Incorrect output from conv2d deep kernel function

[keithm-xmos]

The following integration test config produces incorrect output values for the 2nd output channel group.

NOTE: This test passes when running on the x86 host.

integration_test/test_single_op_models/test_conv2d/test_explicit_padding/test_padded_conv2d.py::test_output[CONFIGS[0]]
  0:
    K_h: 9
    K_w: 7
    height: 8
    input_channels: 32
    num_threads: 1
    output_channels: 28
    pad_b: 2
    pad_l: 1
    pad_r: 0
    pad_t: 0
    strides:
    - 2
    - 2
    width: 7

Predicted and expected output values:

idx=(0, 0, 16): diff=117, expected=10, predicted=127
idx=(0, 0, 17): diff=141, expected=-14, predicted=127
idx=(0, 0, 18): diff=-51, expected=58, predicted=7
idx=(0, 0, 19): diff=113, expected=14, predicted=127
idx=(0, 0, 20): diff=51, expected=76, predicted=127
idx=(0, 0, 21): diff=106, expected=21, predicted=127
idx=(0, 0, 22): diff=157, expected=-30, predicted=127
idx=(0, 0, 23): diff=62, expected=51, predicted=113
idx=(0, 0, 24): diff=-104, expected=115, predicted=11
idx=(0, 0, 25): diff=-199, expected=71, predicted=-128
idx=(0, 0, 26): diff=126, expected=1, predicted=127
idx=(0, 0, 27): diff=-141, expected=13, predicted=-128

[keithm-xmos]

From @astewart-xmos

I'm realizing now that I'm not sure I understand the parameters listed in that ticket.

It says the width (presumably the input width?) is 7. The kernel width is 7, the left padding is 1, and the horizontal stride is 2.
Does that mean that the output width is only 1 pixel?

Same with the output height. height = 8, K_h = 9, pad_t = 0 ---> These suggest that for the first output pixel, the bottom row of the convolution window will be in the padding below the input image.

But then, pad_b = 2 and vertical stride = 2. If the output height is 1 pixel, then the second row of padding on the bottom is unnecessary (i.e. pad_b should be 1). If the output height is 2 pixels, then then for the second row of output pixels, the bottom row of the convolution window is below the padding ( i.e. pad_b should be 3)