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
|
// See docs in ../ops/array_ops.cc.
#include <math.h>
#include <algorithm>
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/public/tensor.h"
#if GOOGLE_CUDA
#include "tensorflow/core/common_runtime/gpu_device_context.h"
#include "tensorflow/stream_executor/stream.h"
#endif // GOOGLE_CUDA
namespace tensorflow {
typedef Eigen::ThreadPoolDevice CPUDevice;
typedef Eigen::GpuDevice GPUDevice;
#if GOOGLE_CUDA
template <typename T>
struct CheckNumericsLaunch {
void Run(const GPUDevice& d, const T* data, int size,
int abnormal_detected[2]);
};
#endif
namespace {
template <typename Device, typename T>
class CheckNumericsOp;
// Partial specialization for CPU
template <typename T>
class CheckNumericsOp<CPUDevice, T> : public OpKernel {
public:
explicit CheckNumericsOp(OpKernelConstruction* context) : OpKernel(context) {
// message_ is used as the prefix for the assertion error message. For
// instance, this can be the name of the input op that produced the tensor.
OP_REQUIRES_OK(context, context->GetAttr("message", &message_));
}
void Compute(OpKernelContext* context) override {
// pass along the input to the output
context->set_output(0, context->input(0));
auto in = context->input(0).flat<T>();
const T* data = in.data();
const int size = in.size();
// Check to see if any element of the tensor is NaN or Inf.
int fp_props =
std::accumulate(data, data + size, 0, [](const int& x, const T& y) {
int prop = std::fpclassify(y);
int result = x;
if (prop == FP_INFINITE) {
result |= kInfBit;
} else if (prop == FP_NAN) {
result |= kNaNBit;
}
return result;
});
string status;
if ((fp_props & kInfBit) && (fp_props & kNaNBit)) {
status = "Inf and NaN";
} else {
if (fp_props & kInfBit) {
status = "Inf";
}
if (fp_props & kNaNBit) {
status = "NaN";
}
}
if (!status.empty()) {
context->SetStatus(errors::InvalidArgument(message_, " : Tensor had ",
status, " values"));
}
}
private:
string message_;
static const int kInfBit = 0x01;
static const int kNaNBit = 0x02;
};
#if GOOGLE_CUDA
// Partial specialization for GPU
template <typename T>
class CheckNumericsOp<GPUDevice, T> : public OpKernel {
public:
typedef GPUDevice Device;
explicit CheckNumericsOp(OpKernelConstruction* context) : OpKernel(context) {
// message_ is used as the prefix for the assertion error message. For
// instance, this can be the name of the input op that produced the tensor.
OP_REQUIRES_OK(context, context->GetAttr("message", &message_));
}
void Compute(OpKernelContext* context) override {
// pass along the input to the output
context->set_output(0, context->input(0));
auto input = context->input(0).flat<T>();
// Allocate and initialize the elements to hold the check results
const int abnormal_detected_size = 2;
Tensor abnormal_detected;
OP_REQUIRES_OK(context, context->allocate_temp(
DT_INT32, TensorShape({abnormal_detected_size}),
&abnormal_detected));
auto* stream = context->op_device_context<GPUDeviceContext>()->stream();
OP_REQUIRES(context, stream, errors::Internal("No GPU stream available."));
perftools::gputools::DeviceMemoryBase abnormal_detected_ptr(
abnormal_detected.flat<int>().data(),
abnormal_detected.flat<int>().size());
stream->ThenMemset32(&abnormal_detected_ptr, 0,
abnormal_detected.flat<int>().size() * sizeof(int));
// Call the Cuda kernels for the numerical checks
const Device& d = context->eigen_device<Device>();
CheckNumericsLaunch<T>().Run(d, input.data(), input.size(),
abnormal_detected.flat<int>().data());
// Copy the results from device to host
AllocatorAttributes attr;
attr.set_on_host(true);
attr.set_gpu_compatible(true);
Tensor abnormal_detected_out;
OP_REQUIRES_OK(context, context->allocate_temp(
DT_INT32, TensorShape({abnormal_detected_size}),
&abnormal_detected_out, attr));
int* abnormal_detected_host = abnormal_detected_out.flat<int>().data();
stream->ThenMemcpy(abnormal_detected_host, abnormal_detected_ptr,
abnormal_detected_size * sizeof(int));
stream->BlockHostUntilDone();
OP_REQUIRES(context, stream->ok(),
errors::Internal("cudaMemcpy from device to host failed"));
int is_nan = abnormal_detected_host[0];
int is_inf = abnormal_detected_host[1];
if (is_nan || is_inf) {
string status;
LOG(ERROR) << "abnormal_detected_host @" << abnormal_detected_host
<< " = {" << is_nan << ", " << is_inf << "} " << message_;
// Results should always be 1 or 0. If we see anything else then
// there has been some GPU memory corruption.
CHECK_GE(is_nan, 0);
CHECK_GE(is_inf, 0);
CHECK_LE(is_nan, 1);
CHECK_LE(is_inf, 1);
if (is_nan && is_inf) {
status = "Inf and NaN";
} else if (is_nan) {
status = "NaN";
} else if (is_inf) {
status = "Inf";
}
context->SetStatus(errors::InvalidArgument(message_, " : Tensor had ",
status, " values"));
}
}
private:
string message_;
};
#endif // GOOGLE_CUDA
} // namespace
REGISTER_KERNEL_BUILDER(Name("CheckNumerics")
.Device(DEVICE_CPU)
.TypeConstraint<float>("T"),
CheckNumericsOp<CPUDevice, float>);
REGISTER_KERNEL_BUILDER(Name("CheckNumerics")
.Device(DEVICE_CPU)
.TypeConstraint<double>("T"),
CheckNumericsOp<CPUDevice, double>);
#if GOOGLE_CUDA
REGISTER_KERNEL_BUILDER(Name("CheckNumerics")
.Device(DEVICE_GPU)
.TypeConstraint<float>("T"),
CheckNumericsOp<GPUDevice, float>);
REGISTER_KERNEL_BUILDER(Name("CheckNumerics")
.Device(DEVICE_GPU)
.TypeConstraint<double>("T"),
CheckNumericsOp<GPUDevice, double>);
#endif // GOOGLE_CUDA
} // namespace tensorflow
|
