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
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
|
/* 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.
==============================================================================*/
#define EIGEN_USE_THREADS
#include "tensorflow/core/kernels/variable_ops.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/platform/types.h"
namespace tensorflow {
// Resource stored by variables in the resource manager
// (legacy, ref-style version).
class LegacyVar : public ResourceBase {
public:
explicit LegacyVar(DataType dtype) : tensor_(dtype) {}
// Not copyable or movable.
LegacyVar(const LegacyVar&) = delete;
LegacyVar& operator=(const LegacyVar&) = delete;
mutex* mu() { return &mu_; }
Tensor* tensor() { return &tensor_; }
string DebugString() override {
return strings::StrCat(DataTypeString(tensor_.dtype()), "/",
tensor_.shape().DebugString());
}
private:
mutex mu_;
Tensor tensor_;
~LegacyVar() override {}
};
VariableOp::VariableOp(OpKernelConstruction* context) : OpKernel(context) {
OP_REQUIRES_OK(context, context->GetAttr("shape", &shape_));
dtype_ = RemoveRefType(context->output_type(0));
}
void VariableOp::Compute(OpKernelContext* ctx) {
mutex_lock l(init_mu_);
if (!initialized_) {
OP_REQUIRES_OK(ctx, cinfo_.Init(ctx->resource_manager(), def(),
true /* use name() */));
initialized_ = true;
}
auto creator = [this](LegacyVar** var) {
*var = new LegacyVar(dtype_);
(*var)->tensor()->set_shape(shape_);
return Status::OK();
};
LegacyVar* var;
OP_REQUIRES_OK(ctx, cinfo_.resource_manager()->LookupOrCreate<LegacyVar>(
cinfo_.container(), cinfo_.name(), &var, creator));
// Output a reference to our tensor, so it may be updated.
//
// As long as the resource manager hasn't been cleared the ref we return
// here is valid because it owns a ref on var.
ctx->set_output_ref(0, var->mu(), var->tensor());
if (ctx->track_allocations() && var->tensor()->IsInitialized()) {
AllocatorAttributes attr;
attr.set_gpu_compatible(true);
attr.set_nic_compatible(true);
ctx->record_persistent_memory_allocation(var->tensor()->AllocatedBytes());
}
var->Unref();
}
class TemporaryVariableOp : public OpKernel {
public:
explicit TemporaryVariableOp(OpKernelConstruction* context)
: OpKernel(context) {
OP_REQUIRES_OK(context, context->GetAttr("shape", &shape_));
OP_REQUIRES_OK(context, context->GetAttr("dtype", &dtype_));
OP_REQUIRES_OK(context, context->GetAttr("var_name", &var_name_));
// Variable name defaults to op name if not specified explicitly.
if (var_name_.empty()) var_name_ = name();
}
void Compute(OpKernelContext* context) override {
Status s;
ResourceMgr* rm = context->resource_manager();
OP_REQUIRES(context, rm, errors::Internal("No per-step resource manager."));
auto* tmp_var = new TmpVar;
OP_REQUIRES(context, tmp_var,
errors::ResourceExhausted("Could not allocate TmpVar."));
tmp_var->name = var_name_;
s = context->allocate_temp(dtype_, shape_, &tmp_var->val);
if (!s.ok()) tmp_var->Unref();
OP_REQUIRES_OK(context, s);
OP_REQUIRES_OK(context, rm->Create(context->step_container()->name(),
var_name_, tmp_var));
context->set_output_ref(0, &tmp_var->mu, &tmp_var->val);
if (context->track_allocations()) {
context->record_persistent_memory_allocation(
tmp_var->val.AllocatedBytes());
}
}
private:
// Refcounted temporary variable resource.
friend class DestroyTemporaryVariableOp;
struct TmpVar : public ResourceBase {
mutex mu;
Tensor val;
string name;
string DebugString() override { return name; }
~TmpVar() override { VLOG(3) << "TmpVar " << name << " deleted"; }
};
TensorShape shape_;
DataType dtype_;
string var_name_;
};
class DestroyTemporaryVariableOp : public OpKernel {
public:
explicit DestroyTemporaryVariableOp(OpKernelConstruction* context)
: OpKernel(context) {
OP_REQUIRES(context, IsRefType(context->input_type(0)),
errors::InvalidArgument("lhs input needs to be a ref type"));
OP_REQUIRES_OK(context, context->GetAttr("var_name", &var_name_));
OP_REQUIRES(context, !var_name_.empty(),
errors::InvalidArgument("Missing var_name attribute"));
}
void Compute(OpKernelContext* context) override {
// NOTE(pbar): All other mutators of the Tensor Ref *must* have completed
// their execution before this DestroyTemporaryVariable op executes.
// This is typically achieved using control dependencies.
CHECK(IsRefType(context->input_dtype(0)));
Tensor tmpvar = context->mutable_input(0, false);
context->set_output(0, tmpvar);
ResourceMgr* rm = context->resource_manager();
OP_REQUIRES(context, rm, errors::Internal("No per-step resource manager."));
OP_REQUIRES_OK(context, rm->Delete<TemporaryVariableOp::TmpVar>(
context->step_container()->name(), var_name_));
if (context->track_allocations()) {
context->record_persistent_memory_allocation(
-static_cast<int64>(tmpvar.AllocatedBytes()));
}
}
private:
string var_name_;
};
class IsVariableInitializedOp : public OpKernel {
public:
explicit IsVariableInitializedOp(OpKernelConstruction* context)
: OpKernel(context) {}
void Compute(OpKernelContext* context) override {
// Get a mutable input tensor of the Ref input.
const Tensor& input_tensor = context->mutable_input(0, false);
Tensor* output = nullptr;
OP_REQUIRES_OK(context,
context->allocate_output(0, TensorShape({}), &output));
auto output_tensor = output->tensor<bool, 0>();
bool result = input_tensor.IsInitialized();
output_tensor() = result;
}
};
REGISTER_KERNEL_BUILDER(Name("Variable").Device(DEVICE_CPU), VariableOp);
REGISTER_KERNEL_BUILDER(Name("VariableV2").Device(DEVICE_CPU), VariableOp);
REGISTER_KERNEL_BUILDER(Name("TemporaryVariable").Device(DEVICE_CPU),
TemporaryVariableOp);
REGISTER_KERNEL_BUILDER(Name("DestroyTemporaryVariable").Device(DEVICE_CPU),
DestroyTemporaryVariableOp);
REGISTER_KERNEL_BUILDER(Name("IsVariableInitialized").Device(DEVICE_CPU),
IsVariableInitializedOp);
#ifdef TENSORFLOW_USE_SYCL
#define REGISTER_SYCL_KERNEL(type) \
REGISTER_KERNEL_BUILDER( \
Name("Variable").Device(DEVICE_SYCL).TypeConstraint<type>("dtype"), \
VariableOp); \
REGISTER_KERNEL_BUILDER( \
Name("VariableV2").Device(DEVICE_SYCL).TypeConstraint<type>("dtype"), \
VariableOp); \
REGISTER_KERNEL_BUILDER(Name("TemporaryVariable") \
.Device(DEVICE_SYCL) \
.TypeConstraint<type>("dtype"), \
TemporaryVariableOp); \
REGISTER_KERNEL_BUILDER(Name("DestroyTemporaryVariable") \
.Device(DEVICE_SYCL) \
.TypeConstraint<type>("T"), \
DestroyTemporaryVariableOp); \
REGISTER_KERNEL_BUILDER(Name("IsVariableInitialized") \
.Device(DEVICE_SYCL) \
.TypeConstraint<type>("dtype") \
.HostMemory("is_initialized"), \
IsVariableInitializedOp);
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SYCL_KERNEL);
#undef REGISTER_SYCL_KERNEL
#endif // TENSORFLOW_USE_SYCL
#if GOOGLE_CUDA
// Only register 'Variable' on GPU for the subset of types also supported by
// 'Assign' (see dense_update_ops.cc.)
#define REGISTER_GPU_KERNELS(type) \
REGISTER_KERNEL_BUILDER( \
Name("Variable").Device(DEVICE_GPU).TypeConstraint<type>("dtype"), \
VariableOp); \
REGISTER_KERNEL_BUILDER( \
Name("VariableV2").Device(DEVICE_GPU).TypeConstraint<type>("dtype"), \
VariableOp); \
REGISTER_KERNEL_BUILDER(Name("TemporaryVariable") \
.Device(DEVICE_GPU) \
.TypeConstraint<type>("dtype"), \
TemporaryVariableOp); \
REGISTER_KERNEL_BUILDER(Name("DestroyTemporaryVariable") \
.Device(DEVICE_GPU) \
.TypeConstraint<type>("T"), \
DestroyTemporaryVariableOp); \
REGISTER_KERNEL_BUILDER(Name("IsVariableInitialized") \
.Device(DEVICE_GPU) \
.TypeConstraint<type>("dtype") \
.HostMemory("is_initialized"), \
IsVariableInitializedOp);
TF_CALL_GPU_NUMBER_TYPES(REGISTER_GPU_KERNELS);
TF_CALL_int64(REGISTER_GPU_KERNELS);
#undef REGISTER_GPU_KERNELS
#endif // GOOGLE_CUDA
} // namespace tensorflow
|
