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
|
#include "caffe2/quantization/server/spatial_batch_norm_dnnlowp_op.h"
#include "caffe2/quantization/server/caffe2_dnnlowp_utils.h"
namespace caffe2 {
template <typename T, bool ReluFused>
SpatialBNDNNLowPOp<T, ReluFused>::SpatialBNDNNLowPOp(
const OperatorDef& operator_def,
Workspace* ws)
: DNNLowPOp<T, SpatialBNOp<CPUContext>>(operator_def, ws),
OP_SINGLE_ARG(double, "epsilon", epsilon_, 1e-5),
order_(StringToStorageOrder(
this->template GetSingleArgument<std::string>("order", "NCHW"))) {
bool is_test = this->template GetSingleArgument<bool>("is_test", false);
OPERATOR_NEEDS_FEATURE(
is_test, "SpatialBN DNNLOWP op only works for inference.");
CAFFE_ENFORCE_NE(
order_,
StorageOrder::UNKNOWN,
"order should be either \"NCHW\" or \"NHWC\".");
CAFFE_ENFORCE(OutputSize() == 1);
CAFFE_ENFORCE_GT(epsilon_, 0);
}
template <typename T, bool ReluFused>
void SpatialBNDNNLowPOp<T, ReluFused>::ComputeFusedParam_(
const int C,
const float* scale,
const float* bias,
const float* mean,
const float* var,
float* alpha,
float* beta) {
EigenVectorArrayMap<float> alpha_arr(alpha, C);
EigenVectorArrayMap<float> beta_arr(beta, C);
alpha_arr = ConstEigenVectorArrayMap<float>(scale, C) *
(ConstEigenVectorArrayMap<float>(var, C) + epsilon_).rsqrt();
beta_arr = ConstEigenVectorArrayMap<float>(bias, C) -
alpha_arr * ConstEigenVectorArrayMap<float>(mean, C);
// Adjust alpha and beta considering quantization scales
alpha_arr = alpha_arr * (in_qparams_[0].scale / out_qparams_.scale);
beta_arr = beta_arr / out_qparams_.scale;
}
template <typename T, bool ReluFused>
bool SpatialBNDNNLowPOp<T, ReluFused>::RunOnDevice() {
if (!this->arguments_parsed_) {
dnnlowp::ParseDNNLowPOperatorArguments(
this, &dequantize_output_, &measure_quantization_error_, &followed_by_);
if (ReluFused) {
// It's actually fused with Relu not followed by but setting this to make
// sure quantization error is correctly measured in
// this->MeasureQuantizationError_
followed_by_ = "Relu";
dnnlowp::AdjustOutputTensorQuantizationParamsWithFollowedBy(
this, followed_by_);
}
this->arguments_parsed_ = true;
}
const auto& X = InputTensorCPU_(INPUT);
const auto& scale = Input(SCALE);
const auto& bias = Input(BIAS);
const int ndim = X.dim();
CAFFE_ENFORCE_GE(ndim, 3);
const int N = X.dim32(0);
const int C = (order_ == StorageOrder::NCHW ? X.dim32(1) : X.dim32(ndim - 1));
const std::vector<int> X_dims(X.sizes().cbegin(), X.sizes().cend());
const int HxW = X.size_from_dim(1) / C;
CAFFE_ENFORCE_EQ(scale.numel(), C);
CAFFE_ENFORCE_EQ(bias.numel(), C);
GetOutputQuantizationParams_();
in_qparams_[0] = GetInputTensorQuantizationParamsOf(this, 0, qfactory_.get());
const float* scale_data = scale.template data<float>();
const float* bias_data = bias.template data<float>();
ReinitializeTensor(
&alpha_, {C}, at::dtype<float>().device(CPUContext::GetDeviceType()));
ReinitializeTensor(
&beta_, {C}, at::dtype<float>().device(CPUContext::GetDeviceType()));
float* alpha_data = alpha_.template mutable_data<float>();
float* beta_data = beta_.template mutable_data<float>();
const auto& mean = Input(EST_MEAN);
const auto& var = Input(EST_VAR);
CAFFE_ENFORCE_EQ(mean.numel(), C);
CAFFE_ENFORCE_EQ(var.numel(), C);
auto* Y = OutputTensorCPU_(OUTPUT);
Y->Resize(X.sizes());
T* Y_data = GetQuantizedOutputData_();
if (N == 0) {
return true;
}
ComputeFusedParam_(
C,
scale_data,
bias_data,
mean.template data<float>(),
var.template data<float>(),
alpha_data,
beta_data);
vector<T> X_temp;
const T* X_data =
dnnlowp::QuantizeInputIfNeeded(this, 0, in_qparams_[0], X_temp);
if (order_ == StorageOrder::NCHW) {
for (int c = 0; c < C; ++c) {
for (int i = 0; i < N; ++i) {
for (int j = 0; j < HxW; ++j) {
long quantized_down = out_qparams_.zero_point +
std::lrintf(alpha_data[c] *
(X_data[(i * C + c) * HxW + j] -
in_qparams_[0].zero_point) +
beta_data[c]);
if (ReluFused) {
quantized_down =
std::max<long>(quantized_down, out_qparams_.zero_point);
}
Y_data[(i * C + c) * HxW + j] =
fbgemm::clamp<long, T>(quantized_down, 8);
}
}
}
} else {
if (GetCpuId().avx2()) {
internal::SpatialBNNHWCAVX2<T>(
N,
C,
HxW,
in_qparams_[0].zero_point,
out_qparams_.zero_point,
X_data,
alpha_data,
beta_data,
Y_data,
ReluFused);
} else {
for (int i = 0; i < N * HxW; ++i) {
for (int c = 0; c < C; ++c) {
long quantized_down = out_qparams_.zero_point +
std::lrintf(alpha_data[c] *
(X_data[i * C + c] - in_qparams_[0].zero_point) +
beta_data[c]);
if (ReluFused) {
quantized_down =
std::max<long>(quantized_down, out_qparams_.zero_point);
}
Y_data[i * C + c] = fbgemm::clamp<long, T>(quantized_down, 8);
}
}
}
}
RunOnDeviceEpilogue_();
return true;
}
REGISTER_CPU_OPERATOR_WITH_ENGINE(
SpatialBN,
DNNLOWP,
SpatialBNDNNLowPOp<uint8_t>);
REGISTER_CPU_OPERATOR_WITH_ENGINE(
Int8SpatialBN,
DNNLOWP,
SpatialBNDNNLowPOp<uint8_t>);
REGISTER_CPU_OPERATOR_WITH_ENGINE(
Int8SpatialBNRelu,
DNNLOWP,
SpatialBNDNNLowPOp<uint8_t, true>);
// NOLINTNEXTLINE(cppcoreguidelines-avoid-magic-numbers,cppcoreguidelines-avoid-non-const-global-variables)
OPERATOR_SCHEMA(Int8SpatialBN).NumInputs(5).NumOutputs(1);
// NOLINTNEXTLINE(cppcoreguidelines-avoid-magic-numbers,cppcoreguidelines-avoid-non-const-global-variables)
OPERATOR_SCHEMA(Int8SpatialBNRelu).NumInputs(5).NumOutputs(1);
} // namespace caffe2
|
