1
2 /* ----------------------------------------------------------------------
3 * Project: CMSIS DSP Library
4 * Title: arm_cityblock_distance_f16.c
5 * Description: Cityblock (Manhattan) distance between two vectors
6 *
7 * $Date: 23 April 2021
8 * $Revision: V1.9.0
9 *
10 * Target Processor: Cortex-M and Cortex-A cores
11 * -------------------------------------------------------------------- */
12 /*
13 * Copyright (C) 2010-2021 ARM Limited or its affiliates. All rights reserved.
14 *
15 * SPDX-License-Identifier: Apache-2.0
16 *
17 * Licensed under the Apache License, Version 2.0 (the License); you may
18 * not use this file except in compliance with the License.
19 * You may obtain a copy of the License at
20 *
21 * www.apache.org/licenses/LICENSE-2.0
22 *
23 * Unless required by applicable law or agreed to in writing, software
24 * distributed under the License is distributed on an AS IS BASIS, WITHOUT
25 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
26 * See the License for the specific language governing permissions and
27 * limitations under the License.
28 */
29
30 #include "dsp/distance_functions_f16.h"
31
32 #if defined(ARM_FLOAT16_SUPPORTED)
33
34 #include <limits.h>
35 #include <math.h>
36
37 /**
38 @ingroup FloatDist
39 */
40
41 /**
42 @defgroup Manhattan Cityblock (Manhattan) distance
43
44 Cityblock (Manhattan) distance
45 */
46
47 /**
48 @addtogroup Manhattan
49 @{
50 */
51
52
53 /**
54 * @brief Cityblock (Manhattan) distance between two vectors
55 * @param[in] pA First vector
56 * @param[in] pB Second vector
57 * @param[in] blockSize vector length
58 * @return distance
59 *
60 */
61 #if defined(ARM_MATH_MVE_FLOAT16) && !defined(ARM_MATH_AUTOVECTORIZE)
62
63 #include "arm_helium_utils.h"
64 #include "arm_vec_math.h"
65
arm_cityblock_distance_f16(const float16_t * pA,const float16_t * pB,uint32_t blockSize)66 float16_t arm_cityblock_distance_f16(const float16_t *pA,const float16_t *pB, uint32_t blockSize)
67 {
68 uint32_t blkCnt;
69 f16x8_t a, b, accumV, tempV;
70
71 accumV = vdupq_n_f16(0.0f);
72
73 blkCnt = blockSize >> 3;
74 while (blkCnt > 0U) {
75 a = vld1q(pA);
76 b = vld1q(pB);
77
78 tempV = vabdq(a, b);
79 accumV = vaddq(accumV, tempV);
80
81 pA += 8;
82 pB += 8;
83 blkCnt--;
84 }
85
86 /*
87 * tail
88 * (will be merged thru tail predication)
89 */
90 blkCnt = blockSize & 7;
91 if (blkCnt > 0U) {
92 mve_pred16_t p0 = vctp16q(blkCnt);
93
94 a = vldrhq_z_f16(pA, p0);
95 b = vldrhq_z_f16(pB, p0);
96
97 tempV = vabdq(a, b);
98 accumV = vaddq_m(accumV, accumV, tempV, p0);
99 }
100
101 return vecAddAcrossF16Mve(accumV);
102 }
103
104 #else
arm_cityblock_distance_f16(const float16_t * pA,const float16_t * pB,uint32_t blockSize)105 float16_t arm_cityblock_distance_f16(const float16_t *pA,const float16_t *pB, uint32_t blockSize)
106 {
107 _Float16 accum,tmpA, tmpB;
108
109 accum = 0.0f16;
110 while(blockSize > 0)
111 {
112 tmpA = *pA++;
113 tmpB = *pB++;
114 accum += (_Float16)fabsf(tmpA - tmpB);
115
116 blockSize --;
117 }
118
119 return(accum);
120 }
121 #endif
122
123 /**
124 * @} end of Manhattan group
125 */
126
127 #endif /* #if defined(ARM_FLOAT16_SUPPORTED) */
128
129