[AIROCMLIR-445] Lower migraphx.backwards_data_convolution

[Mr-Anyone]

Motivation

Being able to support 1D,2D,3D backwards convolution with for CPU with all the attributes (strides, dilation, padding, group).

Technical Details

For backwards convolution is similar to forward convolution, and it into these steps:

1. Expand the input and the kernel in the G dimension. Input goes from NC* into NGC*, and filter goes form CF* into CGF*
2. Emit the linalg.generic that should be like the python loop below.
3. Implements padding through tensor.extract_slice

def my_grouped_impl(input_np, filter_np, output_shape, stride=(1, 1), dilation=(1, 1)):
  result = np.zeros(output_shape, dtype=np.float32)
  batch, group, channel, input_height, input_width = input_np.shape
  group, filter_count, channel, filter_height, filter_width = filter_np.shape

  for n in range(batch):# 0
    for g in range(group):# 1
      for hi in range(input_height):# 2
        for wi in range(input_width): # 3
          for f in range(filter_count):# 4
            # reduction starts here!
            for c in range(channel):# 5
              for hk in range(filter_height):# 6
                for wk in range(filter_width):# 7
                  height_access = hi*stride[0] + dilation[0] * hk
                  width_access = wi*stride[1] + dilation[1] * wk
                  result[n, g, f, height_access, width_access] +=  input_np[n, g, c, hi, wi] * filter_np[c, g, f, hk, wk]
  return result

If we have padding, it should look like the following code structure:

linalg.generic ins (...)  outs(%first) ... // computing the transposed kernel
%output = tensor.extract_slice %first [0, 0, padLow0, padLow1, ....][N, F, Ho, Wo, ....][1, 1, 1, ....] // apply padding

Test Plan

Passing all e2e tests.

Test Result

Passed all e2e test that has mixr-bwd-data-*

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Mr-Anyone]

A lot of the changes are refactor (moving the lambda function into a static function) - it makes more sense to move some of the changes to #2241