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
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
|
/*
* Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved
* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
*
* 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 "FullyConnected.h"
#include "Common.h"
#include "QuantizationHelpers.h"
#include "mir/Tensor.h"
namespace mir_interpreter
{
template <typename T>
static void fullyConnected2D(const mir::TensorVariant &input, const mir::TensorVariant &weights,
mir::TensorVariant &output)
{
assert(input.getShape().rank() == 2);
assert(weights.getShape().rank() == 2);
assert(input.getShape().dim(1) == weights.getShape().dim(0));
auto in_raw = reinterpret_cast<T *>(input.atOffset(0));
auto weight_raw = reinterpret_cast<T *>(weights.atOffset(0));
auto output_raw = reinterpret_cast<T *>(output.atOffset(0));
auto rows = output.getShape().dim(0);
auto cols = output.getShape().dim(1);
auto N = input.getShape().dim(1);
auto wcols = weights.getShape().dim(1);
for (int32_t r = 0; r < rows; ++r)
{
for (int32_t k = 0; k < N; ++k)
{
auto in = in_raw[r * N + k];
for (int32_t c = 0; c < cols; ++c)
{
output_raw[r * cols + c] += in * weight_raw[k * wcols + c];
}
}
}
}
template <typename T> struct FullyConnectedImpl
{
static void run(const mir::TensorVariant &inputv, const mir::TensorVariant &weightsv,
const mir::ops::FullyConnectedOp &op, mir::TensorVariant &res,
const mir::TensorVariant *biasv);
};
template <typename T>
void FullyConnectedImpl<T>::run(const mir::TensorVariant &inputv,
const mir::TensorVariant &weightsv,
const mir::ops::FullyConnectedOp &op, mir::TensorVariant &res,
const mir::TensorVariant *biasv)
{
if (biasv)
{
throw std::runtime_error("non-quantized FullyConnected with fused bias is unsupported");
}
mir::Tensor<T> input{inputv};
mir::Tensor<T> weights{weightsv};
erase<T>(res);
if (input.getShape().rank() == 2 && weights.getShape().rank() == 2 && res.getShape().rank() == 2)
{
// optimized case for 2d matrix multiplication
fullyConnected2D<T>(inputv, weightsv, res);
return;
}
mir::Tensor<T> accessor(res);
const mir::Shape &in_shape = input.getShape();
int32_t in_rank = in_shape.rank();
const mir::Shape &w_shape = weights.getShape();
int32_t w_rank = w_shape.rank();
assert(in_shape.dim(in_rank - 1) == w_shape.dim(w_rank - 2));
(void)in_rank;
mir::ShapeRange out_range(res.getShape());
int32_t len = w_shape.dim(w_rank - 2);
for (auto &out_index : out_range)
{
mir::Index t_index = out_index;
T &output_element = accessor.at(out_index);
int32_t col = t_index.at(w_rank - 1);
int32_t row = t_index.at(w_rank - 2);
for (int32_t i = 0; i < len; ++i)
{
t_index.at(w_rank - 1) = i;
T in = input.at(t_index);
t_index.at(w_rank - 1) = col;
t_index.at(w_rank - 2) = i;
T w = weights.at(t_index);
t_index.at(w_rank - 2) = row;
output_element += in * w;
}
}
}
template <> struct FullyConnectedImpl<uint8_t>
{
static void run(const mir::TensorVariant &inputv, const mir::TensorVariant &weightsv,
const mir::ops::FullyConnectedOp &op, mir::TensorVariant &res,
const mir::TensorVariant *biasv);
};
void FullyConnectedImpl<uint8_t>::run(const mir::TensorVariant &inputv,
const mir::TensorVariant &weightsv,
const mir::ops::FullyConnectedOp &op, mir::TensorVariant &res,
const mir::TensorVariant *biasv)
{
if (!biasv)
{
throw std::runtime_error{"Quantized FullyConnected cannot be executed without fused bias"};
}
const auto &input_type = inputv.getType();
const auto &weights_type = weightsv.getType();
const auto &bias_type = biasv->getType();
const auto &output_type = op.getOutput(0)->getType();
(void)bias_type;
assert(input_type.isQuantized());
assert(weights_type.isQuantized());
assert(bias_type.isQuantized());
assert(output_type.isQuantized());
assert(input_type.getElementType() == mir::DataType::UINT8);
assert(weights_type.getElementType() == mir::DataType::UINT8);
assert(bias_type.getElementType() == mir::DataType::INT32);
int32_t input_offset = -input_type.getQuantization().getZeroPoint();
int32_t weights_offset = -weights_type.getQuantization().getZeroPoint();
int32_t output_offset = output_type.getQuantization().getZeroPoint();
double input_scale = input_type.getQuantization().getScale();
double weights_scale = weights_type.getQuantization().getScale();
double output_scale = output_type.getQuantization().getScale();
double real_multiplier = input_scale * weights_scale / output_scale;
int32_t output_multiplier = 0;
int output_shift = 0;
QuantizeMultiplier(real_multiplier, &output_multiplier, &output_shift);
const mir::Shape &in_shape = inputv.getShape();
const mir::Shape &weights_shape = weightsv.getShape();
const mir::Shape &out_shape = op.getOutputShape(0);
const int32_t batches = in_shape.dim(0);
assert(in_shape.rank() == 2);
assert(weights_shape.rank() == 2);
assert(in_shape.dim(1) == weights_shape.dim(0));
const int32_t accum_depth = weights_shape.dim(0);
const int32_t output_depth = weights_shape.dim(1);
uint8_t *input_data = reinterpret_cast<uint8_t *>(inputv.atOffset(0));
uint8_t *weights_data = reinterpret_cast<uint8_t *>(weightsv.atOffset(0));
int32_t *bias_data = reinterpret_cast<int32_t *>(biasv->atOffset(0));
uint8_t *output_data = reinterpret_cast<uint8_t *>(res.atOffset(0));
int32_t output_min = std::numeric_limits<uint8_t>::min();
int32_t output_max = std::numeric_limits<uint8_t>::max();
for (int32_t b = 0; b < batches; ++b)
{
for (int32_t out_c = 0; out_c < output_depth; ++out_c)
{
int32_t acc = 0;
for (int d = 0; d < accum_depth; ++d)
{
int32_t input_val = input_data[b * accum_depth + d];
int32_t weights_val = weights_data[d * output_depth + out_c];
acc += (weights_val + weights_offset) * (input_val + input_offset);
}
acc += bias_data[out_c];
acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
acc += output_offset;
acc = std::max(acc, output_min);
acc = std::min(acc, output_max);
output_data[out_c + output_depth * b] = static_cast<uint8_t>(acc);
}
}
}
void FullyConnected(const mir::TensorVariant &input, const mir::TensorVariant &weights,
const mir::ops::FullyConnectedOp &op, mir::TensorVariant &res,
const mir::TensorVariant *bias)
{
dispatch<FullyConnectedImpl>(res.getElementType(), input, weights, op, res, bias);
}
} // namespace mir_interpreter
|
