xref: /aosp_15_r20/external/tensorflow/tensorflow/compiler/xla/service/gpu/outfeed_manager.cc (revision b6fb3261f9314811a0f4371741dbb8839866f948)
1 /* Copyright 2018 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/compiler/xla/service/gpu/outfeed_manager.h"
17 
18 #include <memory>
19 
20 #include "tensorflow/compiler/xla/map_util.h"
21 #include "tensorflow/compiler/xla/shape_util.h"
22 #include "tensorflow/core/platform/logging.h"
23 
24 #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
25 #include "tensorflow/compiler/xla/service/gpu/xla_executor_state.h"
26 #include "tensorflow/stream_executor/gpu/gpu_executor.h"
27 #endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
28 
29 namespace xla {
30 namespace gpu {
31 
GetOrCreateOutfeedManager(se::StreamExecutor * executor)32 OutfeedManager *GetOrCreateOutfeedManager(se::StreamExecutor *executor) {
33 #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
34   stream_executor::gpu::GpuExecutor *gpu_executor =
35       stream_executor::gpu::ExtractGpuExecutor(executor);
36   auto *xla_state =
37       gpu_executor->getOrCreateXLAState<GpuExecutorXLAState>(executor);
38   return xla_state->getOrCreateOutfeedManager(executor);
39 #else   // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
40   return nullptr;
41 #endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
42 }
43 
TransferLiteralFromOutfeed(se::StreamExecutor * executor,MutableBorrowingLiteral literal)44 Status OutfeedManager::TransferLiteralFromOutfeed(
45     se::StreamExecutor* executor, MutableBorrowingLiteral literal) {
46   ShapeTree<std::unique_ptr<gpu::OutfeedBuffer>> outfeed_buffers(
47       &literal.shape());
48 
49   for (auto& leaf : outfeed_buffers.leaves()) {
50     const Shape& shape = ShapeUtil::GetSubshape(literal.shape(), leaf.first);
51     CHECK(shape.IsArray()) << ShapeUtil::HumanStringWithLayout(shape);
52     leaf.second =
53         std::make_unique<gpu::OutfeedBuffer>(ShapeUtil::ByteSizeOf(shape));
54     leaf.second->set_destination(
55         std::make_unique<MutableBorrowingLiteral>(literal, leaf.first));
56   }
57 
58   // Give the tree of buffers to the outfeed manager. The device will fill it
59   // while we're waiting for it below.
60   gpu::OutfeedManager* outfeed_manager =
61       gpu::GetOrCreateOutfeedManager(executor);
62   outfeed_manager->EnqueueDestination(&outfeed_buffers);
63 
64   // Now wait till all the buffers are written.
65   for (auto& leaf : outfeed_buffers.leaves()) {
66     const Shape& shape = ShapeUtil::GetSubshape(literal.shape(), leaf.first);
67     CHECK(shape.IsArray()) << ShapeUtil::HumanStringWithLayout(shape);
68     leaf.second->WaitUntilAvailable();
69   }
70 
71   return OkStatus();
72 }
73 
74 }  // namespace gpu
75 }  // namespace xla
76