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
|
#ifndef CAFFE2_OPERATORS_INT8_FC_OP_H_
#define CAFFE2_OPERATORS_INT8_FC_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_op_shared.h"
#include "caffe2/operators/quantized/int8_utils.h"
namespace caffe2 {
namespace int8 {
class Int8FCOp final : public Operator<CPUContext> {
public:
explicit Int8FCOp(const OperatorDef& operator_def, Workspace* ws)
: Operator<CPUContext>(operator_def, ws), ws_(ws) {
createSharedBuffer<CPUContext>(ws_);
}
~Int8FCOp() {
if (this->qnnpackObject_ != nullptr) {
qnnp_delete_operator(this->qnnpackObject_);
this->qnnpackObject_ = nullptr;
}
}
bool RunOnDevice() override {
const auto& X = Inputs()[0]->Get<Int8TensorCPU>();
const auto& W = Inputs()[1]->Get<Int8TensorCPU>();
const auto& B = Inputs()[2]->Get<Int8TensorCPU>();
auto* Y = Outputs()[0]->GetMutable<Int8TensorCPU>();
int32_t Y_offset = 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_offset;
// (NxHxW)xC == MxK x (NxK) -> MxN
const auto K = X.t.size_from_dim(1);
const auto N = W.t.size(0);
TORCH_CHECK_EQ(K, W.t.size(1));
TORCH_CHECK_EQ(N, B.t.numel());
const auto M = X.t.numel() / K;
ReinitializeTensor(&Y->t, {M, N}, at::dtype<uint8_t>().device(CPU));
runWithSharedBuffer<CPUContext>(ws_, [&](Tensor* buffer) {
initQNNPACK();
#if !defined(FBCODE_CAFFE2) && defined(USE_INTERNAL_PTHREADPOOL_IMPL)
pthreadpool_t threadpool =
reinterpret_cast<pthreadpool_t>(ws_->GetThreadPool());
#endif
if (this->qnnpackObject_ == nullptr) {
const qnnp_status createStatus = qnnp_create_fully_connected_nc_q8(
K,
N,
X.zero_point,
X.scale,
W.zero_point,
W.scale,
#ifndef _MSC_VER
W.t.template data<uint8_t>(),
B.t.template data<int32_t>(),
#else
W.t.data<uint8_t>(),
B.t.data<int32_t>(),
#endif
Y->zero_point,
Y->scale,
std::numeric_limits<uint8_t>::min(),
std::numeric_limits<uint8_t>::max(),
0 /* flags */,
&this->qnnpackObject_);
CAFFE_ENFORCE(
createStatus == qnnp_status_success,
"failed to create QNNPACK fully connected operator");
CAFFE_ENFORCE(this->qnnpackObject_ != nullptr);
}
uint8_t* inputPtr = X.t.template mutable_data<uint8_t>();
if (K < 8) {
buffer->Resize(std::vector<int64_t>{X.t.numel() + 8});
inputPtr = buffer->template mutable_data<uint8_t>() + 8;
memcpy(inputPtr, X.t.template data<uint8_t>(), X.t.numel());
}
if (lastBatchSize_ != static_cast<size_t>(M) ||
lastInputPointer_ != inputPtr ||
lastOutputPointer_ != Y->t.template mutable_data<uint8_t>()) {
const qnnp_status setupStatus = qnnp_setup_fully_connected_nc_q8(
this->qnnpackObject_,
M,
inputPtr,
K /* input stride */,
Y->t.template mutable_data<uint8_t>(),
N /* output stride */);
CAFFE_ENFORCE(
setupStatus == qnnp_status_success,
"failed to setup QNNPACK fully connected operator");
lastBatchSize_ = static_cast<size_t>(M);
lastInputPointer_ = inputPtr;
lastOutputPointer_ = Y->t.template mutable_data<uint8_t>();
}
#if defined(FBCODE_CAFFE2) || !defined(USE_INTERNAL_PTHREADPOOL_IMPL)
const qnnp_status runStatus =
qnnp_run_operator(this->qnnpackObject_, nullptr /* thread pool */);
#else
const qnnp_status runStatus =
qnnp_run_operator(this->qnnpackObject_, threadpool);
#endif
CAFFE_ENFORCE(
runStatus == qnnp_status_success, "failed to run QNNPACK operator");
});
return true;
}
private:
Workspace* ws_;
// QNNPACK convolution object
qnnp_operator_t qnnpackObject_{nullptr};
// batch size in the previous call to RunOnDeviceWithOrderNHWC
size_t lastBatchSize_{0};
// input pointer in the previous call to RunOnDeviceWithOrderNHWC
const void* lastInputPointer_{nullptr};
// output pointer in the previous call to RunOnDeviceWithOrderNHWC
void* lastOutputPointer_{nullptr};
};
} // namespace int8
} // namespace caffe2
#endif // CAFFE2_OPERATORS_INT8_FC_OP_H_
|
