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/rendezvous_mgr.h"
17
18 #include <unordered_set>
19
20 #include "tensorflow/core/common_runtime/copy_tensor.h"
21 #include "tensorflow/core/common_runtime/device.h"
22 #include "tensorflow/core/common_runtime/device_mgr.h"
23 #include "tensorflow/core/framework/allocator.h"
24 #include "tensorflow/core/framework/device_factory.h"
25 #include "tensorflow/core/framework/types.h"
26 #include "tensorflow/core/lib/core/errors.h"
27 #include "tensorflow/core/lib/core/notification.h"
28 #include "tensorflow/core/lib/strings/numbers.h"
29 #include "tensorflow/core/lib/strings/str_util.h"
30 #include "tensorflow/core/platform/logging.h"
31 #include "tensorflow/core/platform/mutex.h"
32 #include "tensorflow/core/platform/types.h"
33
34 namespace tensorflow {
35
36 namespace {
SameWorkerRecvDone(const DeviceMgr * device_mgr,const Rendezvous::ParsedKey & parsed,const Rendezvous::Args & send_args,const Rendezvous::Args & recv_args,const Tensor & in,Tensor * out,StatusCallback done)37 void SameWorkerRecvDone(const DeviceMgr* device_mgr,
38 const Rendezvous::ParsedKey& parsed,
39 const Rendezvous::Args& send_args,
40 const Rendezvous::Args& recv_args, const Tensor& in,
41 Tensor* out, StatusCallback done) {
42 // Do a quick copy (sharing the underlying buffer) if both tensors
43 // are on host memory.
44 const bool src_host =
45 (send_args.alloc_attrs.on_host() || parsed.src.type == "CPU");
46 const bool dst_host =
47 (recv_args.alloc_attrs.on_host() || parsed.dst.type == "CPU");
48 if (src_host && dst_host) {
49 *out = in;
50 done(Status::OK());
51 return;
52 }
53
54 // This copy must involve a non-CPU device. Hence, "in" must support DMA
55 // (e.g., string tensors do not work on GPU). Variant copy DMA
56 // checks happen inside CopyTensor::ViaDMA.
57 if (!DataTypeCanUseMemcpy(in.dtype()) && in.dtype() != DT_VARIANT &&
58 in.dtype() != DT_RESOURCE) {
59 done(errors::InvalidArgument(
60 "Non-DMA-safe ", DataTypeString(in.dtype()),
61 " tensor may not be copied from/to a device. Key: ", parsed.FullKey()));
62 return;
63 }
64
65 Device* src_device;
66 Status s = device_mgr->LookupDevice(parsed.src_device, &src_device);
67 if (!s.ok()) {
68 done(s);
69 return;
70 }
71 Device* dst_device;
72 s = device_mgr->LookupDevice(parsed.dst_device, &dst_device);
73 if (!s.ok()) {
74 done(s);
75 return;
76 }
77
78 ScopedMemoryDebugAnnotation op_annotation("SameWorkerRecvDone", 0, "dynamic",
79 in.dtype(), &in.shape());
80 AllocatorAttributes attr = recv_args.alloc_attrs;
81 attr.set_gpu_compatible(send_args.alloc_attrs.gpu_compatible() ||
82 recv_args.alloc_attrs.gpu_compatible());
83 Allocator* out_allocator = dst_device->GetAllocator(attr);
84 bool sync_dst_compute = true;
85 if (in.dtype() != DT_VARIANT) {
86 // Variants are handled by CopyTensor::ViaDMA.
87 AllocationAttributes aa;
88 uint64 safe_alloc_frontier = dst_device->SafeAllocFrontier(0);
89 std::function<uint64()> freed_by_func = [dst_device,
90 &safe_alloc_frontier]() {
91 safe_alloc_frontier = dst_device->SafeAllocFrontier(safe_alloc_frontier);
92 return safe_alloc_frontier;
93 };
94 if ((parsed.dst.type == "GPU" ||
95 DeviceFactory::IsPluggableDevice(parsed.dst.type)) &&
96 safe_alloc_frontier > 0) {
97 // There's a timestamped allocator at work, so use it instead
98 // of sync_dst_compute.
99 aa.freed_by_func = &freed_by_func;
100 sync_dst_compute = false;
101 }
102 Tensor copy(out_allocator, in.dtype(), in.shape(), aa);
103 *out = copy;
104 if (in.shape().num_elements() > 0 && out->data() == nullptr) {
105 done(tensorflow::errors::ResourceExhausted(
106 "SameWorkerRecvDone unable to allocate output tensor. Key: ",
107 parsed.FullKey()));
108 return;
109 }
110 }
111
112 CopyTensor::ViaDMA(
113 parsed.edge_name, send_args.device_context, recv_args.device_context,
114 src_device, dst_device, send_args.alloc_attrs, recv_args.alloc_attrs, &in,
115 out, 0 /*dev_to_dev_stream_index*/, std::move(done), sync_dst_compute);
116 }
117
IntraProcessRecvAsyncImpl(const DeviceMgr * device_mgr,LocalRendezvous * local,const RendezvousInterface::ParsedKey & parsed,const Rendezvous::Args & recv_args,RendezvousInterface::DoneCallback done)118 void IntraProcessRecvAsyncImpl(const DeviceMgr* device_mgr,
119 LocalRendezvous* local,
120 const RendezvousInterface::ParsedKey& parsed,
121 const Rendezvous::Args& recv_args,
122 RendezvousInterface::DoneCallback done) {
123 VLOG(1) << "IntraProcessRendezvous Recv " << local << " " << parsed.FullKey();
124
125 ScopedMemoryDebugAnnotation op_annotation("RecvAsync");
126 // Recv the tensor from local_.
127 local->RecvAsync(
128 parsed, recv_args,
129 [device_mgr, parsed, done = std::move(done)](
130 const Status& status, const Rendezvous::Args& send_args,
131 const Rendezvous::Args& recv_args, const Tensor& in,
132 bool is_dead) mutable {
133 // If "in" is an uninitialized tensor, do copy-construction to
134 // preserve the uninitialized state, along with data type and shape
135 // info, which is useful for debugger purposes.
136 Tensor* out = in.IsInitialized() ? new Tensor : new Tensor(in);
137
138 auto final_callback = [send_args, recv_args, out, is_dead,
139 done = std::move(done)](const Status& s) {
140 done(s, send_args, recv_args, *out, is_dead);
141 delete out;
142 };
143
144 if (status.ok() && in.IsInitialized()) {
145 SameWorkerRecvDone(device_mgr, parsed, send_args, recv_args, in, out,
146 std::move(final_callback));
147 } else {
148 final_callback(status);
149 }
150 });
151 }
152
153 } // namespace
154
RefCountedIntraProcessRendezvous(const DeviceMgr * device_mgr)155 RefCountedIntraProcessRendezvous::RefCountedIntraProcessRendezvous(
156 const DeviceMgr* device_mgr)
157 : device_mgr_(device_mgr), local_(this) {}
158
~RefCountedIntraProcessRendezvous()159 RefCountedIntraProcessRendezvous::~RefCountedIntraProcessRendezvous() {}
160
Send(const ParsedKey & key,const Rendezvous::Args & args,const Tensor & val,const bool is_dead)161 Status RefCountedIntraProcessRendezvous::Send(const ParsedKey& key,
162 const Rendezvous::Args& args,
163 const Tensor& val,
164 const bool is_dead) {
165 VLOG(1) << "IntraProcessRendezvous Send " << this << " " << key.FullKey();
166 return local_.Send(key, args, val, is_dead);
167 }
168
RecvAsync(const ParsedKey & key,const Rendezvous::Args & args,DoneCallback done)169 void RefCountedIntraProcessRendezvous::RecvAsync(const ParsedKey& key,
170 const Rendezvous::Args& args,
171 DoneCallback done) {
172 VLOG(1) << "IntraProcessRendezvous Recv " << this << " " << key.FullKey();
173 IntraProcessRecvAsyncImpl(device_mgr_, &local_, key, args, std::move(done));
174 }
175
StartAbort(const Status & s)176 void RefCountedIntraProcessRendezvous::StartAbort(const Status& s) {
177 local_.StartAbort(s);
178 }
179
PrivateIntraProcessRendezvous(const DeviceMgr * device_mgr)180 PrivateIntraProcessRendezvous::PrivateIntraProcessRendezvous(
181 const DeviceMgr* device_mgr)
182 : device_mgr_(device_mgr), local_(nullptr) {}
183
~PrivateIntraProcessRendezvous()184 PrivateIntraProcessRendezvous::~PrivateIntraProcessRendezvous() {}
185
Send(const ParsedKey & key,const Rendezvous::Args & args,const Tensor & val,const bool is_dead)186 Status PrivateIntraProcessRendezvous::Send(const ParsedKey& key,
187 const Rendezvous::Args& args,
188 const Tensor& val,
189 const bool is_dead) {
190 DVLOG(1) << "IntraProcessRendezvous Send " << this << " " << key.FullKey();
191 return local_.Send(key, args, val, is_dead);
192 }
193
RecvAsync(const ParsedKey & key,const Rendezvous::Args & args,DoneCallback done)194 void PrivateIntraProcessRendezvous::RecvAsync(const ParsedKey& key,
195 const Rendezvous::Args& args,
196 DoneCallback done) {
197 DVLOG(1) << "StackAllocatedIntraProcessRendezvous Recv " << this << " "
198 << key.FullKey();
199 IntraProcessRecvAsyncImpl(device_mgr_, &local_, key, args, std::move(done));
200 }
201
StartAbort(const Status & s)202 void PrivateIntraProcessRendezvous::StartAbort(const Status& s) {
203 local_.StartAbort(s);
204 }
205
206 } // end namespace tensorflow
207