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
17 #include "plugin/device/cpu/hal/device/cpu_memory_manager.h"
18 #include "include/backend/anf_runtime_algorithm.h"
19 #include "include/common/utils/anfalgo.h"
20 #include "utils/ms_context.h"
21 #include "include/common/utils/convert_utils.h"
22 namespace mindspore {
23 namespace device {
24 namespace cpu {
MemMalloc(size_t size)25 uint8_t *CPUMemoryManager::MemMalloc(size_t size) {
26 auto block = std::make_shared<std::vector<uint8_t>>();
27 try {
28 block->resize(size, 0);
29 auto ptr = block->data();
30 mem_block_map_[ptr] = block;
31 return ptr;
32 } catch (const std::exception &e) {
33 MS_LOG(EXCEPTION) << "Malloc memory failed: size " << size;
34 }
35 }
36
MallocStaticMem(size_t size,bool,uint32_t)37 uint8_t *CPUMemoryManager::MallocStaticMem(size_t size, bool, uint32_t) {
38 auto ptr = MemMalloc(size);
39 static_mem_[ptr] = size;
40 return ptr;
41 }
42
MallocDynamicMem(size_t size,bool)43 uint8_t *CPUMemoryManager::MallocDynamicMem(size_t size, bool) {
44 void *ptr = nullptr;
45 size_t min_size = 0;
46 // first find the smallest cached_mem_ which fits the size
47 for (auto &&iter : cached_mem_) {
48 if (iter.second >= size) {
49 if (min_size == 0 || iter.second < min_size) {
50 ptr = iter.first;
51 min_size = iter.second;
52 }
53 }
54 }
55 if (ptr != nullptr) {
56 if (memset_s(ptr, size, 0, size) != EOK) {
57 free(ptr);
58 MS_LOG(EXCEPTION) << "Failed to init memory.";
59 }
60 dynamic_mem_[ptr] = min_size;
61 (void)cached_mem_.erase(ptr);
62 return reinterpret_cast<uint8_t *>(ptr);
63 }
64 // if not found, malloc
65 auto new_ptr = MemMalloc(size);
66 dynamic_mem_[new_ptr] = size;
67 return new_ptr;
68 }
69
ResetDynamicMemory()70 void CPUMemoryManager::ResetDynamicMemory() {
71 // don't free, for multi graph
72 for (auto &&iter : dynamic_mem_) {
73 cached_mem_[iter.first] = iter.second;
74 }
75 dynamic_mem_.clear();
76 }
77
~CPUMemoryManager()78 CPUMemoryManager::~CPUMemoryManager() { MemFree(); }
79
MemFree()80 void CPUMemoryManager::MemFree() noexcept {
81 if (mem_ptr_ != nullptr) {
82 mem_ptr_ = nullptr;
83 mem_size_ = 0;
84 }
85 static_mem_.clear();
86 dynamic_mem_.clear();
87 cached_mem_.clear();
88 mem_block_map_.clear();
89 }
90
AssignMemory(const session::KernelGraph * graph)91 void CPUMemoryManager::AssignMemory(const session::KernelGraph *graph) {
92 size_t graph_mem_size = mem_plan_.MemPlan(graph);
93 if (graph_mem_size > mem_size_) {
94 if (mem_size_ > 0) {
95 dynamic_mem_[mem_ptr_] = mem_size_;
96 mem_size_ = 0;
97 }
98 mem_ptr_ = MemMalloc(graph_mem_size);
99 if (mem_ptr_ != nullptr) {
100 MS_LOG(INFO) << "Simple MemPlan GraphMemSize [" << graph_mem_size << "]";
101 mem_size_ = graph_mem_size;
102 dynamic_malloc_ = false;
103 } else {
104 MS_LOG(INFO) << "Switch to dynamic malloc";
105 dynamic_malloc_ = true;
106 }
107 }
108 if (dynamic_malloc_) {
109 return;
110 }
111 mem_plan_.MemAssign(graph, mem_ptr_);
112 }
113
StaticMemMalloc(size_t mem_size)114 void *CPUMemoryManager::StaticMemMalloc(size_t mem_size) {
115 auto ptr = MemMalloc(mem_size);
116 if (ptr != nullptr) {
117 static_mem_[ptr] = mem_size;
118 return ptr;
119 } else {
120 MS_LOG(EXCEPTION) << "Malloc memory failed: size " << mem_size;
121 }
122 }
123
MemFree(void * ptr)124 void CPUMemoryManager::MemFree(void *ptr) {
125 auto iter = static_mem_.find(ptr);
126 if (iter != static_mem_.end()) {
127 (void)static_mem_.erase(iter);
128 auto block_iter = mem_block_map_.find(ptr);
129 if (block_iter != mem_block_map_.end()) {
130 (void)mem_block_map_.erase(block_iter);
131 }
132 }
133 }
134
IncreaseSummaryRefCount(const session::NamedSummaryOutputs & summary_outputs) const135 void CPUMemoryManager::IncreaseSummaryRefCount(const session::NamedSummaryOutputs &summary_outputs) const {
136 if (!dynamic_malloc_) {
137 return;
138 }
139 if (summary_outputs.empty()) {
140 return;
141 }
142 for (auto &output_item : summary_outputs) {
143 auto node = output_item.second.first;
144 size_t index = IntToSize(output_item.second.second);
145 auto address = AnfAlgo::GetMutableOutputAddr(node, index);
146 MS_EXCEPTION_IF_NULL(address);
147 address->set_ref_count(address->ref_count() + 1);
148 }
149 }
150
DecreaseSummaryRefCount(const session::NamedSummaryOutputs & summary_outputs)151 void CPUMemoryManager::DecreaseSummaryRefCount(const session::NamedSummaryOutputs &summary_outputs) {
152 if (!dynamic_malloc_) {
153 return;
154 }
155 if (summary_outputs.empty()) {
156 return;
157 }
158 for (auto &output_item : summary_outputs) {
159 auto node = output_item.second.first;
160 size_t index = IntToSize(output_item.second.second);
161 auto address = AnfAlgo::GetMutableOutputAddr(node, index);
162 MS_EXCEPTION_IF_NULL(address);
163 address->DecreaseRefCount();
164 if (address->ref_count() == 0 && address->GetDevicePtr() != nullptr) {
165 MemFree(address->GetDevicePtr());
166 address->SetDevicePtr(nullptr);
167 }
168 }
169 }
170
IncreaseAddressRefCount(const session::KernelGraph * graph) const171 void CPUMemoryManager::IncreaseAddressRefCount(const session::KernelGraph *graph) const {
172 if (!dynamic_malloc_) {
173 return;
174 }
175 MS_EXCEPTION_IF_NULL(graph);
176 auto kernels = graph->execution_order();
177 for (const auto &kernel : kernels) {
178 MS_EXCEPTION_IF_NULL(kernel);
179 size_t input_num = common::AnfAlgo::GetInputTensorNum(kernel);
180 for (size_t i = 0; i < input_num; ++i) {
181 auto address = AnfAlgo::GetPrevNodeMutableOutputAddr(kernel, i);
182 MS_EXCEPTION_IF_NULL(address);
183 address->set_ref_count(address->ref_count() + 1);
184 }
185 auto kernel_mod = AnfAlgo::GetKernelMod(kernel);
186 MS_EXCEPTION_IF_NULL(kernel_mod);
187 for (size_t i = 0; i < kernel_mod->GetWorkspaceSizeList().size(); ++i) {
188 auto address = AnfAlgo::GetWorkspaceAddr(kernel, i);
189 MS_EXCEPTION_IF_NULL(address);
190 address->set_ref_count(address->ref_count() + 1);
191 }
192 }
193 }
194
DecreaseAddressRefCount(const AnfNodePtr & kernel)195 void CPUMemoryManager::DecreaseAddressRefCount(const AnfNodePtr &kernel) {
196 if (!dynamic_malloc_) {
197 return;
198 }
199 MS_EXCEPTION_IF_NULL(kernel);
200 size_t input_num = common::AnfAlgo::GetInputTensorNum(kernel);
201 for (size_t i = 0; i < input_num; ++i) {
202 auto address = AnfAlgo::GetPrevNodeMutableOutputAddr(kernel, i);
203 MS_EXCEPTION_IF_NULL(address);
204 address->DecreaseRefCount();
205 if (address->ref_count() == 0 && address->GetDevicePtr() != nullptr) {
206 MemFree(address->GetDevicePtr());
207 address->SetDevicePtr(nullptr);
208 }
209 }
210 auto kernel_mod = AnfAlgo::GetKernelMod(kernel);
211 MS_EXCEPTION_IF_NULL(kernel_mod);
212 for (size_t i = 0; i < kernel_mod->GetWorkspaceSizeList().size(); ++i) {
213 auto address = AnfAlgo::GetWorkspaceAddr(kernel, i);
214 MS_EXCEPTION_IF_NULL(address);
215 address->DecreaseRefCount();
216 if (address->ref_count() == 0 && address->GetDevicePtr() != nullptr) {
217 MemFree(address->GetDevicePtr());
218 address->SetDevicePtr(nullptr);
219 }
220 }
221 }
222 } // namespace cpu
223 } // namespace device
224 } // namespace mindspore
225