[GPU] Padding implementation for same_upper and same_lower in onednn conv and pooling

[wilson-seok]

Description of the issue(symptom, root-cause, how it was resolved)

• Customer's dynamic shape model has accuracy drop than static shape model.
• The model has auto pad(same_upper) type in convolution and pooling. During graph transformation, padding value is updated in static shape model while dynamic shape doesn't have a chance to update. Onednn primitive creator has padding calculation and update, but it didn't consider auto pad(same_upper/same_lower).
• This PR implements same_upper/same_lower cases in convolution and pooling.

The code and line that caused this issue (if it is not changed directly)

• openvino/src/plugins/intel_gpu/src/graph/impls/onednn/convolution_onednn.cpp

  Line 71 in a96a0b1

  pad_r[i] = (os - 1) * stride[i] - is + kernel_range - pad_l[i];

• openvino/src/plugins/intel_gpu/src/graph/impls/onednn/pooling_onednn.cpp

  Line 59 in a96a0b1

  pad_r[i] = (output_md.get_dims()[2 + i] - 1) * stride[i] - input_md.get_dims()[2 + i] + ((kernel[i] - 1) * dilation[i]) - pad_l[i] + 1;

Reproduction step and snapshot (if applicable. Do not attach for customer model)

• $ ./benchmark_app -d GPU.1 -m ~/task/blackmagic/Blackmagic_Dynamic_v2/DynamicModels/PersonMask_ps_mask_faster.xml -i ~/task/blackmagic/Blackmagic_Dynamic_v2/InputNpys/PersonMask_ps_mask_faster_input_0.npy -data_shape [1,14,607,1080] -hint none -nstreams 1 -niter 1

Problematic graph

Checklist

• [v] Is it a proper fix? (not a workaround)
• Did you include test case for this fix, if necessary?
• Did you review existing test that can be extended to cover this scenario? Which test did you review?

Tickets:

• id

[Lyamin-Roman]

Should there be - pad_l[i] here?
It seems to imply that this is a total padding, since we then subtract from this value to obtain the l/r paddings
Or did I misunderstand?