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
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
|
/* Copyright 2016 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.
==============================================================================*/
#define EIGEN_USE_THREADS
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/kernels/fill_functor.h"
#include "tensorflow/core/kernels/inplace_ops_functor.h"
#include "tensorflow/core/lib/core/status.h"
namespace tensorflow {
typedef Eigen::ThreadPoolDevice CPUDevice;
namespace functor {
template <typename T>
Status DoParallelConcatUpdate(const CPUDevice& d, const Tensor& value,
int32 loc, Tensor* output) {
auto Tvalue = value.flat_outer_dims<T>();
auto Toutput = output->flat_outer_dims<T>();
auto nrows = Toutput.dimension(0);
auto r = (loc % nrows + nrows) % nrows; // Guard index range.
Toutput.template chip<0>(r).device(d) = Tvalue.template chip<0>(0);
return Status::OK();
}
template <>
Status DoParallelConcat(const CPUDevice& d, const Tensor& value, int32 loc,
Tensor* output) {
CHECK_EQ(value.dtype(), output->dtype());
switch (value.dtype()) {
#define CASE(type) \
case DataTypeToEnum<type>::value: \
return DoParallelConcatUpdate<type>(d, value, loc, output);
TF_CALL_NUMBER_TYPES(CASE);
TF_CALL_string(CASE);
#undef CASE
default:
return errors::InvalidArgument("Unsupported data type: ", value.dtype());
}
}
} // end namespace functor
namespace {
template <typename Device>
class ParallelConcatUpdate : public OpKernel {
public:
explicit ParallelConcatUpdate(OpKernelConstruction* ctx) : OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("loc", &loc_));
}
void Compute(OpKernelContext* ctx) override {
auto value = ctx->input(0);
auto update = ctx->input(1);
OP_REQUIRES(
ctx, value.dims() == update.dims(),
errors::InvalidArgument("value and update shape doesn't match: ",
value.shape().DebugString(), " vs. ",
update.shape().DebugString()));
for (int i = 1; i < value.dims(); ++i) {
OP_REQUIRES(
ctx, value.dim_size(i) == update.dim_size(i),
errors::InvalidArgument("value and update shape doesn't match ",
value.shape().DebugString(), " vs. ",
update.shape().DebugString()));
}
OP_REQUIRES(ctx, 1 == update.dim_size(0),
errors::InvalidArgument("update shape doesn't match: ",
update.shape().DebugString()));
Tensor output = value; // This creates an alias intentionally.
const auto& d = ctx->eigen_device<Device>();
OP_REQUIRES_OK(
ctx, ::tensorflow::functor::DoParallelConcat(d, update, loc_, &output));
ctx->set_output(0, output);
}
private:
int32 loc_;
};
template <typename Device, typename T>
class ParallelConcatStart : public OpKernel {
public:
explicit ParallelConcatStart(OpKernelConstruction* ctx) : OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("shape", &shape_));
}
void Compute(OpKernelContext* ctx) override {
Tensor* out = nullptr;
// We do not know whether the output will be used on GPU. Setting it to be
// gpu-compatible for now.
AllocatorAttributes attr;
attr.set_gpu_compatible(true);
OP_REQUIRES_OK(ctx, ctx->allocate_output(0, shape_, &out, attr));
}
private:
TensorShape shape_;
};
class FailureKernel : public OpKernel {
public:
explicit FailureKernel(OpKernelConstruction* ctx) : OpKernel(ctx) {
OP_REQUIRES_OK(ctx,
errors::Internal("Found instance of parallel_stack which "
"could not be properly replaced."));
}
void Compute(OpKernelContext*) override {}
};
#define REGISTER(type) \
REGISTER_KERNEL_BUILDER(Name("_ParallelConcatUpdate") \
.Device(DEVICE_CPU) \
.TypeConstraint<type>("T"), \
ParallelConcatUpdate<CPUDevice>);
TF_CALL_POD_STRING_TYPES(REGISTER)
#undef REGISTER
#define REGISTER_EMPTY(type) \
REGISTER_KERNEL_BUILDER(Name("_ParallelConcatStart") \
.Device(DEVICE_CPU) \
.TypeConstraint<type>("dtype"), \
ParallelConcatStart<CPUDevice, type>)
TF_CALL_POD_STRING_TYPES(REGISTER_EMPTY)
#undef REGISTER_EMPTY
#define REGISTER_PARALLEL_CONCAT(type) \
REGISTER_KERNEL_BUILDER( \
Name("ParallelConcat").Device(DEVICE_CPU).TypeConstraint<type>("T"), \
FailureKernel);
TF_CALL_POD_STRING_TYPES(REGISTER_PARALLEL_CONCAT);
#undef REGISTER_PARALLEL_CONCAT
#if GOOGLE_CUDA
typedef Eigen::GpuDevice GPUDevice;
#define REGISTER_EMPTY(type) \
REGISTER_KERNEL_BUILDER(Name("_ParallelConcatStart") \
.Device(DEVICE_GPU) \
.TypeConstraint<type>("dtype"), \
ParallelConcatStart<GPUDevice, type>);
TF_CALL_GPU_NUMBER_TYPES(REGISTER_EMPTY)
#undef REGISTER_EMPTY
#define REGISTER_PARALLEL_CONCAT(type) \
REGISTER_KERNEL_BUILDER( \
Name("ParallelConcat").Device(DEVICE_GPU).TypeConstraint<type>("T"), \
FailureKernel);
TF_CALL_GPU_NUMBER_TYPES(REGISTER_PARALLEL_CONCAT);
#undef REGISTER_PARALLEL_CONCAT
#define REGISTER(type) \
REGISTER_KERNEL_BUILDER(Name("_ParallelConcatUpdate") \
.Device(DEVICE_GPU) \
.TypeConstraint<type>("T"), \
ParallelConcatUpdate<GPUDevice>);
TF_CALL_GPU_NUMBER_TYPES(REGISTER)
#undef REGISTER
// Register versions that operate on int32 data on the CPU even though the op
// has been placed on the GPU
REGISTER_KERNEL_BUILDER(Name("_ParallelConcatUpdate")
.Device(DEVICE_GPU)
.HostMemory("value")
.HostMemory("update")
.HostMemory("output")
.TypeConstraint<int32>("T"),
ParallelConcatUpdate<CPUDevice>);
#endif
} // end namespace
} // end namespace tensorflow
|
