onnxruntime

45f493e8 - Fix incorrect pad indices in AveragePool count_include_pad computation (#27375)

Fix incorrect pad indices in AveragePool count_include_pad computation (#27375) ### Description Fixes #26708 `AveragePool2DTask` and `AveragePool3DTask` in `pool_functors.h` used incorrect pad indices when computing `hend` (and `wend` in 3D) for the `count_include_pad` divisor calculation. ONNX pads format for 2D is `[h_begin, w_begin, h_end, w_end]`. The code was using `pads[1]` (w_begin) instead of `pads[2]` (h_end) to clamp `hend`, causing the padding region to be excluded from the divisor when `count_include_pad=1` with asymmetric pads (e.g. bottom/right only). The same issue existed in `AveragePool3DTask` where pads format is `[h_begin, w_begin, d_begin, h_end, w_end, d_end]`: - `hend` used `pads[1]` instead of `pads[3]` - `wend` used `pads[3]` instead of `pads[4]` ### Example Input `[[1,1],[1,1]]` with `pads=[0,0,1,1]`, `kernel_shape=[2,2]`, `count_include_pad=1`: - **Before (wrong):** `[[1.0, 0.5], [1.0, 0.5]]` - **After (correct):** `[[1.0, 0.5], [0.5, 0.25]]` Verified against ONNX ReferenceEvaluator. ### Changes - `pool_functors.h`: Fix pad indices in `AveragePool2DTask` and `AveragePool3DTask` - `pool_op_test.cc`: Add regression test `AveragePool_CountIncludePad_AsymmetricPads` ### Motivation and Context When using `AveragePool` with `count_include_pad=1` and asymmetric padding (e.g. only bottom and right), the average was computed with the wrong divisor, producing results inconsistent with the ONNX spec.