1 // Copyright 2020 Google LLC
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 #include <assert.h>
7 #include <stdbool.h>
8 #include <stdint.h>
9 #include <stdlib.h>
10
11 #include <xnnpack/memory-planner.h>
12 #include <xnnpack/subgraph.h>
13
14 // Check if two xnn_value's lifecycles overlap.
value_lifecycle_overlap(const struct xnn_value_usage * a,const struct xnn_value_usage * b)15 inline static bool value_lifecycle_overlap(const struct xnn_value_usage* a, const struct xnn_value_usage* b) {
16 assert(a->last_node >= a->first_node);
17 assert(b->last_node >= b->first_node);
18 if (a->first_node < b->first_node) {
19 return a->last_node >= b->first_node;
20 } else {
21 return b->last_node >= a->first_node;
22 }
23 }
24
25 // Use this comparison function to sort xnn_value_usage according to the
26 // tensor_size in decreasing order.
cmp_value_usage_tensor_size(const void * a,const void * b)27 static inline int cmp_value_usage_tensor_size(const void* a, const void* b) {
28 const size_t tensor_size_a = (*(struct xnn_value_usage**)a)->tensor_size;
29 const size_t tensor_size_b = (*(struct xnn_value_usage**)b)->tensor_size;
30 return (tensor_size_b > tensor_size_a) - (tensor_size_b < tensor_size_a);
31 }
32
populate_value_lifecycle(const xnn_subgraph_t subgraph,struct xnn_value_usage * usage)33 static void populate_value_lifecycle(const xnn_subgraph_t subgraph, struct xnn_value_usage* usage) {
34 assert(subgraph != NULL);
35 if (subgraph->num_nodes == 0) {
36 return;
37 }
38 // As we initialized first/last_node in each xnn_value_usage to 0 as in 'xnn_init_value_mem_allocation_tracker',
39 // we start with the second node to tell whether first/last_node have been set or not, and check the first node last.
40 for (uint32_t nid = 1; nid < subgraph->num_nodes; ++nid) {
41 const struct xnn_node* node = subgraph->nodes + nid;
42 for (uint32_t i = 0; i < node->num_inputs; ++i) {
43 if (usage[node->inputs[i]].first_node == 0) {
44 usage[node->inputs[i]].first_node = nid;
45 }
46 usage[node->inputs[i]].last_node = nid;
47 }
48 for (uint32_t i = 0; i < node->num_outputs; ++i) {
49 if (usage[node->outputs[i]].first_node == 0) {
50 usage[node->outputs[i]].first_node = nid;
51 }
52 usage[node->outputs[i]].last_node = nid;
53 }
54 }
55 const struct xnn_node* first_node = subgraph->nodes;
56 for (uint32_t i = 0; i < first_node->num_inputs; ++i) {
57 usage[first_node->inputs[i]].first_node = 0;
58 }
59 for (uint32_t i = 0; i < first_node->num_outputs; ++i) {
60 usage[first_node->outputs[i]].first_node = 0;
61 }
62 }
63
64 // Represent a memory block [start, end)
65 struct memory_block {
66 size_t start;
67 size_t end;
68 };
69
70 // Use this comparison function to sort memory_block according to the 'start'
71 // in increasing order.
cmp_memory_block(const void * a,const void * b)72 static inline int cmp_memory_block(const void* a, const void* b) {
73 const size_t start_a = ((struct memory_block*)a)->start;
74 const size_t start_b = ((struct memory_block*)b)->start;
75 return (start_a > start_b) - (start_a < start_b);
76 }
77
78 // Given the current live memory blocks, return the offset in a memory arena for a to-be-allocated value of size
79 // 'to_alloc_size'.
find_value_alloc_offset(struct memory_block * live_mem_blocks,size_t num_mem_blocks,size_t to_alloc_size)80 static size_t find_value_alloc_offset(struct memory_block* live_mem_blocks,
81 size_t num_mem_blocks,
82 size_t to_alloc_size) {
83 if (num_mem_blocks == 0) {
84 return 0;
85 }
86
87 if (num_mem_blocks == 1) {
88 return live_mem_blocks[0].end;
89 }
90
91 // Sort memory blocks according to 'start' in increasing order.
92 qsort(live_mem_blocks, num_mem_blocks, sizeof(struct memory_block), cmp_memory_block);
93
94 // Coalesce overlapping or immediate adjacent memory blocks to form a list of non-overlapping memory blocks in order
95 // to find the smallest gap.
96 size_t num_coalesced_mem_blocks = 1;
97 for (size_t i = 1; i < num_mem_blocks; ++i) {
98 const size_t current_coalesced_end =
99 live_mem_blocks[num_coalesced_mem_blocks - 1].end;
100 if (live_mem_blocks[i].start > current_coalesced_end) {
101 assert(num_coalesced_mem_blocks <= i);
102 live_mem_blocks[num_coalesced_mem_blocks] = live_mem_blocks[i];
103 num_coalesced_mem_blocks++;
104 continue;
105 }
106 if (live_mem_blocks[i].end > current_coalesced_end) {
107 live_mem_blocks[num_coalesced_mem_blocks - 1].end = live_mem_blocks[i].end;
108 }
109 }
110
111 size_t smallest_gap_size = SIZE_MAX;
112 // The first index to live_mem_blocks that the 'to_alloc_size' should be allocated after.
113 size_t smallest_gap_index = num_coalesced_mem_blocks - 1;
114 for (size_t i = 0; i < num_coalesced_mem_blocks - 1; ++i) {
115 assert(live_mem_blocks[i + 1].start > live_mem_blocks[i].end);
116 const size_t gap = live_mem_blocks[i + 1].start - live_mem_blocks[i].end;
117 if (gap >= to_alloc_size && gap < smallest_gap_size) {
118 smallest_gap_index = i;
119 smallest_gap_size = gap;
120 }
121 }
122 return live_mem_blocks[smallest_gap_index].end;
123 }
124
xnn_init_value_allocation_tracker(struct xnn_value_allocation_tracker * tracker,const xnn_subgraph_t subgraph)125 void xnn_init_value_allocation_tracker(struct xnn_value_allocation_tracker* tracker, const xnn_subgraph_t subgraph) {
126 tracker->subgraph = subgraph;
127 tracker->mem_arena_size = 0;
128 tracker->usage = xnn_allocate_zero_memory(sizeof(struct xnn_value_usage) * subgraph->num_values);
129 #if XNN_ENABLE_MEMOPT
130 populate_value_lifecycle(tracker->subgraph, tracker->usage);
131 #endif
132 tracker->min_value_id = XNN_INVALID_VALUE_ID;
133 tracker->max_value_id = XNN_INVALID_VALUE_ID;
134 }
135
xnn_add_value_allocation_tracker(struct xnn_value_allocation_tracker * tracker,uint32_t value_id,size_t tensor_size)136 void xnn_add_value_allocation_tracker(struct xnn_value_allocation_tracker* tracker,
137 uint32_t value_id,
138 size_t tensor_size) {
139 tracker->usage[value_id].tensor_size = tensor_size;
140 if (tracker->min_value_id == XNN_INVALID_VALUE_ID) {
141 tracker->min_value_id = value_id;
142 } else {
143 // Note that values are expected to be added in increasing order.
144 assert(value_id > tracker->min_value_id);
145 assert(value_id > tracker->max_value_id);
146 }
147
148 tracker->max_value_id = value_id;
149 }
150
xnn_plan_value_allocation_tracker(struct xnn_value_allocation_tracker * tracker)151 void xnn_plan_value_allocation_tracker(struct xnn_value_allocation_tracker* tracker) {
152 #if XNN_ENABLE_MEMOPT
153 if (tracker->min_value_id == XNN_INVALID_VALUE_ID) {
154 assert(tracker->max_value_id == XNN_INVALID_VALUE_ID);
155 return;
156 }
157
158 const uint32_t num_values = tracker->max_value_id - tracker->min_value_id + 1;
159 struct xnn_value_usage** sorted_usage = xnn_allocate_zero_memory(sizeof(struct xnn_value_usage*) * num_values);
160 size_t num_values_to_alloc = 0;
161 for (size_t i = tracker->min_value_id; i <= tracker->max_value_id; ++i) {
162 struct xnn_value_usage* info = tracker->usage + i;
163 if (info->tensor_size != 0) {
164 sorted_usage[num_values_to_alloc++] = info;
165 }
166 }
167 qsort(sorted_usage, num_values_to_alloc, sizeof(struct xnn_value_usage*), cmp_value_usage_tensor_size);
168
169 // Start the allocation planning process.
170 struct memory_block* current_live_mem_blocks = xnn_allocate_zero_memory(
171 sizeof(struct memory_block) * num_values_to_alloc);
172 size_t mem_arena_size = 0;
173 for (size_t i = 0; i < num_values_to_alloc; ++i) {
174 size_t num_live_mem_blocks = 0;
175 struct xnn_value_usage* current = sorted_usage[i];
176 for (size_t j = 0; j < i; ++j) {
177 const struct xnn_value_usage* allocated = sorted_usage[j];
178 if (value_lifecycle_overlap(current, allocated)) {
179 current_live_mem_blocks[num_live_mem_blocks++] = (struct memory_block){
180 .start = allocated->alloc_offset,
181 .end = allocated->alloc_offset + allocated->tensor_size,
182 };
183 }
184 }
185 current->alloc_offset = find_value_alloc_offset(current_live_mem_blocks, num_live_mem_blocks, current->tensor_size);
186 if (mem_arena_size < current->alloc_offset + current->tensor_size) {
187 mem_arena_size = current->alloc_offset + current->tensor_size;
188 }
189 }
190
191 tracker->mem_arena_size = mem_arena_size;
192 xnn_release_memory(sorted_usage);
193 xnn_release_memory(current_live_mem_blocks);
194 #else
195 tracker->mem_arena_size = 0;
196 for (uint32_t i = tracker->min_value_id; i <= tracker->max_value_id; ++i) {
197 if (tracker->usage[i].tensor_size > 0) {
198 tracker->usage[i].alloc_offset = tracker->mem_arena_size;
199 tracker->mem_arena_size += tracker->usage[i].tensor_size;
200 }
201 }
202 #endif
203 }
204