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
|
#include "caffe2/operators/half_float_ops.h"
#include <c10/util/Half.h>
#include "caffe2/utils/cpuid.h"
#ifdef USE_FBGEMM
#include "fbgemm/FbgemmConvert.h"
#endif
namespace caffe2 {
inline void FloatToFloat16_ref(
const float* in,
at::Half* out,
size_t N,
bool do_clip = false) {
if (do_clip) {
constexpr float FP16_MAX = 65504.f;
for (size_t i = 0; i < N; ++i) {
out[i] = std::max(-FP16_MAX, std::min(in[i], FP16_MAX));
}
} else {
for (size_t i = 0; i < N; ++i) {
out[i] = in[i];
}
}
}
inline void Float16ToFloat_ref(const at::Half* in, float* out, size_t N) {
for (size_t i = 0; i < N; ++i) {
out[i] = in[i];
}
}
template <>
bool FloatToHalfOp<CPUContext>::RunOnDevice() {
auto& input = Input(0);
auto* output = Output(0, input.sizes(), at::dtype<at::Half>());
const float* data = input.template data<float>();
at::Half* out = output->template mutable_data<at::Half>();
auto N = input.numel();
#ifdef USE_FBGEMM
// There exists a verion fbgemm::FloatToFloat16_simd which will issue avx-512
// instructions when possible. However, this actually doesn't give perf
// benefits, according to benchmarks on T1/T6. Hence we stick to avx2 versions
// here.
if (GetCpuId().avx2()) {
fbgemm::FloatToFloat16_avx2(
data, reinterpret_cast<fbgemm::float16*>(out), N, clip_);
} else {
FloatToFloat16_ref(data, out, N, clip_);
}
#else
FloatToFloat16_ref(data, out, N, clip_);
#endif
return true;
}
template <>
bool HalfToFloatOp<CPUContext>::RunOnDevice() {
auto& input = Input(0);
auto* output = Output(0, input.sizes(), at::dtype<float>());
const at::Half* data = input.template data<at::Half>();
float* out = output->template mutable_data<float>();
auto N = input.numel();
#ifdef USE_FBGEMM
// Same reasoning of sticking to avx2
if (GetCpuId().avx2()) {
fbgemm::Float16ToFloat_avx2(
reinterpret_cast<const fbgemm::float16*>(data), out, N);
} else {
Float16ToFloat_ref(data, out, N);
}
#else
Float16ToFloat_ref(data, out, N);
#endif
return true;
}
REGISTER_CPU_OPERATOR(FloatToHalf, FloatToHalfOp<CPUContext>);
REGISTER_CPU_OPERATOR(HalfToFloat, HalfToFloatOp<CPUContext>);
OPERATOR_SCHEMA(FloatToHalf)
.NumInputs(1)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* unused */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
const TensorShape& X = in[0];
out.push_back(X);
out[0].set_data_type(TensorProto_DataType_FLOAT16);
return out;
});
OPERATOR_SCHEMA(HalfToFloat)
.NumInputs(1)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* unused */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
const TensorShape& X = in[0];
out.push_back(X);
out[0].set_data_type(TensorProto_DataType_FLOAT);
return out;
});
bool Float16ConstantFillOp::RunOnDevice() {
auto* output = Output(0, shape_, at::dtype<at::Half>());
const float givenValue =
this->template GetSingleArgument<float>("value", 0.0f);
at::Half givenFp16Value = givenValue;
if (output->numel()) {
at::Half* out = output->template mutable_data<at::Half>();
std::fill(out, out + output->numel(), givenFp16Value);
}
return true;
}
template <>
bool Float16UniformFillOp<CPUContext>::RunOnDevice() {
auto* output = Output(0, shape_, at::dtype<at::Half>());
at::Half* out = output->template mutable_data<at::Half>();
// Get a batch row by row and convert
auto leading_dim_sz = output->size(0);
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
int rowsz = output->numel() / output->size(0);
vector<float> intermediate_data_;
intermediate_data_.resize(rowsz);
// NOLINTNEXTLINE(clang-diagnostic-sign-compare)
for (uint64_t i = 0; i < leading_dim_sz; i++) {
math::RandUniform<float, CPUContext>(
rowsz, min_, max_, intermediate_data_.data(), &context_);
// NOLINTNEXTLINE(clang-diagnostic-sign-compare)
for (uint64_t j = 0; j < rowsz; j++) {
out[i * rowsz + j] = intermediate_data_[j];
}
}
return true;
}
REGISTER_CPU_OPERATOR(Float16ConstantFill, Float16ConstantFillOp);
REGISTER_CPU_OPERATOR(Float16UniformFill, Float16UniformFillOp<CPUContext>);
OPERATOR_SCHEMA(Float16UniformFill)
.NumInputs(0)
.NumOutputs(1)
.TensorInferenceFunction(Float16FillerTensorInference)
.SetDoc(
"Fills a half float tensor of a specified shape with"
" values from a uniform distribution[min,max]")
.Arg("shape", "Shape of the tensor")
.Arg("min", "Minimim value to generate")
.Arg("max", "Maximum value to generate");
NO_GRADIENT(Float16UniformFill);
OPERATOR_SCHEMA(Float16ConstantFill)
.NumInputs(0)
.NumOutputs(1)
.TensorInferenceFunction(Float16FillerTensorInference)
.Arg("value", "The value for the elements of the output tensor.")
.Arg("shape", "The shape of the output tensor.")
.Output(
0,
"output",
"Output tensor of constant values specified by 'value'");
class GetFloatToHalfGradient : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
vector<OperatorDef> GetGradientDefs() override {
return SingleGradientDef(
"HalfToFloat", "", vector<string>{GO(0)}, vector<string>{GI(0)});
}
};
REGISTER_GRADIENT(FloatToHalf, GetFloatToHalfGradient);
class GetHalfToFloatGradient : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
vector<OperatorDef> GetGradientDefs() override {
return SingleGradientDef(
"FloatToHalf", "", vector<string>{GO(0)}, vector<string>{GI(0)});
}
};
REGISTER_GRADIENT(HalfToFloat, GetHalfToFloatGradient);
NO_GRADIENT(Float16ConstantFill);
} // namespace caffe2
|
