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
|
/* Copyright 2015 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.
==============================================================================*/
// See docs in ../ops/array_ops.cc.
#define EIGEN_USE_THREADS
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/kernels/bounds_check.h"
#include "tensorflow/core/kernels/ops_util.h"
#include "tensorflow/core/kernels/split_lib.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/lib/gtl/array_slice.h"
namespace tensorflow {
typedef Eigen::ThreadPoolDevice CPUDevice;
typedef Eigen::GpuDevice GPUDevice;
#ifdef TENSORFLOW_USE_SYCL
typedef Eigen::SyclDevice SYCLDevice;
#endif // TENSORFLOW_USE_SYCL
template <typename Device, typename T>
class UnpackOp : public OpKernel {
public:
explicit UnpackOp(OpKernelConstruction* context) : OpKernel(context) {
OP_REQUIRES_OK(context, context->GetAttr("axis", &axis_));
}
void Compute(OpKernelContext* context) override {
const int32 num = num_outputs();
const Tensor& input = context->input(0);
const TensorShape& input_shape = input.shape();
int axis = axis_;
if (axis < 0) axis += input_shape.dims();
OP_REQUIRES(context, 0 <= axis && axis < input_shape.dims(),
errors::InvalidArgument("axis = ", axis_, " not in [",
-input_shape.dims(), ", ",
input_shape.dims(), ")"));
OP_REQUIRES(
context, input_shape.dims() > 0 && input_shape.dim_size(axis) == num,
errors::InvalidArgument("Input shape axis ", axis, " must equal ", num,
", got shape ", input_shape.DebugString()));
auto output_shape = input_shape;
output_shape.RemoveDim(axis);
const int64 output_size = output_shape.num_elements();
OP_REQUIRES(
context, FastBoundsCheck(output_size,
std::numeric_limits<Eigen::DenseIndex>::max()),
errors::InvalidArgument("output size must fit in Eigen DenseIndex"));
// Special case: Aligned, so we can share the underlying buffer.
//
// Apply this optimization conservatively: if input is aligned,
// the resulting tensors must be aligned. It's conservative
// because if the immediate consumer of the resulting tensors are
// not using eigen for computation, its perfectly fine to avoid
// the copying.
if (axis == 0 &&
(output_size == 0 || IsInnerDimsSizeAligned<T>(input_shape))) {
for (int i = 0; i < num; ++i) {
Tensor output;
CHECK(output.CopyFrom(input.Slice(i, i + 1), output_shape));
context->set_output(i, output);
}
return;
}
int64 before_dim = 1;
for (int i = 0; i < axis; ++i) {
before_dim *= input_shape.dim_size(i);
}
int64 after_dim = 1;
for (int i = axis + 1; i < input_shape.dims(); ++i) {
after_dim *= input_shape.dim_size(i);
}
const int64 axis_dim = input_shape.dim_size(axis);
// Except for shape, unpack is a special case of split, so we reuse the
// same computational kernels.
auto input_reshaped =
input.shaped<T, 3>({1, before_dim, axis_dim * after_dim});
for (int i = 0; i < num; ++i) {
Tensor* output;
OP_REQUIRES_OK(context,
context->allocate_output(i, output_shape, &output));
if (output_shape.num_elements() > 0) {
auto output_shaped = output->shaped<T, 3>({1, before_dim, after_dim});
Eigen::DSizes<Eigen::DenseIndex, 3> indices{0, 0, i * after_dim};
Eigen::DSizes<Eigen::DenseIndex, 3> sizes{1, before_dim, after_dim};
functor::Split<Device, T>()(context->eigen_device<Device>(),
output_shaped, input_reshaped, indices,
sizes);
}
}
}
private:
int axis_;
};
#define REGISTER_UNPACK(type) \
REGISTER_KERNEL_BUILDER( \
Name("Unpack").Device(DEVICE_CPU).TypeConstraint<type>("T"), \
UnpackOp<CPUDevice, type>)
TF_CALL_ALL_TYPES(REGISTER_UNPACK);
#undef REGISTER_UNPACK
#if GOOGLE_CUDA
#define REGISTER_GPU(type) \
REGISTER_KERNEL_BUILDER( \
Name("Unpack").Device(DEVICE_GPU).TypeConstraint<type>("T"), \
UnpackOp<GPUDevice, type>)
TF_CALL_GPU_NUMBER_TYPES(REGISTER_GPU);
#undef REGISTER_GPU
// A special GPU kernel for int32.
// TODO(b/25387198): Also enable int32 in device memory. This kernel
// registration requires all int32 inputs and outputs to be in host memory.
REGISTER_KERNEL_BUILDER(Name("Unpack")
.Device(DEVICE_GPU)
.HostMemory("value")
.HostMemory("output")
.TypeConstraint<int32>("T"),
UnpackOp<CPUDevice, int32>);
#endif // GOOGLE_CUDA
#ifdef TENSORFLOW_USE_SYCL
#define REGISTER_SYCL(type) \
REGISTER_KERNEL_BUILDER( \
Name("Unpack").Device(DEVICE_SYCL).TypeConstraint<type>("T"), \
UnpackOp<SYCLDevice, type>)
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SYCL);
REGISTER_KERNEL_BUILDER(Name("Unpack")
.Device(DEVICE_SYCL)
.HostMemory("value")
.HostMemory("output")
.TypeConstraint<int32>("T"),
UnpackOp<CPUDevice, int32>);
#undef REGISTER_SYCL
#endif // TENSORFLOW_USE_SYCL
} // end namespace tensorflow
|
