1 /**
2 * Copyright 2020 Huawei Technologies Co., Ltd
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16 #include "backend/optimizer/ascend/buffer_fusion/multi_output_fusion_pass.h"
17 #include "backend/kernel_compiler/kernel_fusion.h"
18 #include "backend/session/anf_runtime_algorithm.h"
19 #include "base/core_ops.h"
20 #include "utils/ms_context.h"
21 #include "backend/optimizer/common/fusion_id_allocator.h"
22 #include "runtime/device/ascend/lic_manager.h"
23
24 namespace mindspore {
25 namespace opt {
MatchMultiOutputEltwise(const CNodePtr & cnode,const session::KernelGraph & kernel_graph,FusedNodeRecord * candidate_fusion)26 void MultiOutputFusionPass::MatchMultiOutputEltwise(const CNodePtr &cnode, const session::KernelGraph &kernel_graph,
27 FusedNodeRecord *candidate_fusion) {
28 MS_EXCEPTION_IF_NULL(cnode);
29 MS_EXCEPTION_IF_NULL(candidate_fusion);
30 std::unordered_set<AnfNodePtr> record{cnode};
31 auto eltwise_input = cnode->input(kIndex1);
32 MS_EXCEPTION_IF_NULL(eltwise_input);
33 if (CheckMultiOutputEltWiseNode(kernel_graph, eltwise_input)) {
34 (void)record.insert(eltwise_input);
35 auto input_cnode = eltwise_input->cast<CNodePtr>();
36 MS_EXCEPTION_IF_NULL(input_cnode);
37 eltwise_input = input_cnode->input(kIndex1);
38 } else {
39 return;
40 }
41 while (CheckEltWiseNode(kernel_graph, eltwise_input)) {
42 (void)record.insert(eltwise_input);
43 if (record.size() == MULTI_ELTWISE_SIZE) {
44 break;
45 }
46 auto input_cnode = eltwise_input->cast<CNodePtr>();
47 MS_EXCEPTION_IF_NULL(input_cnode);
48 eltwise_input = input_cnode->input(kIndex1);
49 }
50 if (record.size() != MULTI_ELTWISE_SIZE) {
51 return;
52 }
53 candidate_fusion->push_back(record);
54 SetRecordFusionId(record);
55 }
56
MatchSingleFusionPattern(const session::KernelGraph & kernel_graph,FusedNodeRecord * candidate_fusion)57 void MultiOutputFusionPass::MatchSingleFusionPattern(const session::KernelGraph &kernel_graph,
58 FusedNodeRecord *candidate_fusion) {
59 MS_EXCEPTION_IF_NULL(candidate_fusion);
60 if (!LicManager::GetInstance().GetPassSwitch(OptPassEnum::MultiOutputFusionPass)) {
61 return;
62 }
63 std::vector<AnfNodePtr> node_list = TopoSort(kernel_graph.get_return());
64 std::reverse(node_list.begin(), node_list.end());
65 for (auto &node : node_list) {
66 if (!AnfAlgo::IsRealCNodeKernel(node) || fusion_id_allocator->HasFusionIdAttr(node) ||
67 AnfAlgo::CheckPrimitiveType(node, prim::kPrimReturn)) {
68 continue;
69 }
70 auto cnode = node->cast<CNodePtr>();
71 MS_EXCEPTION_IF_NULL(cnode);
72 if (AnfAlgo::GetKernelType(cnode) == KernelType::TBE_KERNEL &&
73 AnfAlgo::GetFusionType(cnode) == kernel::FusionType::ELEMWISE && cnode->inputs().size() == ELTWISE_INPUT_SIZE) {
74 MatchMultiOutputEltwise(cnode, kernel_graph, candidate_fusion);
75 }
76 }
77 }
78 } // namespace opt
79 } // namespace mindspore
80