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
|
#ifndef CAFFE2_OPERATORS_INT8_CHANNEL_SHUFFLE_OP_H_
#define CAFFE2_OPERATORS_INT8_CHANNEL_SHUFFLE_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 {
class Int8ChannelShuffleOp final : public ConvPoolOpBase<CPUContext> {
public:
explicit Int8ChannelShuffleOp(const OperatorDef& operator_def, Workspace* ws)
: ConvPoolOpBase<CPUContext>(operator_def, ws) {
#if !defined(FBCODE_CAFFE2) && defined(USE_INTERNAL_PTHREADPOOL_IMPL)
this->ws_ = ws;
#endif
OPERATOR_NEEDS_FEATURE(
this->order_ == StorageOrder::NHWC,
"Int8ChannelShuffleOp only supports NHWC order");
}
~Int8ChannelShuffleOp() {
if (this->qnnpackOperator_ != nullptr) {
qnnp_delete_operator(this->qnnpackOperator_);
this->qnnpackOperator_ = nullptr;
}
}
bool RunOnDeviceWithOrderNHWC() override {
const auto& X = Inputs()[0]->template Get<Int8TensorCPU>();
auto* Y = Outputs()[0]->template GetMutable<Int8TensorCPU>();
Y->t.ResizeLike(X.t);
Y->scale = X.scale;
Y->zero_point = X.zero_point;
const int32_t Y_offset =
this->template GetSingleArgument<int>("Y_zero_point", 0);
const float Y_scale =
this->template GetSingleArgument<float>("Y_scale", 1.0f);
TORCH_CHECK_EQ(Y_offset, X.zero_point);
TORCH_CHECK_EQ(Y_scale, X.scale);
TORCH_CHECK_GE(X.zero_point, std::numeric_limits<uint8_t>::min());
TORCH_CHECK_LE(X.zero_point, std::numeric_limits<uint8_t>::max());
const auto C = X.t.dim32(3);
const auto G = this->group_;
CAFFE_ENFORCE(C % G == 0, "");
initQNNPACK();
if (this->qnnpackOperator_ == nullptr) {
const qnnp_status createStatus = qnnp_create_channel_shuffle_nc_x8(
G /* groups */,
C / G /* group channels */,
0 /* flags */,
&this->qnnpackOperator_);
CAFFE_ENFORCE(
createStatus == qnnp_status_success,
"failed to create QNNPACK channel shuffle operator");
CAFFE_ENFORCE(this->qnnpackOperator_ != nullptr);
}
const qnnp_status setupStatus = qnnp_setup_channel_shuffle_nc_x8(
this->qnnpackOperator_,
X.t.numel() / C /* batch size */,
X.t.template data<uint8_t>(),
C /* X stride */,
Y->t.template mutable_data<uint8_t>(),
C /* Y stride */);
CAFFE_ENFORCE(
setupStatus == qnnp_status_success,
"failed to setup QNNPACK channel shuffle 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 channel shuffle operator");
return true;
}
private:
#if !defined(FBCODE_CAFFE2) && defined(USE_INTERNAL_PTHREADPOOL_IMPL)
Workspace* ws_;
#endif
// QNNPACK channel shuffle operator
qnnp_operator_t qnnpackOperator_{nullptr};
};
} // namespace int8
} // namespace caffe2
#endif // CAFFE2_OPERATORS_INT8_CHANNEL_SHUFFLE_OP_H_
|
