1 /* ----------------------------------------------------------------------
2  * Project:      CMSIS DSP Library
3  * Title:        arm_mfcc_q15.c
4  * Description:  MFCC function for the q15 version
5  *
6  * $Date:        07 September 2021
7  * $Revision:    V1.10.0
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 
30 
31 #include "dsp/transform_functions.h"
32 #include "dsp/statistics_functions.h"
33 #include "dsp/basic_math_functions.h"
34 #include "dsp/complex_math_functions.h"
35 #include "dsp/fast_math_functions.h"
36 #include "dsp/matrix_functions.h"
37 
38 /* Constants for Q15 implementation */
39 #define LOG2TOLOG_Q15 0x02C5C860
40 #define MICRO_Q15 0x00000219
41 #define SHIFT_MELFILTER_SATURATION_Q15 10
42 /**
43   @ingroup MFCC
44  */
45 
46 
47 
48 /**
49   @addtogroup MFCCQ15
50   @{
51  */
52 
53 /**
54   @brief         MFCC Q15
55   @param[in]    S       points to the mfcc instance structure
56   @param[in]     pSrc points to the input samples in Q15
57   @param[out]     pDst  points to the output MFCC values in q8.7 format
58   @param[inout]     pTmp  points to a temporary buffer of complex
59 
60   @return        none
61 
62   @par           Description
63                    The number of input samples is the FFT length used
64                    when initializing the instance data structure.
65 
66                    The temporary buffer has a 2*fft length.
67 
68                    The source buffer is modified by this function.
69 
70                    The function may saturate. If the FFT length is too
71                    big and the number of MEL filters too small then the fixed
72                    point computations may saturate.
73 
74  */
arm_mfcc_q15(const arm_mfcc_instance_q15 * S,q15_t * pSrc,q15_t * pDst,q31_t * pTmp)75 arm_status arm_mfcc_q15(
76   const arm_mfcc_instance_q15 * S,
77   q15_t *pSrc,
78   q15_t *pDst,
79   q31_t *pTmp
80   )
81 {
82     q15_t m;
83     uint32_t index;
84     uint32_t fftShift=0;
85     q31_t logExponent;
86     q63_t result;
87     arm_matrix_instance_q15 pDctMat;
88     uint32_t i;
89     uint32_t coefsPos;
90     uint32_t filterLimit;
91     q15_t *pTmp2=(q15_t*)pTmp;
92 
93     arm_status status = ARM_MATH_SUCCESS;
94 
95     // q15
96     arm_absmax_q15(pSrc,S->fftLen,&m,&index);
97 
98     if ((m != 0) && (m != 0x7FFF))
99     {
100        q15_t quotient;
101        int16_t shift;
102 
103        status = arm_divide_q15(0x7FFF,m,&quotient,&shift);
104        if (status != ARM_MATH_SUCCESS)
105        {
106           return(status);
107        }
108 
109        arm_scale_q15(pSrc,quotient,shift,pSrc,S->fftLen);
110     }
111 
112 
113     // q15
114     arm_mult_q15(pSrc,S->windowCoefs, pSrc, S->fftLen);
115 
116 
117     /* Compute spectrum magnitude
118     */
119     fftShift = 31 - __CLZ(S->fftLen);
120 #if defined(ARM_MFCC_CFFT_BASED)
121     /* some HW accelerator for CMSIS-DSP used in some boards
122        are only providing acceleration for CFFT.
123        With ARM_MFCC_CFFT_BASED enabled, CFFT is used and the MFCC
124        will be accelerated on those boards.
125 
126        The default is to use RFFT
127     */
128     /* Convert from real to complex */
129     for(i=0; i < S->fftLen ; i++)
130     {
131       pTmp2[2*i] = pSrc[i];
132       pTmp2[2*i+1] = 0;
133     }
134     arm_cfft_q15(&(S->cfft),pTmp2,0,1);
135 #else
136     /* Default RFFT based implementation */
137     arm_rfft_q15(&(S->rfft),pSrc,pTmp2);
138 #endif
139     filterLimit = 1 + (S->fftLen >> 1);
140 
141 
142     // q15 - fftShift
143     arm_cmplx_mag_q15(pTmp2,pSrc,filterLimit);
144     // q14 - fftShift
145 
146     /* Apply MEL filters */
147     coefsPos = 0;
148     for(i=0; i<S->nbMelFilters; i++)
149     {
150       arm_dot_prod_q15(pSrc+S->filterPos[i],
151         &(S->filterCoefs[coefsPos]),
152         S->filterLengths[i],
153         &result);
154 
155       coefsPos += S->filterLengths[i];
156 
157       // q34.29 - fftShift
158       result += MICRO_Q15;
159       result >>= SHIFT_MELFILTER_SATURATION_Q15;
160       // q34.29 - fftShift - satShift
161       pTmp[i] = __SSAT(result,31) ;
162 
163     }
164 
165     if ((m != 0) && (m != 0x7FFF))
166     {
167       arm_scale_q31(pTmp,m<<16,0,pTmp,S->nbMelFilters);
168     }
169 
170     // q34.29 - fftShift - satShift
171     /* Compute the log */
172     arm_vlog_q31(pTmp,pTmp,S->nbMelFilters);
173 
174 
175     // q5.26
176 
177     logExponent = fftShift + 2 + SHIFT_MELFILTER_SATURATION_Q15;
178     logExponent = logExponent * LOG2TOLOG_Q15;
179 
180 
181     // q8.26
182     arm_offset_q31(pTmp,logExponent,pTmp,S->nbMelFilters);
183     arm_shift_q31(pTmp,-19,pTmp,S->nbMelFilters);
184     for(i=0; i<S->nbMelFilters; i++)
185     {
186       pSrc[i] = __SSAT((q15_t)pTmp[i],16);
187     }
188 
189     // q8.7
190 
191     pDctMat.numRows=S->nbDctOutputs;
192     pDctMat.numCols=S->nbMelFilters;
193     pDctMat.pData=(q15_t*)S->dctCoefs;
194 
195     arm_mat_vec_mult_q15(&pDctMat, pSrc, pDst);
196 
197     return(status);
198 }
199 
200 /**
201   @} end of MFCCQ15 group
202  */
203