xref: /btstack/port/stm32-f4discovery-usb/Drivers/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q15.c (revision a8f7f3fcbcd51f8d2e92aca076b6a9f812db358c)
1 /*
2  * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
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_nn_mult_q15.c
22  * Description:  Q15 vector multiplication with variable output shifts
23  *
24  * $Date:        13. July 2018
25  * $Revision:    V.1.0.0
26  *
27  * Target Processor:  Cortex-M cores
28  *
29  * -------------------------------------------------------------------- */
30 
31 #include "arm_nnfunctions.h"
32 
33 /**
34  * @ingroup groupSupport
35  */
36 
37 /**
38  * @addtogroup NNBasicMath
39  * @{
40  */
41 
42 
43 /**
44  * @brief           Q7 vector multiplication with variable output shifts
45  * @param[in]       *pSrcA        pointer to the first input vector
46  * @param[in]       *pSrcB        pointer to the second input vector
47  * @param[out]      *pDst         pointer to the output vector
48  * @param[in]       out_shift     amount of right-shift for output
49  * @param[in]       blockSize     number of samples in each vector
50  * @return none.
51  *
52  * <b>Scaling and Overflow Behavior:</b>
53  * \par
54  * The function uses saturating arithmetic.
55  * Results outside of the allowable Q15 range [0x8000 0x7FFF] will be saturated.
56  */
57 
arm_nn_mult_q15(q15_t * pSrcA,q15_t * pSrcB,q15_t * pDst,const uint16_t out_shift,uint32_t blockSize)58 void arm_nn_mult_q15(
59   q15_t * pSrcA,
60   q15_t * pSrcB,
61   q15_t * pDst,
62   const uint16_t out_shift,
63   uint32_t blockSize)
64 {
65   uint32_t blkCnt;                               /* loop counters */
66 
67 #if defined (ARM_MATH_DSP)
68 
69 /* Run the below code for Cortex-M4 and Cortex-M3 */
70   q31_t inA1, inA2, inB1, inB2;                  /* temporary input variables */
71   q15_t out1, out2, out3, out4;                  /* temporary output variables */
72   q31_t mul1, mul2, mul3, mul4;                  /* temporary variables */
73 
74   /* loop Unrolling */
75   blkCnt = blockSize >> 2U;
76 
77   /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.
78    ** a second loop below computes the remaining 1 to 3 samples. */
79   while (blkCnt > 0U)
80   {
81     /* read two samples at a time from sourceA */
82     inA1 = *__SIMD32(pSrcA)++;
83     /* read two samples at a time from sourceB */
84     inB1 = *__SIMD32(pSrcB)++;
85     /* read two samples at a time from sourceA */
86     inA2 = *__SIMD32(pSrcA)++;
87     /* read two samples at a time from sourceB */
88     inB2 = *__SIMD32(pSrcB)++;
89 
90     /* multiply mul = sourceA * sourceB */
91     mul1 = (q31_t) ((q15_t) (inA1 >> 16) * (q15_t) (inB1 >> 16));
92     mul2 = (q31_t) ((q15_t) inA1 * (q15_t) inB1);
93     mul3 = (q31_t) ((q15_t) (inA2 >> 16) * (q15_t) (inB2 >> 16));
94     mul4 = (q31_t) ((q15_t) inA2 * (q15_t) inB2);
95 
96     /* saturate result to 16 bit */
97     out1 = (q15_t) __SSAT((mul1 + NN_ROUND(out_shift)) >> out_shift, 16);
98     out2 = (q15_t) __SSAT((mul2 + NN_ROUND(out_shift)) >> out_shift, 16);
99     out3 = (q15_t) __SSAT((mul3 + NN_ROUND(out_shift)) >> out_shift, 16);
100     out4 = (q15_t) __SSAT((mul4 + NN_ROUND(out_shift)) >> out_shift, 16);
101 
102     /* store the result */
103 #ifndef ARM_MATH_BIG_ENDIAN
104 
105     *__SIMD32(pDst)++ = __PKHBT(out2, out1, 16);
106     *__SIMD32(pDst)++ = __PKHBT(out4, out3, 16);
107 
108 #else
109 
110     *__SIMD32(pDst)++ = __PKHBT(out2, out1, 16);
111     *__SIMD32(pDst)++ = __PKHBT(out4, out3, 16);
112 
113 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
114 
115     /* Decrement the blockSize loop counter */
116     blkCnt--;
117   }
118 
119   /* If the blockSize is not a multiple of 4, compute any remaining output samples here.
120    ** No loop unrolling is used. */
121   blkCnt = blockSize % 0x4U;
122 
123 #else
124 
125   /* Run the below code for Cortex-M0 */
126 
127   /* Initialize blkCnt with number of samples */
128   blkCnt = blockSize;
129 
130 #endif /* #if defined (ARM_MATH_DSP) */
131 
132 
133   while (blkCnt > 0U)
134   {
135     /* C = A * B */
136     /* Multiply the inputs and store the result in the destination buffer */
137     *pDst++ = (q15_t) __SSAT((((q31_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 16);
138 
139     /* Decrement the blockSize loop counter */
140     blkCnt--;
141   }
142 }
143 
144 /**
145  * @} end of NNBasicMath group
146  */
147 
148