|
| 1 | +# Copyright 2024 Arm Limited and/or its affiliates. |
| 2 | +# All rights reserved. |
| 3 | +# |
| 4 | +# This source code is licensed under the BSD-style license found in the |
| 5 | +# LICENSE file in the root directory of this source tree. |
| 6 | + |
| 7 | +import numpy as np |
| 8 | +from executorch.backends.arm._passes.arm_pass_utils import ( |
| 9 | + create_node, |
| 10 | + get_first_fake_tensor, |
| 11 | +) |
| 12 | +from executorch.backends.arm.tosa_quant_utils import dq_op, q_op |
| 13 | +from executorch.exir.dialects._ops import ops as exir_ops |
| 14 | +from executorch.exir.pass_base import ExportPass, PassResult |
| 15 | + |
| 16 | + |
| 17 | +class DecomposeLinearPass(ExportPass): |
| 18 | + """ |
| 19 | + This pass decomposes linear into a Conv2D with the required view operations. |
| 20 | + linear(x, weights, bias) becomes: |
| 21 | + x_reshaped = view(x) |
| 22 | + weights_reshaped = view(weights) |
| 23 | + conv2d = conv2d(x_reshaped, weights_reshaped, bias) |
| 24 | + output = view(conv2d) |
| 25 | + It also inserts q/dq pairs if the linear node was quantized. |
| 26 | + """ |
| 27 | + |
| 28 | + def call(self, graph_module): |
| 29 | + for node in graph_module.graph.nodes: |
| 30 | + if node.op != "call_function": |
| 31 | + continue |
| 32 | + if node.target != exir_ops.edge.aten.linear.default: |
| 33 | + continue |
| 34 | + args = node.args |
| 35 | + input = args[0] |
| 36 | + weights = args[1] |
| 37 | + bias = args[2] if len(args) > 2 else None |
| 38 | + output_shape = get_first_fake_tensor(node).shape |
| 39 | + input_shape = get_first_fake_tensor(input).shape |
| 40 | + weights_shape = get_first_fake_tensor(weights).shape |
| 41 | + batches = int(np.prod(input_shape[:-1])) if len(input_shape) > 1 else 1 |
| 42 | + # input has shape (..., Ci) |
| 43 | + input_reshaped_shape = [batches, input_shape[-1], 1, 1] |
| 44 | + # weights have shape (Co, Ci) |
| 45 | + weights_reshaped_shape = [weights_shape[0], weights_shape[1], 1, 1] |
| 46 | + |
| 47 | + with graph_module.graph.inserting_before(node): |
| 48 | + quantize = input.op == "call_function" and input.target == dq_op |
| 49 | + q_params = input.args[1:] if quantize else None |
| 50 | + # Reshape input to 4D with shape (N, Ci, 1, 1) |
| 51 | + input_reshaped = create_node( |
| 52 | + graph=graph_module.graph, |
| 53 | + op_target=exir_ops.edge.aten.view_copy.default, |
| 54 | + args=(input, input_reshaped_shape), |
| 55 | + kwargs={}, |
| 56 | + quantize=quantize, |
| 57 | + q_params=q_params, |
| 58 | + ) |
| 59 | + |
| 60 | + quantize = weights.op == "call_function" and weights.target == dq_op |
| 61 | + q_params = weights.args[1:] if quantize else None |
| 62 | + # Reshape weights to 4D with shape (Co, Ci, 1, 1) |
| 63 | + weights_reshaped = create_node( |
| 64 | + graph=graph_module.graph, |
| 65 | + op_target=exir_ops.edge.aten.view_copy.default, |
| 66 | + args=(weights, weights_reshaped_shape), |
| 67 | + kwargs={}, |
| 68 | + quantize=quantize, |
| 69 | + q_params=q_params, |
| 70 | + ) |
| 71 | + |
| 72 | + consumer_node = list(node.users)[0] |
| 73 | + quantize = ( |
| 74 | + consumer_node.op == "call_function" and consumer_node.target == q_op |
| 75 | + ) |
| 76 | + q_params = consumer_node.args[1:] if quantize else None |
| 77 | + conv = create_node( |
| 78 | + graph=graph_module.graph, |
| 79 | + op_target=exir_ops.edge.aten.convolution.default, |
| 80 | + args=( |
| 81 | + input_reshaped, |
| 82 | + weights_reshaped, |
| 83 | + bias, |
| 84 | + [1, 1], # strides |
| 85 | + [0, 0], # padding |
| 86 | + [1, 1], # dilation |
| 87 | + False, # transposed |
| 88 | + [0, 0], # output padding |
| 89 | + 1, # groups |
| 90 | + ), |
| 91 | + kwargs={}, |
| 92 | + quantize=quantize, |
| 93 | + q_params=q_params, |
| 94 | + ) |
| 95 | + |
| 96 | + with graph_module.graph.inserting_after(conv): |
| 97 | + # Reshape output to same rank as original input with shape (..., Co) |
| 98 | + # No need to insert q/dq pair as Conv2D node above has inserted them if |
| 99 | + # required. |
| 100 | + output = create_node( |
| 101 | + graph=graph_module.graph, |
| 102 | + op_target=exir_ops.edge.aten.view_copy.default, |
| 103 | + args=(conv, list(output_shape)), |
| 104 | + kwargs={}, |
| 105 | + ) |
| 106 | + |
| 107 | + node.replace_all_uses_with(output) |
| 108 | + graph_module.graph.erase_node(node) |
| 109 | + graph_module.graph.eliminate_dead_code() |
| 110 | + graph_module.recompile() |
| 111 | + graph_module = super().call(graph_module).graph_module |
| 112 | + return PassResult(graph_module, True) |
0 commit comments