xref: /aosp_15_r20/external/tensorflow/tensorflow/lite/kernels/internal/reference/concatenation.h (revision b6fb3261f9314811a0f4371741dbb8839866f948)
1 /* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_
17 #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_
18 
19 #include <algorithm>
20 
21 #include "tensorflow/lite/kernels/internal/common.h"
22 #include "tensorflow/lite/kernels/internal/compatibility.h"
23 #include "tensorflow/lite/kernels/internal/cppmath.h"
24 #include "tensorflow/lite/kernels/internal/types.h"
25 
26 namespace tflite {
27 namespace reference_ops {
28 
29 template <typename Scalar>
Concatenation(const ConcatenationParams & params,const RuntimeShape * const * input_shapes,const Scalar * const * input_data,const RuntimeShape & output_shape,Scalar * output_data)30 inline void Concatenation(const ConcatenationParams& params,
31                           const RuntimeShape* const* input_shapes,
32                           const Scalar* const* input_data,
33                           const RuntimeShape& output_shape,
34                           Scalar* output_data) {
35   int axis = params.axis;
36   int inputs_count = params.inputs_count;
37   const int concat_dimensions = output_shape.DimensionsCount();
38   TFLITE_DCHECK_LT(axis, concat_dimensions);
39 
40   int64_t concat_size = 0;
41   for (int i = 0; i < inputs_count; i++) {
42     TFLITE_DCHECK_EQ(input_shapes[i]->DimensionsCount(), concat_dimensions);
43     for (int j = 0; j < concat_dimensions; j++) {
44       if (j != axis) {
45         MatchingDim(*input_shapes[i], j, output_shape, j);
46       }
47     }
48     concat_size += input_shapes[i]->Dims(axis);
49   }
50   TFLITE_DCHECK_EQ(concat_size, output_shape.Dims(axis));
51   int64_t outer_size = 1;
52   for (int i = 0; i < axis; ++i) {
53     outer_size *= output_shape.Dims(i);
54   }
55   // For all input arrays,
56   // FlatSize() = outer_size * Dims(axis) * base_inner_size;
57   int64_t base_inner_size = 1;
58   for (int i = axis + 1; i < concat_dimensions; ++i) {
59     base_inner_size *= output_shape.Dims(i);
60   }
61 
62   Scalar* output_ptr = output_data;
63   for (int k = 0; k < outer_size; k++) {
64     for (int i = 0; i < inputs_count; ++i) {
65       const int copy_size = input_shapes[i]->Dims(axis) * base_inner_size;
66       const Scalar* input_ptr = input_data[i] + k * copy_size;
67       memcpy(output_ptr, input_ptr, copy_size * sizeof(Scalar));
68       output_ptr += copy_size;
69     }
70   }
71 }
72 
73 // TODO(b/174275780): The quantized implementation of concatentation isn't fully
74 // quantized as it takes scale as a floating point value. This should be fixed
75 // when optimizng this routine further.
ConcatenationWithScaling(const ConcatenationParams & params,const RuntimeShape * const * input_shapes,const uint8_t * const * input_data,const RuntimeShape & output_shape,uint8_t * output_data)76 inline void ConcatenationWithScaling(const ConcatenationParams& params,
77                                      const RuntimeShape* const* input_shapes,
78                                      const uint8_t* const* input_data,
79                                      const RuntimeShape& output_shape,
80                                      uint8_t* output_data) {
81   int axis = params.axis;
82   const int32_t* input_zeropoint = params.input_zeropoint;
83   const float* input_scale = params.input_scale;
84   int inputs_count = params.inputs_count;
85   const int32_t output_zeropoint = params.output_zeropoint;
86   const float output_scale = params.output_scale;
87 
88   const int concat_dimensions = output_shape.DimensionsCount();
89   TFLITE_DCHECK_LT(axis, concat_dimensions);
90 
91   int64_t concat_size = 0;
92   for (int i = 0; i < inputs_count; i++) {
93     TFLITE_DCHECK_EQ(input_shapes[i]->DimensionsCount(), concat_dimensions);
94     for (int j = 0; j < concat_dimensions; j++) {
95       if (j != axis) {
96         MatchingDim(*input_shapes[i], j, output_shape, j);
97       }
98     }
99     concat_size += input_shapes[i]->Dims(axis);
100   }
101   TFLITE_DCHECK_EQ(concat_size, output_shape.Dims(axis));
102   int64_t outer_size = 1;
103   for (int i = 0; i < axis; ++i) {
104     outer_size *= output_shape.Dims(i);
105   }
106   // For all input arrays,
107   // FlatSize() = outer_size * Dims(axis) * base_inner_size;
108   int64_t base_inner_size = 1;
109   for (int i = axis + 1; i < concat_dimensions; ++i) {
110     base_inner_size *= output_shape.Dims(i);
111   }
112 
113   const float inverse_output_scale = 1.f / output_scale;
114   uint8_t* output_ptr = output_data;
115   for (int k = 0; k < outer_size; k++) {
116     for (int i = 0; i < inputs_count; ++i) {
117       const int copy_size = input_shapes[i]->Dims(axis) * base_inner_size;
118       const uint8_t* input_ptr = input_data[i] + k * copy_size;
119       if (input_zeropoint[i] == output_zeropoint &&
120           input_scale[i] == output_scale) {
121         memcpy(output_ptr, input_ptr, copy_size);
122       } else {
123         const float scale = input_scale[i] * inverse_output_scale;
124         const float bias = -input_zeropoint[i] * scale;
125         for (int j = 0; j < copy_size; ++j) {
126           const int32_t value = static_cast<int32_t>(tflite::TfLiteRound(
127                                     input_ptr[j] * scale + bias)) +
128                                 output_zeropoint;
129           output_ptr[j] = static_cast<uint8_t>(
130               std::max<int32_t>(std::min<int32_t>(255, value), 0));
131         }
132       }
133       output_ptr += copy_size;
134     }
135   }
136 }
137 
138 }  // namespace reference_ops
139 }  // namespace tflite
140 
141 #endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_
142