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 #ifndef ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H
25 #define ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H
26
27 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
28 #include "support/Requires.h"
29 #include "tests/validation/Helpers.h"
30 #include "tests/validation/reference/UtilsQuantizedAsymm.h"
31
32 namespace arm_compute
33 {
34 namespace test
35 {
36 namespace convolution_3d
37 {
38 namespace detail
39 {
is_valid_pixel(int i,int min,int max)40 inline bool is_valid_pixel(int i, int min, int max)
41 {
42 return (i >= min && i < max);
43 }
44
45 // 3D convolution for floating point type
46 template < typename T, typename TW, typename TB, typename std::enable_if < validation::is_floating_point<T>::value &&validation::is_floating_point<TW>::value
47 &&validation::is_floating_point<TB>::value,
48 int >::type = 0 >
49 inline void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<TW> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &out,
50 int i_offset, int w_offset, int b_offset, int o_offset,
51 int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights, int dilation_x = 1, int dilation_y = 1, int filter_id = 0)
52 {
53 ARM_COMPUTE_UNUSED(filter_id);
54 const T *in_ptr = in.data() + i_offset;
55 const TW *w_ptr = weights.data() + w_offset;
56 const TB *b_ptr = bias.data() + b_offset;
57 T *out_ptr = out.data() + o_offset;
58
59 const int half_width_weights_start = width_weights / 2;
60 const int half_width_weights_end = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start;
61 const int half_height_weights_start = height_weights / 2;
62 const int half_height_weights_end = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start;
63
64 // Reset accumulator
65 T acc(0);
66
67 // Compute a 2D convolution for each IFM and accumulate the result
68 for(int ifm = 0; ifm < depth_in; ++ifm)
69 {
70 // Compute the offset for the input slice
71 const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in;
72
73 // Compute 2D convolution
74 for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk)
75 {
76 for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk)
77 {
78 // Check if the pixel is out-of-bound
79 if(is_valid_pixel(xi + xk * dilation_x, 0, width_in) && is_valid_pixel(yi + yk * dilation_y, 0, height_in))
80 {
81 const int idx = xk + half_width_weights_start;
82 const int idy = yk + half_height_weights_start;
83
84 const T i_value = in_ptr[offset_slice_in + xk * dilation_x + yk * dilation_y * width_in];
85 const TW w_value = w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights];
86
87 acc += i_value * w_value;
88 }
89 }
90 }
91 }
92
93 // Accumulate the bias and store the result
94 *out_ptr = acc + (*b_ptr);
95 }
96
97 // 3D convolution for QASYMM8 type
98 template < typename T, typename TW, typename TB, ARM_COMPUTE_REQUIRES_TA((std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value) &&(std::is_same<TW, uint8_t>::value
99 || std::is_same<TW, int8_t>::value)) >
100 inline void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<TW> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &out,
101 int i_offset, int w_offset, int b_offset, int o_offset,
102 int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights, int dilation_x = 1, int dilation_y = 1, int filter_id = 0)
103 {
104 const T *in_ptr = in.data() + i_offset;
105 const TW *w_ptr = weights.data() + w_offset;
106 const TB *b_ptr = bias.data() + b_offset;
107 T *out_ptr = out.data() + o_offset;
108
109 const UniformQuantizationInfo iq_info = in.quantization_info().uniform();
110 const UniformQuantizationInfo wq_info = weights.quantization_info().uniform();
111 const UniformQuantizationInfo oq_info = out.quantization_info().uniform();
112
113 const int input_offset = -iq_info.offset;
114 const float input_scale = iq_info.scale;
115 int weights_offset = -wq_info.offset;
116 float weights_scale = wq_info.scale;
117 if(is_data_type_quantized_per_channel(weights.data_type()))
118 {
119 if(is_data_type_quantized_asymmetric(weights.data_type()))
120 {
121 weights_offset = weights.quantization_info().offset()[filter_id];
122 }
123 else
124 {
125 weights_offset = 0;
126 }
127 weights_scale = weights.quantization_info().scale()[filter_id];
128 }
129 const int output_offset = oq_info.offset;
130 const float output_scale = oq_info.scale;
131
132 int output_multiplier = 0;
133 int output_shift = 0;
134 const float multiplier = input_scale * weights_scale / output_scale;
135 arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
136
137 const int half_width_weights_start = width_weights / 2;
138 const int half_width_weights_end = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start;
139 const int half_height_weights_start = height_weights / 2;
140 const int half_height_weights_end = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start;
141
142 // Reset accumulator
143 int32_t acc(0);
144
145 // Compute a 2D convolution for each IFM and accumulate the result
146 for(int ifm = 0; ifm < depth_in; ++ifm)
147 {
148 // Compute the offset for the input slice
149 const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in;
150
151 // Compute 2D convolution
152 for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk)
153 {
154 for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk)
155 {
156 // Check if the pixel is out-of-bound
157 if(is_valid_pixel(xi + xk * dilation_x, 0, width_in) && is_valid_pixel(yi + yk * dilation_y, 0, height_in))
158 {
159 const int idx = xk + half_width_weights_start;
160 const int idy = yk + half_height_weights_start;
161
162 const int32_t i_value = in_ptr[offset_slice_in + xk * dilation_x + yk * dilation_y * width_in];
163 const int32_t w_value = w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights];
164 acc += (i_value + input_offset) * (w_value + weights_offset);
165 }
166 }
167 }
168 }
169
170 // Accumulate the bias
171 acc += (*b_ptr);
172
173 // Quantize down
174 acc = validation::quantize_down_scale_by_fixedpoint(acc, output_multiplier, output_shift, output_offset,
175 std::numeric_limits<T>::lowest(), std::numeric_limits<T>::max());
176
177 // Store the result
178 *out_ptr = acc;
179 }
180 } // namespace detail
181 } // namespace convolution_3d
182 } // namespace test
183 } // namespace arm_compute
184 #endif /* ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H */
185