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
|
#ifndef CAFFE2_OPERATORS_INT8_AVERAGE_POOL_OP_H_
#define CAFFE2_OPERATORS_INT8_AVERAGE_POOL_OP_H_
#include <qnnpack.h>
#include "caffe2/core/context.h"
#include "caffe2/core/operator.h"
#include "caffe2/core/tensor_int8.h"
#include "caffe2/operators/conv_pool_op_base.h"
#include "caffe2/operators/quantized/int8_utils.h"
namespace caffe2 {
namespace int8 {
template <Activation Ac>
class Int8AveragePoolOp final : public ConvPoolOpBase<CPUContext> {
public:
template <class... Args>
explicit Int8AveragePoolOp(Args&&... args)
: ConvPoolOpBase<CPUContext>(std::forward<Args>(args)...) {
OPERATOR_NEEDS_FEATURE(
this->order_ == StorageOrder::NHWC, "Int8 only supports NHWC order.");
}
~Int8AveragePoolOp() {
if (this->qnnpackOperator_ != nullptr) {
qnnp_delete_operator(this->qnnpackOperator_);
this->qnnpackOperator_ = nullptr;
}
if (this->qnnpackGlobalOperator_ != nullptr) {
qnnp_delete_operator(this->qnnpackGlobalOperator_);
this->qnnpackGlobalOperator_ = nullptr;
}
}
bool RunOnDeviceWithOrderNHWC() override {
const auto& X = Inputs()[0]->template Get<Int8TensorCPU>();
auto* Y = Outputs()[0]->template GetMutable<Int8TensorCPU>();
int32_t Y_zero_point =
this->template GetSingleArgument<int>("Y_zero_point", 0);
auto Y_scale = this->template GetSingleArgument<float>("Y_scale", 1);
Y->scale = Y_scale;
Y->zero_point = Y_zero_point;
TORCH_CHECK_EQ(X.t.dim(), 4);
const int channels = X.t.dim32(3);
ConvPoolOpBase<CPUContext>::SetOutputSize(X.t, &(Y->t), channels);
initQNNPACK();
const bool anyPadding =
pad_t() != 0 || pad_r() != 0 || pad_b() != 0 || pad_l() != 0;
const bool anyStride = stride_h() > 1 || stride_w() > 1;
const bool globalPooling = !anyPadding && !anyStride &&
(X.t.dim32(1) == kernel_h() && X.t.dim32(2) == kernel_w());
if (globalPooling) {
if (this->qnnpackGlobalOperator_ == nullptr) {
const qnnp_status createStatus =
qnnp_create_global_average_pooling_nwc_q8(
channels,
X.zero_point,
X.scale,
Y->zero_point,
Y->scale,
activationLimits(Y->scale, Y->zero_point, Ac).first,
activationLimits(Y->scale, Y->zero_point, Ac).second,
0 /* flags */,
&this->qnnpackGlobalOperator_);
CAFFE_ENFORCE(
createStatus == qnnp_status_success,
"failed to create QNNPACK Global Average Pooling operator");
CAFFE_ENFORCE(this->qnnpackGlobalOperator_ != nullptr);
}
const qnnp_status setupStatus = qnnp_setup_global_average_pooling_nwc_q8(
this->qnnpackGlobalOperator_,
X.t.dim32(0),
X.t.dim32(1) * X.t.dim32(2),
X.t.template data<uint8_t>(),
channels,
Y->t.template mutable_data<uint8_t>(),
channels);
CAFFE_ENFORCE(
setupStatus == qnnp_status_success,
"failed to setup QNNPACK Global Average Pooling operator");
#if defined(FBCODE_CAFFE2) || !defined(USE_INTERNAL_PTHREADPOOL_IMPL)
const qnnp_status runStatus = qnnp_run_operator(
this->qnnpackGlobalOperator_, nullptr /* thread pool */);
#else
pthreadpool_t threadpool =
reinterpret_cast<pthreadpool_t>(ws_->GetThreadPool());
const qnnp_status runStatus =
qnnp_run_operator(this->qnnpackGlobalOperator_, threadpool);
#endif
CAFFE_ENFORCE(
runStatus == qnnp_status_success,
"failed to run QNNPACK Global Average Pooling operator");
} else {
if (this->qnnpackOperator_ == nullptr) {
const qnnp_status createStatus = qnnp_create_average_pooling2d_nhwc_q8(
pad_t(),
pad_r(),
pad_b(),
pad_l(),
kernel_h(),
kernel_w(),
stride_h(),
stride_w(),
channels,
X.zero_point,
X.scale,
Y->zero_point,
Y->scale,
activationLimits(Y->scale, Y->zero_point, Ac).first,
activationLimits(Y->scale, Y->zero_point, Ac).second,
0 /* flags */,
&this->qnnpackOperator_);
CAFFE_ENFORCE(
createStatus == qnnp_status_success,
"failed to create QNNPACK Average Pooling operator");
CAFFE_ENFORCE(this->qnnpackOperator_ != nullptr);
}
const qnnp_status setupStatus = qnnp_setup_average_pooling2d_nhwc_q8(
this->qnnpackOperator_,
X.t.dim32(0),
X.t.dim32(1),
X.t.dim32(2),
X.t.template data<uint8_t>(),
channels,
Y->t.template mutable_data<uint8_t>(),
channels,
nullptr /* thread pool */);
CAFFE_ENFORCE(
setupStatus == qnnp_status_success,
"failed to setup QNNPACK Average Pooling operator");
#if defined(FBCODE_CAFFE2) || !defined(USE_INTERNAL_PTHREADPOOL_IMPL)
const qnnp_status runStatus =
qnnp_run_operator(this->qnnpackOperator_, nullptr /* thread pool */);
#else
pthreadpool_t threadpool =
reinterpret_cast<pthreadpool_t>(ws_->GetThreadPool());
const qnnp_status runStatus =
qnnp_run_operator(this->qnnpackOperator_, threadpool);
#endif
CAFFE_ENFORCE(
runStatus == qnnp_status_success,
"failed to run QNNPACK Average Pooling operator");
}
return true;
}
private:
// QNNPACK Average Pooling operator
qnnp_operator_t qnnpackOperator_{nullptr};
// QNNPACK Global Average Pooling operator
qnnp_operator_t qnnpackGlobalOperator_{nullptr};
};
} // namespace int8
} // namespace caffe2
#endif // CAFFE2_OPERATORS_INT8_AVERAGE_POOL_OP_H_
|
