1 | /* Copyright 2015 The TensorFlow Authors. All Rights Reserved. |
2 | |
3 | Licensed under the Apache License, Version 2.0 (the "License"); |
4 | you may not use this file except in compliance with the License. |
5 | You may obtain a copy of the License at |
6 | |
7 | http://www.apache.org/licenses/LICENSE-2.0 |
8 | |
9 | Unless required by applicable law or agreed to in writing, software |
10 | distributed under the License is distributed on an "AS IS" BASIS, |
11 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
12 | See the License for the specific language governing permissions and |
13 | limitations under the License. |
14 | ==============================================================================*/ |
15 | |
16 | #include "tensorflow/core/common_runtime/graph_view.h" |
17 | |
18 | #include <atomic> |
19 | #include <deque> |
20 | #include <memory> |
21 | #include <string> |
22 | #include <unordered_map> |
23 | #include <vector> |
24 | |
25 | #include "tensorflow/core/common_runtime/device.h" |
26 | #include "tensorflow/core/framework/node_def_util.h" |
27 | #include "tensorflow/core/framework/op_kernel.h" |
28 | #include "tensorflow/core/framework/tensor.h" |
29 | #include "tensorflow/core/graph/algorithm.h" |
30 | #include "tensorflow/core/graph/edgeset.h" |
31 | #include "tensorflow/core/graph/graph.h" |
32 | #include "tensorflow/core/lib/core/errors.h" |
33 | #include "tensorflow/core/lib/gtl/inlined_vector.h" |
34 | #include "tensorflow/core/lib/strings/str_util.h" |
35 | #include "tensorflow/core/util/determinism.h" |
36 | #include "tensorflow/core/util/device_name_utils.h" |
37 | |
38 | namespace tensorflow { |
39 | |
40 | string NodeItem::DebugString() const { |
41 | string ret = strings::StrCat("{name:'" , kernel->name(), "' id:" , node_id); |
42 | if (is_source) { |
43 | strings::StrAppend(&ret, " source}" ); |
44 | } else { |
45 | strings::StrAppend(&ret, " def:{" , SummarizeNodeDef(kernel->def()), "}}" ); |
46 | } |
47 | return ret; |
48 | } |
49 | |
50 | GraphView::~GraphView() { |
51 | static_assert(std::is_trivially_destructible<AllocatorAttributes>::value, |
52 | "Update code if AllocatorAttributes gains a destructor" ); |
53 | static_assert(std::is_trivially_destructible<EdgeInfo>::value, |
54 | "Update code if EdgeInfo gains a destructor" ); |
55 | for (int i = 0; i < num_nodes_; i++) { |
56 | NodeItem* n = node(i); |
57 | if (n != nullptr) { |
58 | n->NodeItem::~NodeItem(); |
59 | // Memory for "n" itself is held in space_ & gets cleaned up below |
60 | } |
61 | } |
62 | delete[] node_offsets_; |
63 | delete[] space_; |
64 | } |
65 | |
66 | namespace { |
67 | typedef std::tuple<int32, int32> OutputAndControlEdges; |
68 | |
69 | OutputAndControlEdges CountOutputEdges(const Node* n) { |
70 | DCHECK_LE(n->out_edges().size(), kint32max); |
71 | int32_t num_output_edges = 0; |
72 | int32_t num_output_control_edges = 0; |
73 | for (auto e : n->out_edges()) { |
74 | if (IsSink(e->dst())) continue; |
75 | if (e->IsControlEdge()) { |
76 | ++num_output_control_edges; |
77 | } else { |
78 | ++num_output_edges; |
79 | } |
80 | } |
81 | return OutputAndControlEdges(num_output_edges, num_output_control_edges); |
82 | } |
83 | } // namespace |
84 | |
85 | size_t GraphView::NodeItemBytes(const Node* n) { |
86 | int32_t num_output_edges; |
87 | int32_t num_output_control_edges; |
88 | std::tie(num_output_edges, num_output_control_edges) = CountOutputEdges(n); |
89 | const int num_inputs = n->num_inputs(); |
90 | const int num_outputs = n->num_outputs(); |
91 | |
92 | // Compute number of bytes needed for NodeItem and variable length data. |
93 | // We do not subtract sizeof(var) since num_inputs/num_outputs might |
94 | // both be zero. |
95 | const size_t raw_bytes = |
96 | sizeof(NodeItem) // Fixed |
97 | + num_output_edges * sizeof(EdgeInfo) // output_edges[...] |
98 | + num_output_control_edges * // |
99 | sizeof(ControlEdgeInfo) // output_control_edges[...] |
100 | + num_outputs * sizeof(AllocatorAttributes) // output_attr[...] |
101 | + num_outputs * sizeof(int) // forward_from[num_outputs] |
102 | + num_inputs * sizeof(uint8) // input_type[num_inputs] |
103 | + num_outputs * sizeof(uint8); // output_type[num_outputs] |
104 | static constexpr size_t kItemAlignment = sizeof(NodeItem*); |
105 | static_assert(kItemAlignment % alignof(NodeItem) == 0, |
106 | "NodeItem must be aligned with kItemAlignment" ); |
107 | static_assert(kItemAlignment % alignof(EdgeInfo) == 0, |
108 | "EdgeInfo must be aligned with kItemAlignment" ); |
109 | static_assert(kItemAlignment % alignof(ControlEdgeInfo) == 0, |
110 | "ControlEdgeInfo must be aligned with kItemAlignment" ); |
111 | static_assert(kItemAlignment % alignof(AllocatorAttributes) == 0, |
112 | "AllocatorAttributes must be aligned with kItemAlignment" ); |
113 | static_assert(sizeof(NodeItem) % alignof(EdgeInfo) == 0, |
114 | "NodeItem must be aligned with EdgeInfo" ); |
115 | static_assert(sizeof(NodeItem) % alignof(AllocatorAttributes) == 0, |
116 | "NodeItem must be aligned with AllocatorAttributes" ); |
117 | static_assert(sizeof(EdgeInfo) % alignof(AllocatorAttributes) == 0, |
118 | "EdgeInfo must be aligned with AllocatorAttributes" ); |
119 | const size_t bytes = |
120 | ((raw_bytes + kItemAlignment - 1) / kItemAlignment) * kItemAlignment; |
121 | return bytes; |
122 | } |
123 | |
124 | char* GraphView::InitializeNode(char* ptr, const Node* n) { |
125 | const int id = n->id(); |
126 | CHECK(node_offsets_[id] == kuint32max); // Initial value in constructor |
127 | |
128 | const size_t bytes = NodeItemBytes(n); |
129 | constexpr size_t kItemAlignment = sizeof(NodeItem*); |
130 | CHECK_EQ(reinterpret_cast<uintptr_t>(ptr) % kItemAlignment, 0); |
131 | NodeItem* item = reinterpret_cast<NodeItem*>(ptr); |
132 | |
133 | // We store a 32-bit offset relative to the beginning of space_, so that we |
134 | // only need an array of 32-bit values to map from node id to the NodeItem*, |
135 | // (versus 64 bits on most machines if we just stored an array of NodeItem* |
136 | // pointers). Casting to int64 is needed on 32bit CPU to avoid comparing |
137 | // values as "int" vs "size_t" in CHECK_LE. |
138 | CHECK_LE(static_cast<int64_t>(ptr - space_), kuint32max); |
139 | const uint32 offset = static_cast<uint32>(ptr - space_); |
140 | node_offsets_[id] = offset; |
141 | ptr += bytes; |
142 | |
143 | int32_t num_output_edges; |
144 | int32_t num_output_control_edges; |
145 | std::tie(num_output_edges, num_output_control_edges) = CountOutputEdges(n); |
146 | const int num_inputs = n->num_inputs(); |
147 | const int num_outputs = n->num_outputs(); |
148 | |
149 | new (item) NodeItem(); |
150 | item->num_inputs = num_inputs; |
151 | item->num_outputs = num_outputs; |
152 | item->num_output_edges = num_output_edges; |
153 | item->num_output_control_edges = num_output_control_edges; |
154 | |
155 | // Fill output edges. |
156 | // Keep track of the last EdgeInfo in the EdgeInfo array that references |
157 | // a given output slot. For all but the last, we need to do a copy of the |
158 | // Tensor when propagating results downstream in the graph, but for the |
159 | // last one, we can just do a move of the Tensor object to propagate it. |
160 | gtl::InlinedVector<EdgeInfo*, 4> last_indices(num_outputs, nullptr); |
161 | EdgeInfo* dst_edge = item->output_edge_base(); |
162 | for (auto e : n->out_edges()) { |
163 | if (e->IsControlEdge()) continue; |
164 | dst_edge->dst_id = e->dst()->id(); |
165 | CHECK_LE(e->src_output(), 0x3FFFFFFF); // Must fit in 31 bits |
166 | dst_edge->output_slot = e->src_output(); |
167 | dst_edge->is_last = false; |
168 | const int output_slot = dst_edge->output_slot; |
169 | if (output_slot >= 0) { |
170 | last_indices[output_slot] = dst_edge; |
171 | } |
172 | // NOTE: The `input_slot` will be rewritten to the frame-wide offset later |
173 | // in `ExecutorImpl::Initialize()`. |
174 | dst_edge->input_slot = e->dst_input(); |
175 | dst_edge++; |
176 | } |
177 | for (EdgeInfo* edge_info : last_indices) { |
178 | if (edge_info != nullptr) { |
179 | edge_info->is_last = true; |
180 | } |
181 | } |
182 | ControlEdgeInfo* dst_control_edge = item->output_control_edge_base(); |
183 | for (auto e : n->out_edges()) { |
184 | if (!e->IsControlEdge() || IsSink(e->dst())) continue; |
185 | dst_control_edge->dst_id = e->dst()->id(); |
186 | dst_control_edge++; |
187 | } |
188 | |
189 | AllocatorAttributes* output_attrs = item->output_attr_base(); |
190 | for (int i = 0; i < num_outputs; i++) { |
191 | new (&output_attrs[i]) AllocatorAttributes(); |
192 | } |
193 | |
194 | DCHECK_LT(DataType_MAX, 255); // Must fit in uint8 |
195 | uint8* input_types = item->input_type_base(); |
196 | item->is_any_input_ref_typed = false; |
197 | for (int i = 0; i < num_inputs; i++) { |
198 | input_types[i] = static_cast<uint8>(n->input_type(i)); |
199 | DCHECK_EQ(item->input_type(i), n->input_type(i)); |
200 | item->is_any_input_ref_typed |= IsRefType(n->input_type(i)); |
201 | } |
202 | |
203 | // Check ScopedAllocatorAttrs and forward_from. Also assign output_types. |
204 | { |
205 | std::vector<int> forward_input; |
206 | Status fwd_status = |
207 | GetNodeAttr(n->attrs(), "_forward_input" , &forward_input); |
208 | std::vector<int> scoped_allocator_attrs; |
209 | Status sa_status = |
210 | GetNodeAttr(n->attrs(), "_scoped_allocator" , &scoped_allocator_attrs); |
211 | |
212 | int* forward_from = item->forward_from_base(); |
213 | uint8* output_types = item->output_type_base(); |
214 | for (int i = 0; i < num_outputs; ++i) { |
215 | output_types[i] = static_cast<uint8>(n->output_type(i)); |
216 | DCHECK_EQ(item->output_type(i), n->output_type(i)); |
217 | |
218 | forward_from[i] = OpKernelContext::Params::kNoReservation; |
219 | if (sa_status.ok()) { |
220 | for (int j = 0; j < scoped_allocator_attrs.size(); j += 2) { |
221 | if (scoped_allocator_attrs[j] == i) { |
222 | // This output slot must be explicitly allocated from a |
223 | // ScopedAllocator. |
224 | forward_from[i] = OpKernelContext::Params::kNeverForward; |
225 | DCHECK_EQ(output_attrs[i].scope_id, 0); |
226 | output_attrs[i].scope_id = scoped_allocator_attrs[j + 1]; |
227 | } |
228 | } |
229 | } |
230 | if (fwd_status.ok() && |
231 | forward_from[i] == OpKernelContext::Params::kNoReservation) { |
232 | DCHECK_EQ(forward_input.size() % 2, 0); |
233 | for (int j = 0; j < forward_input.size(); j += 2) { |
234 | if (forward_input[j + 1] == i) { |
235 | DCHECK_EQ(forward_from[i], OpKernelContext::Params::kNoReservation); |
236 | forward_from[i] = forward_input[j]; |
237 | break; |
238 | } |
239 | } |
240 | } |
241 | } |
242 | } |
243 | |
244 | return ptr; |
245 | } |
246 | |
247 | Status GraphView::Initialize(const Graph* g) { |
248 | CHECK(node_offsets_ == nullptr); |
249 | const int num_nodes = g->num_node_ids(); |
250 | num_nodes_ = num_nodes; |
251 | size_t total_bytes = 0; |
252 | for (const Node* n : g->nodes()) { |
253 | if (n->out_edges().size() > kint32max) { |
254 | return errors::InvalidArgument( |
255 | "The executor cannot handle nodes with more than " , kint32max, |
256 | " output edges. Node " , n->name(), " had " , n->out_edges().size(), |
257 | " output edges." ); |
258 | } |
259 | total_bytes += NodeItemBytes(n); |
260 | } |
261 | |
262 | node_offsets_ = new uint32[num_nodes]; |
263 | for (int i = 0; i < num_nodes; i++) { |
264 | node_offsets_[i] = kuint32max; |
265 | } |
266 | |
267 | space_ = new char[total_bytes]; // NodeItem objects are allocated here |
268 | char* ptr = space_; |
269 | auto it = g->nodes(); |
270 | if (OpOrderDeterminismRequired()) { |
271 | // For OpOrder determinism, we need node_id's to be stable across runs. We |
272 | // assign node_ids in the order in which `InitializeNode` is called on each |
273 | // node. However, `g` exposes a NodeIter of nodes, which does not guarantee |
274 | // a deterministic ordering across runs. Since NodeIter is immutable, we |
275 | // must sort a local copy. We sort by node_name, which is set in the |
276 | // GraphDef, so must be stable across runs. |
277 | std::vector<Node*> nodes(it.begin(), it.end()); |
278 | std::sort(nodes.begin(), nodes.end(), NodeComparatorName()); |
279 | for (const Node* n : nodes) { |
280 | ptr = InitializeNode(ptr, n); |
281 | } |
282 | } else { |
283 | for (const Node* n : it) { |
284 | ptr = InitializeNode(ptr, n); |
285 | } |
286 | } |
287 | CHECK_EQ(ptr, space_ + total_bytes); |
288 | return OkStatus(); |
289 | } |
290 | |
291 | namespace { |
292 | // If a Node has been marked to use a ScopedAllocator x for output i, then |
293 | // sc_attr will contain the subsequence (i, x) at an even offset. This function |
294 | // extracts and transfers that ScopedAllocator id to alloc_attr. For now, we |
295 | // only allow one ScopedAllocator use per Node. |
296 | bool (const std::vector<int>& sc_attr, |
297 | int output_index, |
298 | AllocatorAttributes* alloc_attr) { |
299 | DCHECK_LE(2, sc_attr.size()); |
300 | for (int i = 0; i < sc_attr.size(); i += 2) { |
301 | if (sc_attr[i] == output_index) { |
302 | CHECK_EQ(alloc_attr->scope_id, 0); |
303 | alloc_attr->scope_id = sc_attr[i + 1]; |
304 | return true; |
305 | } |
306 | } |
307 | return false; |
308 | } |
309 | } // namespace |
310 | |
311 | void GraphView::SetScopedAllocatorAttrs( |
312 | const std::vector<const Node*>& sa_nodes) { |
313 | for (const Node* sa : sa_nodes) { |
314 | NodeItem* sa_item = node(sa->id()); |
315 | AllocatorAttributes* sa_attrs = sa_item->output_attr_base(); |
316 | // Control edges out of the ScopedAllocator should be use instances, but may |
317 | // include a few other nodes. |
318 | for (const auto& e : sa->out_edges()) { |
319 | if (IsSink(e->dst()) || !e->IsControlEdge()) { |
320 | continue; |
321 | } |
322 | Node* use_node = e->dst(); |
323 | NodeItem* item = node(use_node->id()); |
324 | AllocatorAttributes* use_attrs = item->output_attr_base(); |
325 | std::vector<int> scoped_allocator_attrs; |
326 | Status s = GetNodeAttr(use_node->attrs(), "_scoped_allocator" , |
327 | &scoped_allocator_attrs); |
328 | if (!s.ok()) { |
329 | VLOG(2) << "Failed to find expected ScopedAllocator attr on " |
330 | << use_node->name(); |
331 | continue; |
332 | } |
333 | // There can be more than one output using ScopedAllocation, but this |
334 | // analysis assumes they use the same ScopedAllocator. |
335 | for (const auto& e : use_node->out_edges()) { |
336 | if (IsSink(e->dst()) || !e->IsControlEdge()) { |
337 | AllocatorAttributes attr; |
338 | if (ExtractScopedAllocatorAttr(scoped_allocator_attrs, |
339 | e->src_output(), &attr)) { |
340 | // Set the scope_id on this use instance node. |
341 | (use_attrs + e->src_output())->Merge(attr); |
342 | // Propagate the other attributes of this node back to the SA node. |
343 | attr = *(use_attrs + e->src_output()); |
344 | attr.scope_id = 0; |
345 | sa_attrs->Merge(attr); |
346 | } |
347 | } |
348 | } |
349 | } |
350 | } |
351 | } |
352 | |
353 | namespace { |
354 | Status InferAllocAttr(const Node* n, const Node* dst, |
355 | const DeviceNameUtils::ParsedName& local_dev_name, |
356 | AllocatorAttributes* attr) { |
357 | Status s; |
358 | // Note that it's possible for *n to be a Recv and *dst to be a Send, |
359 | // so these two cases are not mutually exclusive. |
360 | if (IsRecv(n)) { |
361 | string src_name; |
362 | s = GetNodeAttr(n->attrs(), "send_device" , &src_name); |
363 | if (!s.ok()) return s; |
364 | DeviceNameUtils::ParsedName parsed_src_name; |
365 | if (!DeviceNameUtils::ParseFullName(src_name, &parsed_src_name)) { |
366 | s = errors::Internal("Bad send_device attr '" , src_name, "' in node " , |
367 | n->name()); |
368 | return s; |
369 | } |
370 | if (!DeviceNameUtils::IsSameAddressSpace(parsed_src_name, local_dev_name)) { |
371 | // Value is going to be the sink of an RPC. |
372 | attr->set_nic_compatible(true); |
373 | VLOG(2) << "node " << n->name() << " is the sink of an RPC in" ; |
374 | } else if ((local_dev_name.type == "CPU" || n->IsHostRecv()) && |
375 | parsed_src_name.type != "CPU" ) { |
376 | // Value is going to be the sink of a local DMA from GPU to CPU (or |
377 | // other types of accelerators). |
378 | attr->set_gpu_compatible(true); |
379 | VLOG(2) << "node " << n->name() << " is the sink of a gpu->cpu copy" ; |
380 | } else { |
381 | VLOG(2) << "default alloc case local type " << local_dev_name.type |
382 | << " remote type " << parsed_src_name.type; |
383 | } |
384 | } |
385 | if (IsSend(dst)) { |
386 | string dst_name; |
387 | s = GetNodeAttr(dst->attrs(), "recv_device" , &dst_name); |
388 | if (!s.ok()) return s; |
389 | DeviceNameUtils::ParsedName parsed_dst_name; |
390 | if (!DeviceNameUtils::ParseFullName(dst_name, &parsed_dst_name)) { |
391 | s = errors::Internal("Bad recv_device attr '" , dst_name, "' in node " , |
392 | n->name()); |
393 | return s; |
394 | } |
395 | if (!DeviceNameUtils::IsSameAddressSpace(parsed_dst_name, local_dev_name)) { |
396 | // Value is going to be the source of an RPC. |
397 | attr->set_nic_compatible(true); |
398 | VLOG(2) << "node " << n->name() << " is the source of an RPC out" ; |
399 | } else if ((local_dev_name.type == "CPU" || dst->IsHostSend()) && |
400 | parsed_dst_name.type != "CPU" ) { |
401 | // Value is going to be the source of a local DMA from CPU to GPU (or |
402 | // other types of accelerators). |
403 | // Note that this does not cover the case where the allocation of the |
404 | // output tensor is not generated by the src: n. |
405 | attr->set_gpu_compatible(true); |
406 | VLOG(2) << "node " << n->name() << " is the source of a cpu->gpu copy" ; |
407 | } else { |
408 | VLOG(2) << "default alloc case local type " << local_dev_name.type |
409 | << " remote type " << parsed_dst_name.type; |
410 | } |
411 | } |
412 | if (n->IsCollective()) { |
413 | // We'll make the sweeping assumption that any collective op is going |
414 | // to be involved in network i/o. |
415 | attr->set_nic_compatible(true); |
416 | } |
417 | return s; |
418 | } |
419 | } // namespace |
420 | |
421 | Status GraphView::SetAllocAttrs(const Graph* g, const Device* device) { |
422 | Status s; |
423 | const DeviceNameUtils::ParsedName& local_dev_name = device->parsed_name(); |
424 | |
425 | std::vector<const Node*> scoped_allocator_instances; |
426 | for (const Node* n : g->nodes()) { |
427 | NodeItem* item = node(n->id()); |
428 | AllocatorAttributes* attrs = item->output_attr_base(); |
429 | if (IsScopedAllocator(n)) { |
430 | scoped_allocator_instances.push_back(n); |
431 | } |
432 | |
433 | // Examine the out edges of each node looking for special use |
434 | // cases that may affect memory allocation attributes. |
435 | for (const auto& e : n->out_edges()) { |
436 | if (!e->IsControlEdge()) { |
437 | AllocatorAttributes attr; |
438 | s = InferAllocAttr(n, e->dst(), local_dev_name, &attr); |
439 | if (!s.ok()) return s; |
440 | if (attr.value != 0 || attr.scope_id != 0) { |
441 | attrs[e->src_output()].Merge(attr); |
442 | } |
443 | } |
444 | } |
445 | |
446 | for (int out = 0; out < n->num_outputs(); out++) { |
447 | const OpKernel* op_kernel = item->kernel; |
448 | DCHECK_LT(out, op_kernel->output_memory_types().size()); |
449 | bool on_host = op_kernel->output_memory_types()[out] == HOST_MEMORY; |
450 | if (on_host) { |
451 | AllocatorAttributes h; |
452 | h.set_on_host(on_host); |
453 | attrs[out].Merge(h); |
454 | } |
455 | } |
456 | } |
457 | SetScopedAllocatorAttrs(scoped_allocator_instances); |
458 | return s; |
459 | } |
460 | |
461 | } // namespace tensorflow |
462 | |