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
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
|
/* Copyright 2018 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.
==============================================================================*/
// TODO(ycling): Refactoring. Remove BroadcastLogical and use the more
// generalized and efficient BroadcastBinaryFunction.
//
// Also appears to duplicte MinimumMaximum.
//
// R: Result type. T1: Input 1 type. T2: Input 2 type.
template <typename R, typename T1, typename T2>
inline void BroadcastBinaryFunction4DSlow(
const RuntimeShape& unextended_input1_shape, const T1* input1_data,
const RuntimeShape& unextended_input2_shape, const T2* input2_data,
const RuntimeShape& unextended_output_shape, R* output_data,
R (* func)(T1, T2)) {
TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4);
TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
const RuntimeShape output_shape =
RuntimeShape::ExtendedShape(4, unextended_output_shape);
NdArrayDesc<4> desc1;
NdArrayDesc<4> desc2;
NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
unextended_input2_shape, &desc1, &desc2);
for (int b = 0; b < output_shape.Dims(0); ++b) {
for (int y = 0; y < output_shape.Dims(1); ++y) {
for (int x = 0; x < output_shape.Dims(2); ++x) {
for (int c = 0; c < output_shape.Dims(3); ++c) {
auto out_idx = Offset(output_shape, b, y, x, c);
auto in1_idx = SubscriptToIndex(desc1, b, y, x, c);
auto in2_idx = SubscriptToIndex(desc2, b, y, x, c);
auto in1_val = input1_data[in1_idx];
auto in2_val = input2_data[in2_idx];
output_data[out_idx] = func(in1_val, in2_val);
}
}
}
}
}
// R: Result type. T1: Input 1 type. T2: Input 2 type.
// TODO(renjieliu): Refactor other binary functions to use this one.
template <typename R, typename T1, typename T2>
inline void BinaryFunction(const RuntimeShape& input1_shape,
const T1* input1_data,
const RuntimeShape& input2_shape,
const T2* input2_data,
const RuntimeShape& output_shape, R* output_data,
R (* func)(T1, T2)) {
const int flat_size =
MatchingFlatSize(input1_shape, input2_shape, output_shape);
for (int i = 0; i < flat_size; ++i) {
output_data[i] = func(input1_data[i], input2_data[i]);
}
}
struct Sub {
static inline void Sub_(const float* input1_data, const float* input2_data,
float* output_data, const int size) {
int i = 0;
#ifdef USE_NEON
for (; i <= size - 16; i += 16) {
auto a10 = vld1q_f32(input1_data + i);
auto a11 = vld1q_f32(input1_data + i + 4);
auto a12 = vld1q_f32(input1_data + i + 8);
auto a13 = vld1q_f32(input1_data + i + 12);
auto a20 = vld1q_f32(input2_data + i);
auto a21 = vld1q_f32(input2_data + i + 4);
auto a22 = vld1q_f32(input2_data + i + 8);
auto a23 = vld1q_f32(input2_data + i + 12);
auto x0 = vsubq_f32(a10, a20);
auto x1 = vsubq_f32(a11, a21);
auto x2 = vsubq_f32(a12, a22);
auto x3 = vsubq_f32(a13, a23);
vst1q_f32(output_data + i, x0);
vst1q_f32(output_data + i + 4, x1);
vst1q_f32(output_data + i + 8, x2);
vst1q_f32(output_data + i + 12, x3);
}
for (; i <= size - 4; i += 4) {
auto a1 = vld1q_f32(input1_data + i);
auto a2 = vld1q_f32(input2_data + i);
auto x = vsubq_f32(a1, a2);
vst1q_f32(output_data + i, x);
}
#endif // NEON
for (; i < size; i++) {
output_data[i] = input1_data[i] - input2_data[i];
}
}
static inline void Call(
const float* input1_data, const RuntimeShape& in1_shape,
const float* input2_data, const RuntimeShape& in2_shape,
float* output_data, const RuntimeShape& out_shape) {
if (in1_shape != in2_shape) {
BroadcastBinaryFunction4DSlow<float, float, float>(
in1_shape, input1_data,
in2_shape, input2_data,
out_shape, output_data,
[](float a, float b) { return a - b; }
);
} else {
Sub_(input1_data, input2_data, output_data, out_shape.FlatSize());
}
}
};
struct Add {
static inline void Add_(const float* input1_data, const float* input2_data,
float* output_data, const int size) {
int i = 0;
#ifdef USE_NEON
for (; i <= size - 16; i += 16) {
auto a10 = vld1q_f32(input1_data + i);
auto a11 = vld1q_f32(input1_data + i + 4);
auto a12 = vld1q_f32(input1_data + i + 8);
auto a13 = vld1q_f32(input1_data + i + 12);
auto a20 = vld1q_f32(input2_data + i);
auto a21 = vld1q_f32(input2_data + i + 4);
auto a22 = vld1q_f32(input2_data + i + 8);
auto a23 = vld1q_f32(input2_data + i + 12);
auto x0 = vaddq_f32(a10, a20);
auto x1 = vaddq_f32(a11, a21);
auto x2 = vaddq_f32(a12, a22);
auto x3 = vaddq_f32(a13, a23);
vst1q_f32(output_data + i, x0);
vst1q_f32(output_data + i + 4, x1);
vst1q_f32(output_data + i + 8, x2);
vst1q_f32(output_data + i + 12, x3);
}
for (; i <= size - 4; i += 4) {
auto a1 = vld1q_f32(input1_data + i);
auto a2 = vld1q_f32(input2_data + i);
auto x = vaddq_f32(a1, a2);
vst1q_f32(output_data + i, x);
}
#endif // NEON
for (; i < size; i++) {
output_data[i] = input1_data[i] + input2_data[i];
}
}
static inline void Call(
const float* input1_data, const RuntimeShape& in1_shape,
const float* input2_data, const RuntimeShape& in2_shape,
float* output_data, const RuntimeShape& out_shape) {
if (in1_shape != in2_shape) {
BroadcastBinaryFunction4DSlow<float, float, float>(
in1_shape, input1_data,
in2_shape, input2_data,
out_shape, output_data,
[](float a, float b) { return a + b; }
);
} else {
Add_(input1_data, input2_data, output_data, out_shape.FlatSize());
}
}
};
struct Max {
static inline void Call(
const float* input1_data, const RuntimeShape& in1_shape,
const float* input2_data, const RuntimeShape& in2_shape,
float* output_data, const RuntimeShape& out_shape) {
if (in1_shape != in2_shape) {
BroadcastBinaryFunction4DSlow<float, float, float>(
in1_shape, input1_data,
in2_shape, input2_data,
out_shape, output_data,
[](float a, float b) { return std::max(a, b); }
);
} else {
auto input1 = MapAsVector(input1_data, in1_shape.FlatSize());
auto input2 = MapAsVector(input2_data, in2_shape.FlatSize());
auto output = MapAsVector(output_data, out_shape.FlatSize());
output = input1.cwiseMax(input2);
}
}
};
struct Mul {
static inline void Call(const float* input1_data, const RuntimeShape& in1_shape,
const float* input2_data, const RuntimeShape& in2_shape,
float* output_data, const RuntimeShape& out_shape) {
if (in1_shape != in2_shape) {
BroadcastBinaryFunction4DSlow<float, float, float>(
in1_shape, input1_data,
in2_shape, input2_data,
out_shape, output_data,
[](float a, float b) { return a * b; });
} else {
Mul_(input1_data, input2_data, output_data, out_shape.FlatSize());
}
}
static inline void Mul_(const float* input1_data, const float* input2_data,
float* output_data, const int size) {
int i = 0;
#ifdef USE_NEON
for (; i <= size - 16; i += 16) {
auto a10 = vld1q_f32(input1_data + i);
auto a11 = vld1q_f32(input1_data + i + 4);
auto a12 = vld1q_f32(input1_data + i + 8);
auto a13 = vld1q_f32(input1_data + i + 12);
auto a20 = vld1q_f32(input2_data + i);
auto a21 = vld1q_f32(input2_data + i + 4);
auto a22 = vld1q_f32(input2_data + i + 8);
auto a23 = vld1q_f32(input2_data + i + 12);
auto x0 = vmulq_f32(a10, a20);
auto x1 = vmulq_f32(a11, a21);
auto x2 = vmulq_f32(a12, a22);
auto x3 = vmulq_f32(a13, a23);
vst1q_f32(output_data + i, x0);
vst1q_f32(output_data + i + 4, x1);
vst1q_f32(output_data + i + 8, x2);
vst1q_f32(output_data + i + 12, x3);
}
for (; i <= size - 4; i += 4) {
auto a1 = vld1q_f32(input1_data + i);
auto a2 = vld1q_f32(input2_data + i);
auto x = vmulq_f32(a1, a2);
vst1q_f32(output_data + i, x);
}
#endif // NEON
for (; i < size; i++) {
output_data[i] = input1_data[i] * input2_data[i];
}
}
};
//TODO maybe move to a separate file since everything else here is extracted from TF Lite
//23.11.2018
struct Div {
static inline void Call(
const float* input1_data, const RuntimeShape& in1_shape,
const float* input2_data, const RuntimeShape& in2_shape,
float* output_data, const RuntimeShape& out_shape) {
if (in1_shape != in2_shape) {
BroadcastBinaryFunction4DSlow<float, float, float>(
in1_shape, input1_data,
in2_shape, input2_data,
out_shape, output_data,
[](float a, float b) { return a / b; }
);
} else {
auto input1 = MapAsVector(input1_data, in1_shape.FlatSize());
auto input2 = MapAsVector(input2_data, in2_shape.FlatSize());
auto output = MapAsVector(output_data, out_shape.FlatSize());
output = input1.cwiseQuotient(input2);
}
}
};
|
