|
| 1 | +# Copyright 2025 Arm Limited and/or its affiliates. |
| 2 | +# |
| 3 | +# This source code is licensed under the BSD-style license found in the |
| 4 | +# LICENSE file in the root directory of this source tree. |
| 5 | + |
| 6 | + |
| 7 | +import torch |
| 8 | +from executorch.backends.arm.operators.operator_validation_utils import ( |
| 9 | + adjust_pooling_pad_if_needed, |
| 10 | +) |
| 11 | +from executorch.exir.dialects._ops import ops as exir_ops |
| 12 | +from executorch.exir.pass_base import ExportPass |
| 13 | + |
| 14 | +edge_div_ops = (exir_ops.edge.aten.avg_pool2d.default,) |
| 15 | +aten_div_ops = (torch.ops.aten.avg_pool2d.default,) |
| 16 | + |
| 17 | + |
| 18 | +def get_decomposition(op) -> tuple: |
| 19 | + if op in edge_div_ops: |
| 20 | + return ( |
| 21 | + exir_ops.edge.aten.full.default, |
| 22 | + exir_ops.edge.aten.cat.default, |
| 23 | + exir_ops.edge.aten.avg_pool2d.default, |
| 24 | + exir_ops.edge.aten.mul.Tensor, |
| 25 | + ) |
| 26 | + if op in aten_div_ops: |
| 27 | + return ( |
| 28 | + torch.ops.aten.full.default, |
| 29 | + torch.ops.aten.cat.default, |
| 30 | + torch.ops.aten.avg_pool2d.default, |
| 31 | + torch.ops.aten.mul.Tensor, |
| 32 | + ) |
| 33 | + raise RuntimeError(f"Can't get div decomposition for op {op}") |
| 34 | + |
| 35 | + |
| 36 | +class DecomposeAvgPool2d(ExportPass): |
| 37 | + """ """ |
| 38 | + |
| 39 | + def call_operator(self, op, args, kwargs, meta): |
| 40 | + if op not in (edge_div_ops + aten_div_ops): |
| 41 | + return super().call_operator(op, args, kwargs, meta) |
| 42 | + |
| 43 | + full_op, cat_op, avgpool_op, mul_op = get_decomposition(op) |
| 44 | + |
| 45 | + x = args[0] |
| 46 | + kernel_h, kernel_w = args[1] |
| 47 | + kernel_size = kernel_h * kernel_w |
| 48 | + stride_h, stride_w = args[2] |
| 49 | + pad_h, pad_w = new_pad_h, new_pad_w = args[3] if len(args) > 3 else (0, 0) |
| 50 | + ceil_mode = args[4] if len(args) > 4 else False |
| 51 | + count_include_pad = args[5] if len(args) > 5 else True |
| 52 | + divisor_override = args[6] if len(args) > 6 else None |
| 53 | + |
| 54 | + n, c, h, w = x.data.shape |
| 55 | + post_pad_w, post_pad_h = (0, 0) |
| 56 | + |
| 57 | + # Count_include_pad == False means that we use a different divisor for edge elements |
| 58 | + # When divisor_override is set, this will be overriden anyways. |
| 59 | + # It is easier to replace a constant divisor, so set count_include_pad == True |
| 60 | + if divisor_override is not None: |
| 61 | + count_include_pad = True |
| 62 | + |
| 63 | + # Add width padding manually if count_include_pad |
| 64 | + if count_include_pad and pad_w > 0: |
| 65 | + pre_pad_shape = [n, c, h, pad_w] |
| 66 | + pre_pad = super().call_operator(full_op, (pre_pad_shape, 0.0), kwargs, meta) |
| 67 | + |
| 68 | + if ceil_mode and divisor_override is None: |
| 69 | + post_pad_w = pad_w |
| 70 | + else: |
| 71 | + post_pad_w = adjust_pooling_pad_if_needed( |
| 72 | + w, kernel_w, stride_w, pad_w, ceil_mode |
| 73 | + ) |
| 74 | + |
| 75 | + if post_pad_w > 0: |
| 76 | + post_pad_shape = [n, c, h, post_pad_w] |
| 77 | + post_pad = super().call_operator( |
| 78 | + full_op, (post_pad_shape, 0.0), kwargs, meta |
| 79 | + ) |
| 80 | + cat_nodes = [pre_pad, x, post_pad] |
| 81 | + else: |
| 82 | + cat_nodes = [pre_pad, x] |
| 83 | + |
| 84 | + x = super().call_operator(cat_op, (cat_nodes, 3), kwargs, meta) |
| 85 | + new_pad_w = 0 |
| 86 | + |
| 87 | + # Add height padding manually if count_include_pad |
| 88 | + if count_include_pad and pad_h > 0: |
| 89 | + pre_pad_shape = [n, c, pad_h, w + pad_w + post_pad_w] |
| 90 | + pre_pad = super().call_operator(full_op, (pre_pad_shape, 0.0), kwargs, meta) |
| 91 | + |
| 92 | + if ceil_mode and divisor_override is None: |
| 93 | + post_pad_h = pad_h |
| 94 | + else: |
| 95 | + post_pad_h = adjust_pooling_pad_if_needed( |
| 96 | + h, kernel_h, stride_h, pad_h, ceil_mode |
| 97 | + ) |
| 98 | + |
| 99 | + if post_pad_h > 0: |
| 100 | + post_pad_shape = [n, c, post_pad_h, w + pad_w + post_pad_w] |
| 101 | + post_pad = super().call_operator( |
| 102 | + full_op, (post_pad_shape, 0.0), kwargs, meta |
| 103 | + ) |
| 104 | + cat_nodes = [pre_pad, x, post_pad] |
| 105 | + else: |
| 106 | + cat_nodes = [pre_pad, x] |
| 107 | + |
| 108 | + x = super().call_operator(cat_op, (cat_nodes, 2), kwargs, meta) |
| 109 | + new_pad_h = 0 |
| 110 | + |
| 111 | + avgpool_args = (x, args[1], args[2], [new_pad_h, new_pad_w], ceil_mode, False) |
| 112 | + x = super().call_operator(avgpool_op, avgpool_args, kwargs, meta) |
| 113 | + |
| 114 | + # Multiply by factor (kernel_size / divisor_override) if divisor_override |
| 115 | + if divisor_override is not None and divisor_override != kernel_size: |
| 116 | + override_multiplier = super().call_operator( |
| 117 | + full_op, ([1, 1, 1, 1], kernel_size / divisor_override), kwargs, meta |
| 118 | + ) |
| 119 | + x = super().call_operator(mul_op, (x, override_multiplier), kwargs, meta) |
| 120 | + |
| 121 | + return x |
0 commit comments