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
|
// Copyright (c) 2020-2022, ARM Limited.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "reduction.h"
#include "quant_util.h"
using namespace TosaReference;
using namespace Eigen;
using namespace tosa;
template <int Rank, DType Dtype>
ReduceNode<Rank, Dtype>::ReduceNode(SubgraphTraverser* sgt_, const Op& op_, TosaAttributeBase* attribute_, uint64_t id_)
: GraphNode(sgt_, op_, id_)
{
setRequiredOperands(1, 1);
setRequiredRank(0, 4);
INIT_ATTRIBUTE(Axis);
}
template <int Rank, DType Dtype>
ReduceNode<Rank, Dtype>::~ReduceNode()
{
if (attribute)
delete attribute;
}
template <int Rank, DType Dtype>
int ReduceNode<Rank, Dtype>::checkTensorAttributes()
{
if (validateRequiredOperands())
return 1;
if (validateRequiredRank(inputs[0]) || validateRequiredRank(outputs[0]))
{
return 1;
}
if (attribute->axis() < 0 || attribute->axis() >= inputs[0]->getRank())
{
printNodeValidationError("ReduceOp: axis must between [0, input_rank - 1]");
return 1;
}
if (inputs[0]->matchRankType(*outputs[0]))
{
printNodeValidationError("ReduceOp: Input and output tensor ranks must match");
return 1;
}
if (outputs[0]->getShape()[attribute->axis()] != 1)
{
printNodeValidationError("ReduceOp: Output tensor shape[axis] needs to be 1.");
return 1;
}
in = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[0]);
out = dynamic_cast<TosaReference::TensorTemplate<TOut>*>(outputs[0]);
if ((!in) || (!out))
{
printNodeValidationError("ReduceOp: Input or output fail to cast to Eigen tensor since rank/type not expected");
return 1;
}
dims[0] = this->attribute->axis();
return 0;
}
template <int Rank, DType Dtype>
int OpReduceAll<Rank, Dtype>::eval()
{
this->out->getTensor() = this->in->getTensor().all(this->dims).reshape(this->out->getTensor().dimensions());
return GraphNode::eval();
}
template <int Rank, DType Dtype>
int OpReduceAny<Rank, Dtype>::eval()
{
this->out->getTensor() = this->in->getTensor().any(this->dims).reshape(this->out->getTensor().dimensions());
return GraphNode::eval();
}
template <int Rank, DType Dtype>
int OpReduceMax<Rank, Dtype>::eval()
{
this->out->getTensor() = this->in->getTensor().maximum(this->dims).reshape(this->out->getTensor().dimensions());
return GraphNode::eval();
}
template <int Rank, DType Dtype>
int OpReduceMin<Rank, Dtype>::eval()
{
this->out->getTensor() = this->in->getTensor().minimum(this->dims).reshape(this->out->getTensor().dimensions());
return GraphNode::eval();
}
template <int Rank, DType Dtype>
int OpReduceProduct<Rank, Dtype>::eval()
{
this->out->getTensor() = this->in->getTensor().prod(this->dims).reshape(this->out->getTensor().dimensions());
return GraphNode::eval();
}
template <int Rank, DType Dtype>
int OpReduceSum<Rank, Dtype>::eval()
{
this->out->getTensor() = this->in->getTensor().sum(this->dims).reshape(this->out->getTensor().dimensions());
return GraphNode::eval();
}
struct SumRequiresReducer {
static const bool PacketAccess = false;
SumRequiresReducer(SubgraphTraverser* parent_sgt) : parent_sgt(parent_sgt) {}
void reduce(const int32_t val, int32_t* accum) {
int64_t res_in_64 = static_cast<int64_t>(*accum) + val;
int64_t i32_max_in_64 = static_cast<int64_t>(std::numeric_limits<int32_t>::max());
int64_t i32_min_in_64 = static_cast<int64_t>(std::numeric_limits<int32_t>::min());
REQUIRE(res_in_64 <= i32_max_in_64 && res_in_64 >= i32_min_in_64, "OpReduceSum: result not in i32 range");
*accum = static_cast<int32_t>(res_in_64);
}
int32_t initialize() const { return 0; }
int32_t finalize(const int32_t accum) const { return accum; }
private:
SubgraphTraverser* parent_sgt;
};
template <int Rank, DType Dtype>
int OpReduceSumInt<Rank, Dtype>::eval()
{
this->out->getTensor() = this->in->getTensor().reduce(this->dims, SumRequiresReducer(this->parent_sgt)).reshape(this->out->getTensor().dimensions());
return GraphNode::eval();
}
// template explicit instantiation
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceAll, BOOL);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceAny, BOOL);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, FP16);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, FP32);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, INT8);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, INT16);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, INT32);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, FP16);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, FP32);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, INT8);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, INT16);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, INT32);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceProduct, FP16);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceProduct, FP32);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSum, FP16);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSum, FP32);
DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSumInt, INT32);
|
