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 2017 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 "tensorflow/contrib/lite/c/builtin_op_data.h"
#include "tensorflow/contrib/lite/c/c_api_internal.h"
#include "tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h"
#include "tensorflow/contrib/lite/kernels/internal/quantization_util.h"
#include "tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h"
#include "tensorflow/contrib/lite/kernels/internal/tensor.h"
#include "tensorflow/contrib/lite/kernels/kernel_util.h"
#include "tensorflow/contrib/lite/kernels/op_macros.h"
namespace tflite {
namespace ops {
namespace builtin {
namespace mul {
// This file has three implementation of Mul.
enum KernelType {
kReference,
kGenericOptimized, // Neon-free
kNeonOptimized,
};
constexpr int kInputTensor1 = 0;
constexpr int kInputTensor2 = 1;
constexpr int kOutputTensor = 0;
struct OpData {
bool requires_broadcast;
// Parameters used in the quantized paths where the output is 8bit
int32 output_activation_min;
int32 output_activation_max;
// Parameters used in all quantized paths
int32_t output_multiplier;
int output_shift;
};
void* Init(TfLiteContext* context, const char* buffer, size_t length) {
auto* data = new OpData;
data->requires_broadcast = false;
return data;
}
void Free(TfLiteContext* context, void* buffer) {
delete reinterpret_cast<OpData*>(buffer);
}
TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
auto* params = reinterpret_cast<TfLiteMulParams*>(node->builtin_data);
OpData* data = reinterpret_cast<OpData*>(node->user_data);
TF_LITE_ENSURE_EQ(context, NumInputs(node), 2);
TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);
const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
TF_LITE_ENSURE_EQ(context, input1->type, input2->type);
data->requires_broadcast = !HaveSameShapes(input1, input2);
TfLiteIntArray* output_size = nullptr;
if (data->requires_broadcast) {
TF_LITE_ENSURE_OK(context, CalculateShapeForBroadcast(
context, input1, input2, &output_size));
} else {
output_size = TfLiteIntArrayCopy(input1->dims);
}
if (output->type == kTfLiteUInt8) {
CalculateActivationRangeUint8(params->activation, output,
&data->output_activation_min,
&data->output_activation_max);
}
if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt16) {
double real_multiplier =
input1->params.scale * input2->params.scale / output->params.scale;
QuantizeMultiplierSmallerThanOneExp(
real_multiplier, &data->output_multiplier, &data->output_shift);
}
return context->ResizeTensor(context, output, output_size);
}
template <KernelType kernel_type>
void EvalMul(TfLiteContext* context, TfLiteNode* node, TfLiteMulParams* params,
const OpData* data, const TfLiteTensor* input1,
const TfLiteTensor* input2, TfLiteTensor* output) {
#define TF_LITE_MUL(type, opname, data_type) \
data_type output_activation_min, output_activation_max; \
CalculateActivationRange(params->activation, &output_activation_min, \
&output_activation_max); \
tflite::ArithmeticParams op_params; \
SetActivationParams(output_activation_min, output_activation_max, \
&op_params); \
type::opname(op_params, GetTensorShape(input1), \
GetTensorData<data_type>(input1), GetTensorShape(input2), \
GetTensorData<data_type>(input2), GetTensorShape(output), \
GetTensorData<data_type>(output))
if (output->type == kTfLiteInt32) {
if (kernel_type == kReference) {
if (data->requires_broadcast) {
TF_LITE_MUL(reference_ops, BroadcastMul4DSlow, int32_t);
} else {
TF_LITE_MUL(reference_ops, Mul, int32_t);
}
} else {
if (data->requires_broadcast) {
TF_LITE_MUL(optimized_ops, BroadcastMul4DSlow, int32_t);
} else {
TF_LITE_MUL(optimized_ops, Mul, int32_t);
}
}
} else if (output->type == kTfLiteFloat32) {
if (kernel_type == kReference) {
if (data->requires_broadcast) {
TF_LITE_MUL(reference_ops, BroadcastMul4DSlow, float);
} else {
TF_LITE_MUL(reference_ops, Mul, float);
}
} else {
if (data->requires_broadcast) {
TF_LITE_MUL(optimized_ops, BroadcastMul4DSlow, float);
} else {
TF_LITE_MUL(optimized_ops, Mul, float);
}
}
}
#undef TF_LITE_MUL
}
template <KernelType kernel_type>
TfLiteStatus EvalQuantized(TfLiteContext* context, TfLiteNode* node,
TfLiteMulParams* params, const OpData* data,
const TfLiteTensor* input1,
const TfLiteTensor* input2, TfLiteTensor* output) {
if (input1->type == kTfLiteUInt8 && input2->type == kTfLiteUInt8 &&
output->type == kTfLiteUInt8) {
#define TF_LITE_MUL(type, opname) \
tflite::ArithmeticParams op_params; \
SetActivationParams(data->output_activation_min, \
data->output_activation_max, &op_params); \
op_params.input1_offset = -input1->params.zero_point; \
op_params.input2_offset = -input2->params.zero_point; \
op_params.output_offset = output->params.zero_point; \
op_params.output_multiplier = data->output_multiplier; \
op_params.output_shift = data->output_shift; \
type::opname(op_params, GetTensorShape(input1), \
GetTensorData<uint8_t>(input1), GetTensorShape(input2), \
GetTensorData<uint8_t>(input2), GetTensorShape(output), \
GetTensorData<uint8_t>(output))
// The quantized version of Mul doesn't support activations, so we
// always use BroadcastMul.
if (kernel_type == kReference) {
TF_LITE_MUL(reference_ops, BroadcastMul4DSlow);
} else {
TF_LITE_MUL(optimized_ops, BroadcastMul4DSlow);
}
#undef TF_LITE_MUL
} else if (input1->type == kTfLiteInt16 && input2->type == kTfLiteInt16 &&
output->type == kTfLiteInt16) {
#define TF_LITE_MUL(type, opname) \
tflite::ArithmeticParams op_params; \
type::opname(op_params, GetTensorShape(input1), \
GetTensorData<int16_t>(input1), GetTensorShape(input2), \
GetTensorData<int16_t>(input2), GetTensorShape(output), \
GetTensorData<int16_t>(output))
if (kernel_type == kReference) {
TF_LITE_MUL(reference_ops, Mul);
} else {
TF_LITE_MUL(optimized_ops, Mul);
}
#undef TF_LITE_MUL
} else if (input1->type == kTfLiteInt16 && input2->type == kTfLiteInt16 &&
output->type == kTfLiteUInt8) {
#define TF_LITE_MUL(type, opname) \
tflite::ArithmeticParams op_params; \
SetActivationParams(data->output_activation_min, \
data->output_activation_max, &op_params); \
op_params.output_offset = output->params.zero_point; \
type::opname(op_params, GetTensorShape(input1), \
GetTensorData<int16_t>(input1), GetTensorShape(input2), \
GetTensorData<int16_t>(input2), GetTensorShape(output), \
GetTensorData<uint8_t>(output))
if (kernel_type == kReference) {
TF_LITE_MUL(reference_ops, Mul);
} else {
TF_LITE_MUL(optimized_ops, Mul);
}
#undef TF_LITE_MUL
} else {
context->ReportError(
context, "Unsupported combination of input and output types in Mul.");
return kTfLiteError;
}
return kTfLiteOk;
}
template <KernelType kernel_type>
TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
auto* params = reinterpret_cast<TfLiteMulParams*>(node->builtin_data);
OpData* data = reinterpret_cast<OpData*>(node->user_data);
const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
if (output->type == kTfLiteFloat32 || output->type == kTfLiteInt32) {
EvalMul<kernel_type>(context, node, params, data, input1, input2, output);
} else if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt16) {
TF_LITE_ENSURE_OK(
context, EvalQuantized<kernel_type>(context, node, params, data, input1,
input2, output));
} else {
context->ReportError(context,
"Mul only supports FLOAT32, INT32 and quantized UINT8 "
"and INT16 now, got %d.",
output->type);
return kTfLiteError;
}
return kTfLiteOk;
}
} // namespace mul
TfLiteRegistration* Register_MUL_REF() {
static TfLiteRegistration r = {mul::Init, mul::Free, mul::Prepare,
mul::Eval<mul::kReference>};
return &r;
}
TfLiteRegistration* Register_MUL_GENERIC_OPT() {
static TfLiteRegistration r = {mul::Init, mul::Free, mul::Prepare,
mul::Eval<mul::kGenericOptimized>};
return &r;
}
TfLiteRegistration* Register_MUL_NEON_OPT() {
static TfLiteRegistration r = {mul::Init, mul::Free, mul::Prepare,
mul::Eval<mul::kNeonOptimized>};
return &r;
}
TfLiteRegistration* Register_MUL() {
#ifdef USE_NEON
return Register_MUL_NEON_OPT();
#else
return Register_MUL_GENERIC_OPT();
#endif
}
} // namespace builtin
} // namespace ops
} // namespace tflite
|
