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
|
#include "caffe2/quantization/server/channel_shuffle_dnnlowp_op.h"
#include "caffe2/quantization/server/caffe2_dnnlowp_utils.h"
#include "caffe2/quantization/server/transpose.h"
#include "caffe2/utils/eigen_utils.h"
namespace caffe2 {
template <typename T>
ChannelShuffleDNNLowPOp<T>::ChannelShuffleDNNLowPOp(
const OperatorDef& operator_def,
Workspace* ws)
: BaseType(operator_def, ws),
order_(StringToStorageOrder(
this->template GetSingleArgument<std::string>("order", "NCHW"))),
OP_SINGLE_ARG(int, "group", group_, 1) {
CAFFE_ENFORCE_NE(order_, StorageOrder::UNKNOWN);
}
template <typename T>
bool ChannelShuffleDNNLowPOp<T>::RunOnDevice() {
return order_ == StorageOrder::NCHW ? RunOnDeviceWithOrderNCHW()
: RunOnDeviceWithOrderNHWC();
}
template <typename T>
bool ChannelShuffleDNNLowPOp<T>::RunOnDeviceWithOrderNCHW() {
using namespace dnnlowp;
this->ParseDNNLowPOperatorArguments_();
// Choose quantization params
TensorQuantizationParams in_qparams =
GetInputTensorQuantizationParamsOf(this, 0, qfactory_.get());
const auto& X = InputTensorCPU_(0);
auto* Y = OutputTensorCPU_(0);
Y->ResizeLike(X);
const int N = X.dim32(0);
const int C = X.dim32(1);
const int G = group_;
CAFFE_ENFORCE_EQ(C % G, 0);
const int K = C / G;
const int HxW = X.size_from_dim(2);
const int stride = C * HxW;
const T* X_data = X.template data<T>();
T* Y_data = Y->template mutable_data<T>();
#ifdef _OPENMP
#pragma omp parallel for
#endif
for (int i = 0; i < N; ++i) {
ConstEigenMatrixMap<T> X_mat(X_data + i * stride, K * HxW, G);
for (int j = 0; j < K; ++j) {
EigenMatrixMap<T>(Y_data + i * stride + j * G * HxW, HxW, G) =
X_mat.block(j * HxW, 0, HxW, G);
}
}
// Even if there is a pre-chosen quantization parameters for the output,
// it is ignored because channel shuffle output quantization should be same
// as the input.
PropagateOutputTensorQuantizationParams(this, 0, in_qparams);
return true;
}
template <typename T>
bool ChannelShuffleDNNLowPOp<T>::RunOnDeviceWithOrderNHWC() {
using namespace dnnlowp;
this->ParseDNNLowPOperatorArguments_();
// Choose quantization params
TensorQuantizationParams in_qparams =
GetInputTensorQuantizationParamsOf(this, 0, qfactory_.get());
const auto& X = InputTensorCPU_(0);
auto* Y = OutputTensorCPU_(0);
Y->ResizeLike(X);
const auto C = X.dim32(X.ndim() - 1);
const auto G = this->group_;
CAFFE_ENFORCE(C % G == 0, "");
const auto K = C / G;
std::array<int, 2> dims = {G, K};
std::array<int, 2> axes = {1, 0};
const T* X_data = X.template data<T>();
T* Y_data = Y->template mutable_data<T>();
if (G == 4 && std::is_same<T, std::uint8_t>::value && GetCpuId().avx2()) {
#ifdef _OPENMP
#pragma omp parallel for
#endif
for (auto i = 0; i < X.numel(); i += C) {
// Transpose each C = GxK matrix
fbgemm::transpose_4rows(
K,
reinterpret_cast<const std::uint8_t*>(X_data + i),
reinterpret_cast<std::uint8_t*>(Y_data + i));
}
} else {
#ifdef _OPENMP
#pragma omp parallel for
#endif
for (auto i = 0; i < X.numel(); i += C) {
// Transpose each C = GxK matrix
math::Transpose(
2, dims.data(), axes.data(), X_data + i, Y_data + i, &context_);
}
}
// Even if there is a pre-chosen quantization parameters for the output,
// it is ignored because channel shuffle output quantization should be same
// as the input.
PropagateOutputTensorQuantizationParams(this, 0, in_qparams);
return true;
}
REGISTER_CPU_OPERATOR_WITH_ENGINE(
ChannelShuffle,
DNNLOWP,
ChannelShuffleDNNLowPOp<uint8_t>);
REGISTER_CPU_OPERATOR_WITH_ENGINE(
Int8ChannelShuffle,
DNNLOWP,
ChannelShuffleDNNLowPOp<uint8_t>);
REGISTER_CPU_OPERATOR_WITH_ENGINE(
ChannelShuffle,
DNNLOWP_16,
ChannelShuffleDNNLowPOp<uint16_t>);
} // namespace caffe2
|
