1 /*
2 * Copyright (c) 2018-2020 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "arm_compute/graph/mutators/DepthConcatSubTensorMutator.h"
25
26 #include "arm_compute/graph/Graph.h"
27 #include "arm_compute/graph/Logger.h"
28 #include "arm_compute/graph/Utils.h"
29 #include "arm_compute/graph/algorithms/TopologicalSort.h"
30 #include "arm_compute/graph/backends/BackendRegistry.h"
31 #include "arm_compute/graph/nodes/ConcatenateLayerNode.h"
32
33 #include "support/Cast.h"
34 #include "support/Iterable.h"
35
36 namespace arm_compute
37 {
38 namespace graph
39 {
name()40 const char *DepthConcatSubTensorMutator::name()
41 {
42 return "DepthConcatSubTensorMutator";
43 }
44
type() const45 IGraphMutator::MutationType DepthConcatSubTensorMutator::type() const
46 {
47 return IGraphMutator::MutationType::Backend;
48 }
49
mutate(Graph & g)50 void DepthConcatSubTensorMutator::mutate(Graph &g)
51 {
52 // Early exit if no Concatenation layers exist in graph
53 if(g.nodes(NodeType::ConcatenateLayer).empty())
54 {
55 return;
56 }
57
58 // Perform topological sort
59 std::vector<NodeID> topological_sorted_node_ids = dfs(g);
60
61 // Should be in reverse order of execution
62 for(auto &node_id : arm_compute::utils::iterable::reverse_iterate(topological_sorted_node_ids))
63 {
64 INode *node = g.node(node_id);
65 if(node != nullptr && node->type() == NodeType::ConcatenateLayer && node->output(0) != nullptr)
66 {
67 // Get output tensor
68 auto output_tensor = node->output(0);
69
70 // Check concatenation axis (Sub-tensor optimization is supported for concatenation axis >=2)
71 auto *concat_node = arm_compute::utils::cast::polymorphic_downcast<ConcatenateLayerNode *>(node);
72 if(output_tensor == nullptr || get_dimension_idx(output_tensor->desc().layout, concat_node->concatenation_axis()) < 2)
73 {
74 continue;
75 }
76
77 // Check that all tensor have the same target, valid inputs and same quantization info
78 bool is_valid = std::all_of(node->input_edges().cbegin(), node->input_edges().cend(),
79 [&](const EdgeID & eid)
80 {
81 return (g.edge(eid) != nullptr) && (g.edge(eid)->tensor() != nullptr) && (g.edge(eid)->tensor()->desc().target == output_tensor->desc().target)
82 && (g.edge(eid)->tensor()->desc().quant_info == output_tensor->desc().quant_info);
83 });
84
85 // Create subtensors
86 if(is_valid && is_target_supported(output_tensor->desc().target))
87 {
88 ARM_COMPUTE_LOG_GRAPH_VERBOSE("Using sub-tensors for the node with ID : "
89 << node->id() << " and name : " << node->name() << std::endl);
90 // Create sub-tensor handles
91 unsigned depth = 0;
92 for(unsigned int i = 0; i < node->input_edges().size(); ++i)
93 {
94 auto input_tensor = node->input(i);
95 const auto input_shape = input_tensor->desc().shape;
96
97 backends::IDeviceBackend &backend = backends::BackendRegistry::get().get_backend(input_tensor->desc().target);
98 std::unique_ptr<ITensorHandle> handle = backend.create_subtensor(output_tensor->handle(), input_shape, Coordinates(0, 0, depth), false);
99 input_tensor->set_handle(std::move(handle));
100
101 depth += input_shape.z();
102 }
103
104 auto *dc_node = arm_compute::utils::cast::polymorphic_downcast<ConcatenateLayerNode *>(node);
105 dc_node->set_enabled(false);
106 }
107 }
108 }
109 }
110 } // namespace graph
111 } // namespace arm_compute
112