1 /*
2 * SPDX-FileCopyrightText: Copyright 2023-2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
3 *
4 * SPDX-License-Identifier: Apache-2.0
5 *
6 * Licensed under the Apache License, Version 2.0 (the License); you may
7 * not use this file except in compliance with the License.
8 * You may obtain a copy of the License at
9 *
10 * www.apache.org/licenses/LICENSE-2.0
11 *
12 * Unless required by applicable law or agreed to in writing, software
13 * distributed under the License is distributed on an AS IS BASIS, WITHOUT
14 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15 * See the License for the specific language governing permissions and
16 * limitations under the License.
17 */
18
19 /* ----------------------------------------------------------------------
20 * Project: CMSIS NN Library
21 * Title: arm_vector_sum_s8_s64
22 * Description: Generic function for calculating vector sums
23 *
24 * $Date: 26 March 2024
25 * $Revision: V.1.0.0
26 *
27 * Target : Arm(R) M-Profile Architecture
28 *
29 * -------------------------------------------------------------------- */
30
31 #include "arm_nnfunctions.h"
32 #include "arm_nnsupportfunctions.h"
33 /**
34 * @ingroup Public
35 */
36
37 /**
38 * @addtogroup FC
39 * @{
40 */
41
42 /*
43 * S8 vector sum fuction in preparation for e.g. kernel sums in fully connected and matrix multiplication layer function
44 *
45 * Refer header file for details.
46 *
47 */
arm_vector_sum_s8_s64(int64_t * vector_sum_buf,const int32_t vector_cols,const int32_t vector_rows,const int8_t * vector_data,const int32_t lhs_offset,const int64_t * bias_data)48 arm_cmsis_nn_status arm_vector_sum_s8_s64(int64_t *vector_sum_buf,
49 const int32_t vector_cols,
50 const int32_t vector_rows,
51 const int8_t *vector_data,
52 const int32_t lhs_offset,
53 const int64_t *bias_data)
54 {
55
56 if (bias_data)
57 {
58 memcpy(vector_sum_buf, bias_data, vector_rows * sizeof(int64_t));
59 }
60 else
61 {
62 memset(vector_sum_buf, 0, vector_rows * sizeof(int64_t));
63 }
64 if (lhs_offset)
65 {
66 #if defined(ARM_MATH_MVEI)
67
68 const int32_t row_loop_cnt = vector_rows / 5;
69 for (int i_row_loop_cnt = 0; i_row_loop_cnt < row_loop_cnt; i_row_loop_cnt++)
70 {
71 const int32_t col_loop_cnt = (vector_cols + 15) / 16;
72 const int8_t *vector_0 = vector_data;
73 const int8_t *vector_1 = vector_data + vector_cols;
74 const int8_t *vector_2 = vector_data + 2 * vector_cols;
75 const int8_t *vector_3 = vector_data + 3 * vector_cols;
76 const int8_t *vector_4 = vector_data + 4 * vector_cols;
77 int32_t vector_sum_0 = 0;
78 int32_t vector_sum_1 = 0;
79 int32_t vector_sum_2 = 0;
80 int32_t vector_sum_3 = 0;
81 int32_t vector_sum_4 = 0;
82 uint32_t col_cnt = (uint32_t)vector_cols;
83 for (int i = 0; i < col_loop_cnt; i++)
84 {
85 mve_pred16_t p = vctp8q(col_cnt);
86 col_cnt -= 16;
87 const int8x16_t ker_0 = vldrbq_z_s8(vector_0, p);
88 vector_sum_0 = vaddvaq_s8(vector_sum_0, ker_0);
89 const int8x16_t ker_1 = vldrbq_z_s8(vector_1, p);
90 vector_sum_1 = vaddvaq_s8(vector_sum_1, ker_1);
91 const int8x16_t ker_2 = vldrbq_z_s8(vector_2, p);
92 vector_sum_2 = vaddvaq_s8(vector_sum_2, ker_2);
93 const int8x16_t ker_3 = vldrbq_z_s8(vector_3, p);
94 vector_sum_3 = vaddvaq_s8(vector_sum_3, ker_3);
95 const int8x16_t ker_4 = vldrbq_z_s8(vector_4, p);
96 vector_sum_4 = vaddvaq_s8(vector_sum_4, ker_4);
97 vector_0 += 16;
98 vector_1 += 16;
99 vector_2 += 16;
100 vector_3 += 16;
101 vector_4 += 16;
102 }
103 vector_data += 5 * vector_cols;
104
105 vector_sum_0 *= lhs_offset;
106 vector_sum_1 *= lhs_offset;
107 vector_sum_2 *= lhs_offset;
108 vector_sum_3 *= lhs_offset;
109 vector_sum_4 *= lhs_offset;
110
111 vector_sum_buf[0] += vector_sum_0;
112 vector_sum_buf[1] += vector_sum_1;
113 vector_sum_buf[2] += vector_sum_2;
114 vector_sum_buf[3] += vector_sum_3;
115 vector_sum_buf[4] += vector_sum_4;
116 vector_sum_buf += 5;
117 }
118 const int32_t loop_cnt = vector_rows % 5;
119 for (int i_row_loop_cnt = 0; i_row_loop_cnt < loop_cnt; i_row_loop_cnt++)
120 {
121 const int32_t col_loop_cnt = (vector_cols + 15) / 16;
122 const int8_t *vector_0 = vector_data;
123 int32_t vector_sum_0 = 0;
124 uint32_t col_cnt = (uint32_t)vector_cols;
125 for (int i = 0; i < col_loop_cnt; i++)
126 {
127 mve_pred16_t p = vctp8q(col_cnt);
128 col_cnt -= 16;
129 const int8x16_t ker_0 = vldrbq_z_s8(vector_0, p);
130 vector_sum_0 = vaddvaq_s8(vector_sum_0, ker_0);
131 vector_0 += 16;
132 }
133 vector_data += vector_cols;
134 vector_sum_0 *= lhs_offset;
135
136 vector_sum_buf[i_row_loop_cnt] += vector_sum_0;
137 }
138 #else
139 for (int i = 0; i < vector_rows; i++)
140 {
141 int64_t sum = 0;
142 for (int j = 0; j < vector_cols; j++)
143 {
144 sum += *vector_data++;
145 }
146 *vector_sum_buf++ += sum * (int64_t)lhs_offset;
147 }
148 #endif
149 }
150
151 return (ARM_CMSIS_NN_SUCCESS);
152 }
153
154 /**
155 * @} end of FC group
156 */
157
