1 /* ----------------------------------------------------------------------
2 * Project: CMSIS DSP Library
3 * Title: arm_vlog_f16.c
4 * Description: Fast vectorized log
5 *
6 * $Date: 23 April 2021
7 * $Revision: V1.9.0
8 *
9 * Target Processor: Cortex-M and Cortex-A cores
10 * -------------------------------------------------------------------- */
11 /*
12 * Copyright (C) 2010-2021 ARM Limited or its affiliates. All rights reserved.
13 *
14 * SPDX-License-Identifier: Apache-2.0
15 *
16 * Licensed under the Apache License, Version 2.0 (the License); you may
17 * not use this file except in compliance with the License.
18 * You may obtain a copy of the License at
19 *
20 * www.apache.org/licenses/LICENSE-2.0
21 *
22 * Unless required by applicable law or agreed to in writing, software
23 * distributed under the License is distributed on an AS IS BASIS, WITHOUT
24 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
25 * See the License for the specific language governing permissions and
26 * limitations under the License.
27 */
28
29 #include "dsp/fast_math_functions_f16.h"
30 #include "dsp/support_functions_f16.h"
31
32 #if defined(ARM_FLOAT16_SUPPORTED)
33
34 /* Degree of the polynomial approximation */
35 #define NB_DEG_LOGF16 3
36
37 /*
38 Related to the Log2 of the number of approximations.
39 For instance, with 3 there are 1 + 2^3 polynomials
40 */
41 #define NB_DIV_LOGF16 3
42
43 /* Length of the LUT table */
44 #define NB_LUT_LOGF16 (NB_DEG_LOGF16+1)*(1 + (1<<NB_DIV_LOGF16))
45
46
47 /*
48
49 LUT of polynomial approximations.
50
51 Could be generated with:
52
53 ClearAll[lut, coefs, nb, deg];
54 nb = 3;
55 deg = 3;
56 lut = Table[
57 MiniMaxApproximation[
58 Log[x/2^nb + i], {x, {10^-6, 1.0/2^nb}, deg, 0},
59 MaxIterations -> 1000][[2, 1]], {i, 1, 2, (1.0/2^nb)}];
60 coefs = Chop@Flatten[CoefficientList[lut, x]];
61
62 */
63 static float16_t lut_logf16[NB_LUT_LOGF16]={
64 0,0.125,-0.00781197,0.00063974,0.117783,
65 0.111111,-0.00617212,0.000447935,0.223144,
66 0.1,-0.00499952,0.000327193,0.318454,0.0909091,
67 -0.00413191,0.000246234,0.405465,0.0833333,
68 -0.00347199,0.000189928,0.485508,0.0769231,
69 -0.00295841,0.00014956,0.559616,0.0714286,
70 -0.0025509,0.000119868,0.628609,0.0666667,
71 -0.00222213,0.0000975436,0.693147,
72 0.0625,-0.00195305,0.0000804357};
73
74
logf16_scalar(float16_t x)75 float16_t logf16_scalar(float16_t x)
76 {
77 int16_t i = arm_typecast_s16_f16(x);
78
79 int32_t vecExpUnBiased = (i >> 10) - 15;
80 i = i - (vecExpUnBiased << 10);
81 float16_t vecTmpFlt1 = arm_typecast_f16_s16(i);
82
83 float16_t *lut;
84 int n;
85 float16_t tmp,v;
86
87 tmp = ((_Float16)vecTmpFlt1 - 1.0f16) * (1 << NB_DIV_LOGF16);
88 n = (int)floor((double)tmp);
89 v = (_Float16)tmp - (_Float16)n;
90
91 lut = lut_logf16 + n * (1+NB_DEG_LOGF16);
92
93 float16_t res = lut[NB_DEG_LOGF16-1];
94 for(int j=NB_DEG_LOGF16-2; j >=0 ; j--)
95 {
96 res = (_Float16)lut[j] + (_Float16)v * (_Float16)res;
97 }
98
99 res = (_Float16)res + 0.693147f16 * (_Float16)vecExpUnBiased;
100
101
102 return(res);
103 }
104
105 #if defined(ARM_MATH_MVE_FLOAT16) && !defined(ARM_MATH_AUTOVECTORIZE)
106
107 #include "arm_common_tables.h"
108 #include "arm_vec_math_f16.h"
109
110
vlogq_lut_f16(float16x8_t vecIn)111 float16x8_t vlogq_lut_f16(float16x8_t vecIn)
112 {
113 int16x8_t i = vreinterpretq_s16_f16(vecIn);
114
115 int16x8_t vecExpUnBiased = vsubq_n_s16(vshrq_n_s16(i,10), 15);
116 i = vsubq_s16(i,vshlq_n_s16(vecExpUnBiased,10));
117 float16x8_t vecTmpFlt1 = vreinterpretq_f16_s16(i);
118
119
120 float16x8_t lutV;
121 int16x8_t n;
122 int16x8_t offset;
123
124 float16x8_t tmp,v,res;
125
126 tmp = vmulq_n_f16(vsubq_n_f16(vecTmpFlt1,1.0f16),(_Float16)(1 << NB_DIV_LOGF16));
127
128 n = vcvtq_s16_f16(tmp);
129 v = vsubq_f16(tmp,vcvtq_f16_s16(n));
130
131
132 offset = vmulq_n_s16(n,(1+NB_DEG_LOGF16));
133 offset = vaddq_n_s16(offset,NB_DEG_LOGF16-1);
134
135 res = vldrhq_gather_shifted_offset_f16(lut_logf16,(uint16x8_t)offset);
136 offset = vsubq_n_s16(offset,1);
137
138 for(int j=NB_DEG_LOGF16-2; j >=0 ; j--)
139 {
140 lutV = vldrhq_gather_shifted_offset_f16(lut_logf16,(uint16x8_t)offset);
141 res = vfmaq_f16(lutV,v,res);
142 offset = vsubq_n_s16(offset,1);
143
144 }
145
146 res = vfmaq_n_f16(res,vcvtq_f16_s16(vecExpUnBiased),0.693147f16);
147
148
149 return(res);
150
151 }
152
153 #endif
154
155 /**
156 @ingroup groupFastMath
157 */
158
159 /**
160 @addtogroup vlog
161 @{
162 */
163
164 /**
165 @brief Floating-point vector of log values.
166 @param[in] pSrc points to the input vector
167 @param[out] pDst points to the output vector
168 @param[in] blockSize number of samples in each vector
169 @return none
170 */
171
172
arm_vlog_f16(const float16_t * pSrc,float16_t * pDst,uint32_t blockSize)173 void arm_vlog_f16(
174 const float16_t * pSrc,
175 float16_t * pDst,
176 uint32_t blockSize)
177 {
178 uint32_t blkCnt;
179
180 #if defined(ARM_MATH_MVE_FLOAT16) && !defined(ARM_MATH_AUTOVECTORIZE)
181 f16x8_t src;
182 f16x8_t dst;
183
184 blkCnt = blockSize >> 3;
185
186 while (blkCnt > 0U)
187 {
188 src = vld1q(pSrc);
189 dst = vlogq_lut_f16(src);
190 vst1q(pDst, dst);
191
192 pSrc += 8;
193 pDst += 8;
194 /* Decrement loop counter */
195 blkCnt--;
196 }
197
198 blkCnt = blockSize & 7;
199 #else
200 blkCnt = blockSize;
201 #endif
202
203 while (blkCnt > 0U)
204 {
205 /* C = log(A) */
206
207 /* Calculate log and store result in destination buffer. */
208 *pDst++ = logf16_scalar(*pSrc++);
209
210 /* Decrement loop counter */
211 blkCnt--;
212 }
213 }
214
215
216
217 /**
218 @} end of vlog group
219 */
220
221
222 #endif /* #if defined(ARM_FLOAT16_SUPPORTED) */
223