• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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