• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2020 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 #include "tensorflow/core/common_runtime/permuter.h"
16 
17 #include "tensorflow/core/common_runtime/collective_rma_local.h"
18 #include "tensorflow/core/common_runtime/collective_util.h"
19 #include "tensorflow/core/common_runtime/copy_tensor.h"
20 #include "tensorflow/core/common_runtime/device.h"
21 #include "tensorflow/core/common_runtime/device_mgr.h"
22 #include "tensorflow/core/common_runtime/dma_helper.h"
23 #include "tensorflow/core/common_runtime/process_util.h"
24 #include "tensorflow/core/framework/allocator.h"
25 #include "tensorflow/core/framework/device_base.h"
26 #include "tensorflow/core/framework/op_kernel.h"
27 #include "tensorflow/core/framework/tensor.h"
28 #include "tensorflow/core/framework/types.h"
29 #include "tensorflow/core/lib/core/errors.h"
30 #include "tensorflow/core/lib/core/notification.h"
31 #include "tensorflow/core/lib/core/status.h"
32 #include "tensorflow/core/lib/strings/str_util.h"
33 #include "tensorflow/core/lib/strings/strcat.h"
34 #include "tensorflow/core/platform/env.h"
35 #include "tensorflow/core/platform/types.h"
36 
37 namespace tensorflow {
38 
Permuter()39 Permuter::Permuter()
40     : col_ctx_(nullptr), col_params_(nullptr), done_(nullptr), counter_(0) {}
41 
CheckCounterAndCallDone()42 StatusCallback Permuter::CheckCounterAndCallDone() {
43   return [this](const Status& s) {
44     mu_.lock();
45     status_.Update(s);
46     int counter = ++counter_;
47     Status status = status_;
48     mu_.unlock();
49     if (counter == 2) done_(status);
50   };
51 }
52 
InitializeCollectiveContext(std::shared_ptr<CollectiveContext> col_ctx)53 Status Permuter::InitializeCollectiveContext(
54     std::shared_ptr<CollectiveContext> col_ctx) {
55   DCHECK(col_ctx->dev_mgr);
56   col_ctx_ = col_ctx;
57   col_params_ = col_ctx->col_params;
58   return collective_util::InitializeDeviceAndLocality(
59       col_ctx->dev_mgr, col_ctx->device_name, &col_ctx->device,
60       &col_ctx->device_locality);
61 }
62 
Run(StatusCallback done)63 void Permuter::Run(StatusCallback done) {
64   if (col_params_->instance.permutation.size() !=
65       col_params_->instance.devices.size()) {
66     done(errors::Internal("Permutation must be the same size as devices"));
67   }
68   done_ = std::move(done);
69   DispatchSend(col_params_->default_rank,
70                col_params_->instance.permutation[col_params_->default_rank],
71                col_ctx_->input, CheckCounterAndCallDone());
72   for (int i = 0; i < col_params_->instance.permutation.size(); ++i) {
73     if (col_params_->default_rank == col_params_->instance.permutation[i]) {
74       DispatchRecv(i, col_params_->instance.permutation[i], col_ctx_->output,
75                    CheckCounterAndCallDone());
76     }
77   }
78 }
79 
DispatchSend(int src_rank,int target_rank,const Tensor * tensor,const StatusCallback & done)80 void Permuter::DispatchSend(int src_rank, int target_rank, const Tensor* tensor,
81                             const StatusCallback& done) {
82   string send_buf_key =
83       strings::StrCat(col_ctx_->exec_key, src_rank, target_rank);
84   VLOG(1) << "DispatchSend " << send_buf_key << " from_device "
85           << col_ctx_->device_name << " to_device "
86           << col_params_->instance.devices[target_rank]
87           << " target_rank=" << target_rank << " src_rank=" << src_rank;
88   col_ctx_->col_exec->remote_access()->PostToPeer(
89       col_params_->instance.devices[target_rank],
90       col_params_->group.task_names[target_rank], send_buf_key,
91       col_ctx_->device, col_ctx_->op_ctx->op_device_context(),
92       col_ctx_->op_ctx->output_alloc_attr(0), tensor, col_ctx_->device_locality,
93       col_ctx_->op_ctx->cancellation_manager(), done);
94 }
95 
DispatchRecv(int src_rank,int target_rank,Tensor * tensor,const StatusCallback & done)96 void Permuter::DispatchRecv(int src_rank, int target_rank, Tensor* tensor,
97                             const StatusCallback& done) {
98   string recv_buf_key =
99       strings::StrCat(col_ctx_->exec_key, src_rank, target_rank);
100   VLOG(1) << "DispatchRecv " << recv_buf_key << " to_device "
101           << col_ctx_->device_name << " from_device "
102           << col_params_->instance.devices[src_rank]
103           << " target_rank=" << target_rank << " src_rank=" << src_rank;
104   col_ctx_->col_exec->remote_access()->RecvFromPeer(
105       col_params_->instance.devices[src_rank],
106       col_params_->group.task_names[src_rank],
107       col_params_->task.is_local[src_rank], recv_buf_key, col_ctx_->device,
108       col_ctx_->op_ctx->op_device_context(),
109       col_ctx_->op_ctx->output_alloc_attr(0), tensor, col_ctx_->device_locality,
110       0, col_ctx_->op_ctx->cancellation_manager(), done);
111 }
112 namespace {
113 REGISTER_COLLECTIVE(Permute, Permuter);
114 }  // namespace
115 
116 }  // namespace tensorflow
117