations returns all zeros) when running on my Jetson Orin AGX (Jetpack 6.0).

To Reproduce
Run the following code on a Jetson

#include <matx.h>

int main(int argc, char **argv) {
    auto a = matx::make_tensor<float>({3, 3});
    auto b = matx::make_tensor<float>({3, 3});

    a.SetVals({{1,2,3}, {4,5,6}, {7,8,9}});

    (b = 2*a).run();

    matx::print(a);
    matx::print(b);
}

Expected Behavior
The code should output (result on my desktop machine):

Tensor{float} Rank: 2, Sizes:[3, 3], Strides:[3,1]
000000:  1.0000e+00  2.0000e+00  3.0000e+00 
000001:  4.0000e+00  5.0000e+00  6.0000e+00 
000002:  7.0000e+00  8.0000e+00  9.0000e+00 
Tensor{float} Rank: 2, Sizes:[3, 3], Strides:[3,1]
000000:  2.0000e+00  4.0000e+00  6.0000e+00 
000001:  8.0000e+00  1.0000e+01  1.2000e+01 
000002:  1.4000e+01  1.6000e+01  1.8000e+01 

Instead, it outputs (on Jetson):

tensor_2_f32: Tensor{float} Rank: 2, Sizes:[3, 3], Strides:[3,1]
000000:  1.0000e+00  2.0000e+00  3.0000e+00 
000001:  4.0000e+00  5.0000e+00  6.0000e+00 
000002:  7.0000e+00  8.0000e+00  9.0000e+00 
tensor_2_f32: Tensor{float} Rank: 2, Sizes:[3, 3], Strides:[3,1]
000000:  0.0000e+00  0.0000e+00  0.0000e+00 
000001:  0.0000e+00  0.0000e+00  0.0000e+00 
000002:  0.0000e+00  0.0000e+00  0.0000e+00 

System Details (please complete the following information):

• OS: Jetpack 6.0
• CUDA version: 12.6.68
• g++ version: 11.4.0

[cliffburdick]

Hi @SamDSchofield , can you give us your compiler command? I just wanted to make sure you're using the correct architecture. You can also check for CUDA errors using cuda-gdb.

[SamDSchofield]

Hi @cliffburdick, it turns out the issue was caused by having my CMAKE_CUDA_ARCHITECTURES set to the wrong value. Changing set(CMAKE_CUDA_ARCHITECTURES 70) to set(CMAKE_CUDA_ARCHITECTURES 87) in my CMakeLists.txt solved the issue. Thanks for your help!