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
|
/* 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.
==============================================================================*/
// Defines the XlaCompileOnDemandOp.
#include "tensorflow/compiler/jit/xla_compile_on_demand_op.h"
#include "tensorflow/compiler/jit/xla_device.h"
#include "tensorflow/compiler/jit/xla_launch_util.h"
#include "tensorflow/compiler/tf2xla/xla_compiler.h"
#include "tensorflow/compiler/tf2xla/xla_op_registry.h"
namespace tensorflow {
namespace {
std::map<int, OptionalTensor> GetVariables(OpKernelContext* ctx) {
std::map<int, OptionalTensor> variables;
for (int64 i = 0; i < ctx->num_inputs(); ++i) {
if (ctx->input(i).dtype() == DT_RESOURCE) {
Var* variable = nullptr;
ResourceHandle handle = HandleFromInput(ctx, i);
OptionalTensor& optional = variables[i];
optional.name = handle.name();
if (LookupResource(ctx, handle, &variable).ok()) {
tf_shared_lock lock(*variable->mu());
optional.present = true;
optional.value = *variable->tensor();
}
}
}
return variables;
}
} // namespace
Status XlaCompileOnDemandOp::Run(OpKernelContext* ctx,
const XlaDevice::Metadata& metadata,
const XlaCompiler::CompilationResult* result,
xla::LocalExecutable* executable) {
std::map<int, OptionalTensor> variables = GetVariables(ctx);
xla::LocalClient* client = metadata.client();
// Builds an XLA allocator for the device.
XlaComputationLaunchContext launch_context(
client, client->backend().memory_allocator(), true);
launch_context.PopulateInputs(ctx, result, variables);
se::Stream* stream =
ctx->op_device_context() ? ctx->op_device_context()->stream() : nullptr;
TF_RET_CHECK(stream);
VLOG(2) << "Executing computation: " << name();
for (const xla::ShapedBuffer* arg : launch_context.arguments()) {
VLOG(2) << name() << ": " << *arg;
}
xla::ExecutableRunOptions run_options;
run_options.set_stream(stream);
run_options.set_allocator(client->backend().memory_allocator());
run_options.set_intra_op_thread_pool(&ctx->eigen_cpu_device());
run_options.set_rng_seed(ctx->step_id());
xla::StatusOr<xla::ScopedShapedBuffer> run_result =
executable->Run(launch_context.arguments(), run_options);
TF_RETURN_IF_ERROR(run_result.status());
launch_context.PopulateOutputs(ctx, result, run_result.ConsumeValueOrDie());
return Status::OK();
}
bool XlaCompileOnDemandOp::MustArgumentBeConstant(const OpKernel* op_kernel,
int64 argument_idx) {
// TODO(jmolloy): This could be expensive, so memoize.
auto* constant_inputs = tensorflow::XlaOpRegistry::CompileTimeConstantInputs(
op_kernel->def().op());
CHECK(constant_inputs);
std::set<int64> constant_input_indices;
for (const auto& name : *constant_inputs) {
int start, stop;
TF_CHECK_OK(op_kernel->InputRange(name, &start, &stop));
for (int i = start; i < stop; ++i) {
constant_input_indices.insert(i);
}
}
return constant_input_indices.count(argument_idx) > 0;
}
bool XlaCompileOnDemandOp::ShouldArgumentBeConstant(const OpKernel* op_kernel,
int64 argument_idx) {
// Right now we only create kConstant arguments when absolutely required, but
// there may be benefit in eagerly constant-folding a larger subset of
// arguments in the future.
return MustArgumentBeConstant(op_kernel, argument_idx);
}
Status XlaCompileOnDemandOp::Compile(
OpKernelContext* ctx, const XlaDevice::Metadata& metadata,
const XlaCompiler::CompilationResult** result,
xla::LocalExecutable** executable) {
std::map<int, Tensor> constant_arguments;
for (int64 i = 0; i < ctx->num_inputs(); ++i) {
const Tensor& device_tensor = ctx->input(i);
if (const XlaTensor* xla_tensor = XlaTensor::FromTensor(&device_tensor)) {
if (xla_tensor->has_host_tensor() &&
ShouldArgumentBeConstant(&ctx->op_kernel(), i)) {
constant_arguments[i] = xla_tensor->host_tensor();
}
}
if (constant_arguments.count(i) == 0 &&
MustArgumentBeConstant(&ctx->op_kernel(), i)) {
// Slow path; the argument is not available as a host constant so we must
// fetch it synchronously.
Tensor host_tensor;
AllocatorAttributes attrs;
attrs.set_on_host(true);
TF_RETURN_IF_ERROR(ctx->allocate_temp(
device_tensor.dtype(), device_tensor.shape(), &host_tensor, attrs));
Notification n;
ctx->op_device_context()->CopyDeviceTensorToCPU(
&device_tensor, "ConstantArgument",
reinterpret_cast<Device*>(ctx->device()), &host_tensor,
[&](Status status) { n.Notify(); });
n.WaitForNotification();
constant_arguments[i] = host_tensor;
}
}
// We store information about the JIT-compiled XLA computation
// in the ResourceMgr.
ResourceMgr* rm = ctx->resource_manager();
CHECK(rm);
XlaCompilationCache* cache;
TF_RETURN_IF_ERROR(rm->LookupOrCreate<XlaCompilationCache>(
rm->default_container(), "xla_cache", &cache,
[&](XlaCompilationCache** cache) {
*cache = new XlaCompilationCache(metadata.client(),
metadata.jit_device_type());
return Status::OK();
}));
// Hold the reference to the JIT during evaluation. (We could probably
// free it sooner because the ResourceMgr will retain a reference, but
// this is more obviously correct.)
core::ScopedUnref cache_ref(cache);
XlaCompiler::Options options;
options.device_type = metadata.jit_device_type();
options.client = metadata.client();
options.flib_def =
new FunctionLibraryDefinition(OpRegistry::Global(), FunctionDefLibrary{});
options.shape_representation_fn = metadata.shape_representation_fn();
XlaCompiler::CompileOptions compile_options;
compile_options.is_entry_computation = true;
// Optimization: don't resolve constants. If we resolve constants we never
// emit them on the device, meaning that if they are needed by a following
// computation the host has to transfer them.
compile_options.resolve_compile_time_constants = false;
// Optimization: where possible, have the computation return a naked array
// rather than a one-element tuple.
compile_options.always_return_tuple = false;
std::map<int, OptionalTensor> variable_args = GetVariables(ctx);
return cache->CompileSingleOp(options, constant_arguments, variable_args, ctx,
result, executable, &compile_options);
}
void XlaCompileOnDemandOp::Compute(OpKernelContext* ctx) {
const XlaCompiler::CompilationResult* result;
xla::LocalExecutable* executable;
const XlaDevice::Metadata* metadata;
OP_REQUIRES_OK(ctx, XlaDevice::GetMetadata(ctx, &metadata));
OP_REQUIRES_OK(ctx, Compile(ctx, *metadata, &result, &executable));
OP_REQUIRES_OK(ctx, Run(ctx, *metadata, result, executable));
}
} // namespace tensorflow
|
