zero padding not working in convolve2_nn [BUG]

[MichaelMcCulloch]

Description

• Details regarding the bug
  • Zero Padding not allowed by Convolve2_nn
• Did you build ArrayFire yourself or did you use the official installers
  • Official Installers
• Which backend is experiencing this issue? (CPU, CUDA, OpenCL)
  • all of them
• Do you have a workaround?
  • No
• Can the bug be reproduced reliably on your system?
  • Yes
• A clear and concise description of what you expected to happen.
  • My understanding was this was resolved in fix zero padding in convolve2NN arrayfire#2820
• Run your executable with AF_TRACE=all and AF_PRINT_ERRORS=1 environment
  variables set.
• Screenshot or terminal output of the results from above run

Reproducible Code and/or Steps

#[test]
    fn mnist__convolutional_neural_net__learns_mnist_fail() {
        let signal = randu::<f32>(dim4!(8, 8, 3, 1));

        let quadrant1 = Array::new(&vec![1f32, 0.0, 0.0, 0.0], dim4![2, 2, 1, 1]);

        let q1 = convolve2_nn(
            &signal,
            &quadrant1,
            dim4!(2, 2, 1, 1),
            dim4!(0, 0, 1, 1),
            dim4!(1, 1, 1, 1),
        );

        af_print!("q1", q1);
    }

System Information

Please provide the following information:

1. ArrayFire version
   • 3.8
2. Devices installed on the system
   *. RTX 3090, R9 5950X
3. (optional) Output from the af::info() function if applicable.

ArrayFire v3.8.1 (CUDA, 64-bit Linux, build default)
Platform: CUDA Runtime 11.5, Driver: 510.47.03
[0] NVIDIA GeForce RTX 3090, 24252 MB, CUDA Compute 8.6
test tests::info____ ... ok

4. Output from the following scripts:

LSB Version: n/a
Distributor ID: ManjaroLinux
Description: Manjaro Linux
Release: 21.2.3
Codename: Qonos

name, memory.total [MiB], driver_version
NVIDIA GeForce RTX 3090, 24576 MiB, 510.47.03

Checklist

• [Y] Using the latest available ArrayFire release
• [Y] GPU drivers are up to date

[syurkevi]

For devs reference, issue is on this line:

arrayfire-rust/src/ml/mod.rs

Line 81 in f53bf4c

padding.ndims() as c_uint,

and the corresponding line for the gradient function as well.
ndims() will return 0 for 0 padding when it should pass a "2" (or matching signal.ndims() ) instead

[9prady9]

I am contemplating if I should avoid using dim4 for padding arguments all together. I am looking into it, will send in a fix soon. Thank you for reporting it @MichaelMcCulloch