1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
|
/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "tensorflow/contrib/tensorrt/kernels/trt_engine_op.h"
#include "tensorflow/contrib/tensorrt/log/trt_logger.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/stream_executor.h"
#include "tensorflow/core/platform/types.h"
#if GOOGLE_CUDA
#if GOOGLE_TENSORRT
#include "cuda/include/cuda_runtime_api.h"
namespace tensorflow {
namespace tensorrt {
static ::tensorflow::tensorrt::Logger logger;
TRTEngineOp::TRTEngineOp(OpKernelConstruction* context) : OpKernel(context) {
// read serialized_engine
string serialized_engine;
OP_REQUIRES_OK(context,
context->GetAttr("serialized_engine", &serialized_engine));
// register input output node name in trt_sub_graph
OP_REQUIRES_OK(context, context->GetAttr("input_nodes", &input_nodes_));
OP_REQUIRES_OK(context, context->GetAttr("output_nodes", &output_nodes_));
// TODO(samikama) runtime should be taken from a resourcemanager as well.
// Only engine should be in the op and context and runtime should be taken
// from resourcemanager
nvinfer1::IRuntime* infer = nvinfer1::createInferRuntime(logger);
trt_engine_ptr_.reset(infer->deserializeCudaEngine(
serialized_engine.c_str(), serialized_engine.size(), nullptr));
trt_execution_context_ptr_.reset(trt_engine_ptr_->createExecutionContext());
// Runtime is safe to delete after engine creation
infer->destroy();
}
void TRTEngineOp::Compute(OpKernelContext* context) {
int num_binding = context->num_inputs() + context->num_outputs();
std::vector<void*> buffers(num_binding);
size_t binding_index;
int num_batch = 0;
bool valid = true;
for (int i = 0; i < context->num_inputs(); i++) {
// Grab the input tensor
binding_index = trt_engine_ptr_->getBindingIndex(input_nodes_[i].c_str());
const Tensor& input_tensor = context->input(i);
const TensorShape& input_shape = input_tensor.shape();
if (i == 0) {
num_batch = input_shape.dim_size(0);
} else if (num_batch != input_shape.dim_size(0)) {
valid = false;
break;
}
switch (trt_engine_ptr_->getBindingDataType(binding_index)) {
case nvinfer1::DataType::kFLOAT:
buffers[binding_index] = (void*)(input_tensor.flat<float>().data());
break;
case nvinfer1::DataType::kHALF:
LOG(FATAL) << "half size is not supported yet!";
break;
case nvinfer1::DataType::kINT8:
LOG(FATAL) << "int8 is not supported yet!";
break;
}
}
// Might want a different way to inform the user of batch size inconsistency
if (!valid) LOG(WARNING) << "input data inconsistent batch size";
for (int i = 0; i < static_cast<int>(output_nodes_.size()); i++) {
// This is bad that we have to reallocate output buffer every run.
// Create an output tensor
binding_index = trt_engine_ptr_->getBindingIndex(output_nodes_[i].c_str());
Tensor* output_tensor = nullptr;
TensorShape output_shape;
if (binding_index != -1) {
auto dims = trt_engine_ptr_->getBindingDimensions(binding_index);
std::vector<int> trt_shape(dims.nbDims + 1);
trt_shape[0] = num_batch;
for (int j = 0; j < dims.nbDims; j++) trt_shape[j + 1] = dims.d[j];
OP_REQUIRES_OK(context,
TensorShapeUtils::MakeShape(
trt_shape.data(), trt_shape.size(), &output_shape));
} else {
LOG(FATAL) << "output node not found, at " << output_nodes_[i];
break;
}
OP_REQUIRES_OK(context,
context->allocate_output(i, output_shape, &output_tensor));
switch (trt_engine_ptr_->getBindingDataType(binding_index)) {
case nvinfer1::DataType::kFLOAT:
buffers[binding_index] =
reinterpret_cast<void*>(output_tensor->flat<float>().data());
break;
case nvinfer1::DataType::kHALF:
LOG(FATAL) << "half size is not supported yet!";
break;
case nvinfer1::DataType::kINT8:
LOG(FATAL) << "int8 is not supported yet!";
break;
}
}
// copied from cuda_kernel_helper since it seems only valid in *.cu.cc files
const cudaStream_t* stream = CHECK_NOTNULL(
reinterpret_cast<const cudaStream_t*>(context->op_device_context()
->stream()
->implementation()
->CudaStreamMemberHack()));
// execution handled by TF since we are getting stream from TF.
// it is safe for CPU pointer array (buffers) to go out of scope after enqueue
trt_execution_context_ptr_->enqueue(num_batch, &buffers[0], *stream, nullptr);
}
REGISTER_KERNEL_BUILDER(Name("TRTEngineOp").Device(DEVICE_GPU), TRTEngineOp);
} // namespace tensorrt
} // namespace tensorflow
#endif // GOOGLE_TENSORRT
#endif // GOOGLE_CUDA
|
