• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #include "FullyConnectedLayer.h"
25 
26 #include "arm_compute/core/Types.h"
27 #include "tests/validation/reference/UtilsQuantizedAsymm.h"
28 
29 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
30 
31 #include <numeric>
32 
33 namespace arm_compute
34 {
35 namespace test
36 {
37 namespace validation
38 {
39 namespace reference
40 {
41 namespace
42 {
43 // Vector matrix multiply for floating point
44 template < typename T, typename TB, typename std::enable_if < is_floating_point<T>::value &&is_floating_point<TB>::value, int >::type = 0 >
vector_matrix_multiply(const SimpleTensor<T> & src,const SimpleTensor<T> & weights,const SimpleTensor<TB> & bias,SimpleTensor<T> & dst,int offset_src,int offset_dst,int cols_weights,int rows_weights)45 void vector_matrix_multiply(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &dst, int offset_src, int offset_dst, int cols_weights,
46                             int rows_weights)
47 {
48     const T *src_ptr     = src.data() + offset_src;
49     const T *weights_ptr = weights.data();
50     const TB *bias_ptr    = bias.data();
51     T        *dst_ptr     = dst.data() + offset_dst;
52 #if defined(_OPENMP)
53     #pragma omp parallel for
54 #endif /* _OPENMP */
55     for(int y = 0; y < rows_weights; ++y)
56     {
57         dst_ptr[y] = std::inner_product(src_ptr, src_ptr + cols_weights, &weights_ptr[cols_weights * y], static_cast<T>(0)) + bias_ptr[y];
58     }
59 }
60 
61 // Vector matrix multiply for quantized type
62 template < typename T, typename TB, typename std::enable_if < (std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value) &&std::is_same<TB, int32_t>::value, int >::type = 0 >
vector_matrix_multiply(const SimpleTensor<T> & src,const SimpleTensor<T> & weights,const SimpleTensor<TB> & bias,SimpleTensor<T> & dst,int offset_src,int offset_dst,int cols_weights,int rows_weights)63 void vector_matrix_multiply(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &dst, int offset_src, int offset_dst,
64                             int cols_weights, int rows_weights)
65 {
66     const T *src_ptr     = src.data() + offset_src;
67     const T *weights_ptr = weights.data();
68     const TB *bias_ptr    = bias.data();
69     T        *dst_ptr     = dst.data() + offset_dst;
70 
71     const UniformQuantizationInfo iq_info = src.quantization_info().uniform();
72     const UniformQuantizationInfo wq_info = weights.quantization_info().uniform();
73     const UniformQuantizationInfo oq_info = dst.quantization_info().uniform();
74 
75     const int   input_offset   = -iq_info.offset;
76     const float input_scale    = iq_info.scale;
77     const int   weights_offset = -wq_info.offset;
78     const float weights_scale  = wq_info.scale;
79     const int   output_offset  = oq_info.offset;
80     const float output_scale   = oq_info.scale;
81 
82     int         output_multiplier = 0;
83     int         output_shift      = 0;
84     const float multiplier        = input_scale * weights_scale / output_scale;
85     arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
86 
87     const int min = std::numeric_limits<T>::lowest();
88     const int max = std::numeric_limits<T>::max();
89 #if defined(_OPENMP)
90     #pragma omp parallel for
91 #endif /* _OPENMP */
92     for(int y = 0; y < rows_weights; ++y)
93     {
94         // Reset accumulator
95         int32_t acc = 0;
96 
97         for(int x = 0; x < cols_weights; ++x)
98         {
99             acc += (src_ptr[x] + input_offset) * (weights_ptr[x + y * cols_weights] + weights_offset);
100         }
101 
102         // Accumulate the bias
103         acc += bias_ptr[y];
104 
105         // Quantize down
106         acc = quantize_down_scale_by_fixedpoint(acc, output_multiplier, output_shift, output_offset, min, max);
107 
108         // Store the result
109         dst_ptr[y] = static_cast<T>(acc);
110     }
111 }
112 } // namespace
113 
114 template <typename T, typename TB>
fully_connected_layer(const SimpleTensor<T> & src,const SimpleTensor<T> & weights,const SimpleTensor<TB> & bias,const TensorShape & dst_shape,QuantizationInfo out_quant_info)115 SimpleTensor<T> fully_connected_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, const TensorShape &dst_shape, QuantizationInfo out_quant_info)
116 {
117     // if no explicit quantization has been set you the same as src
118     if(out_quant_info == QuantizationInfo())
119     {
120         out_quant_info = src.quantization_info();
121     }
122 
123     // Create reference
124     SimpleTensor<T> dst{ TensorShape{ dst_shape }, src.data_type(), 1, out_quant_info };
125 
126     // Health checks
127     const int          num_batch_dimensions = std::max(0, static_cast<int>(dst_shape.num_dimensions()) - 1);
128     const int          num_input_dimensions = src.shape().num_dimensions() - num_batch_dimensions;
129     const unsigned int linear_input_size    = src.shape().total_size_lower(num_input_dimensions);
130 
131     ARM_COMPUTE_UNUSED(num_batch_dimensions);
132     ARM_COMPUTE_UNUSED(num_input_dimensions);
133     ARM_COMPUTE_UNUSED(linear_input_size);
134     ARM_COMPUTE_ERROR_ON(weights.shape().x() != linear_input_size);
135     ARM_COMPUTE_ERROR_ON(weights.shape().y() != bias.shape().x());
136     ARM_COMPUTE_ERROR_ON(weights.shape().y() != dst.shape().x());
137 
138     // Compute reference
139     const int cols_weights = weights.shape().x();
140     const int rows_weights = weights.shape().y();
141     const int num_batches  = dst_shape.total_size_upper(1);
142 
143     for(int k = 0; k < num_batches; ++k)
144     {
145         const int offset_in  = k * cols_weights;
146         const int offset_out = k * rows_weights;
147 
148         vector_matrix_multiply<T>(src,
149                                   weights,
150                                   bias,
151                                   dst,
152                                   offset_in,
153                                   offset_out,
154                                   cols_weights,
155                                   rows_weights);
156     }
157 
158     return dst;
159 }
160 
161 template SimpleTensor<float> fully_connected_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &bias, const TensorShape &dst_shape,
162                                                    QuantizationInfo out_quant_info);
163 template SimpleTensor<half> fully_connected_layer(const SimpleTensor<half> &src, const SimpleTensor<half> &weights, const SimpleTensor<half> &bias, const TensorShape &dst_shape,
164                                                   QuantizationInfo out_quant_info);
165 template SimpleTensor<uint8_t> fully_connected_layer(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &bias, const TensorShape &dst_shape,
166                                                      QuantizationInfo out_quant_info);
167 template SimpleTensor<int8_t> fully_connected_layer(const SimpleTensor<int8_t> &src, const SimpleTensor<int8_t> &weights, const SimpleTensor<int32_t> &bias, const TensorShape &dst_shape,
168                                                     QuantizationInfo out_quant_info);
169 } // namespace reference
170 } // namespace validation
171 } // namespace test
172 } // namespace arm_compute
173