1 /* ----------------------------------------------------------------------
2 * Project: CMSIS DSP Library
3 * Title: arm_dot_prod_f16.c
4 * Description: Floating-point dot product
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/basic_math_functions_f16.h"
30
31 /**
32 @ingroup groupMath
33 */
34
35
36 /**
37 @addtogroup BasicDotProd
38 @{
39 */
40
41 /**
42 @brief Dot product of floating-point vectors.
43 @param[in] pSrcA points to the first input vector.
44 @param[in] pSrcB points to the second input vector.
45 @param[in] blockSize number of samples in each vector.
46 @param[out] result output result returned here.
47 */
48
49 #if defined(ARM_MATH_MVE_FLOAT16) && !defined(ARM_MATH_AUTOVECTORIZE)
50
51 #include "arm_helium_utils.h"
52
53
arm_dot_prod_f16(const float16_t * pSrcA,const float16_t * pSrcB,uint32_t blockSize,float16_t * result)54 void arm_dot_prod_f16(
55 const float16_t * pSrcA,
56 const float16_t * pSrcB,
57 uint32_t blockSize,
58 float16_t * result)
59 {
60 f16x8_t vecA, vecB;
61 f16x8_t vecSum;
62 uint32_t blkCnt;
63 float16_t sum = 0.0f;
64 vecSum = vdupq_n_f16(0.0f);
65
66 /* Compute 4 outputs at a time */
67 blkCnt = blockSize >> 3U;
68 while (blkCnt > 0U)
69 {
70 /*
71 * C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1]
72 * Calculate dot product and then store the result in a temporary buffer.
73 * and advance vector source and destination pointers
74 */
75 vecA = vld1q(pSrcA);
76 pSrcA += 8;
77
78 vecB = vld1q(pSrcB);
79 pSrcB += 8;
80
81 vecSum = vfmaq(vecSum, vecA, vecB);
82 /*
83 * Decrement the blockSize loop counter
84 */
85 blkCnt --;
86 }
87
88
89 blkCnt = blockSize & 7;
90 if (blkCnt > 0U)
91 {
92 /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */
93
94 mve_pred16_t p0 = vctp16q(blkCnt);
95 vecA = vld1q(pSrcA);
96 vecB = vld1q(pSrcB);
97 vecSum = vfmaq_m(vecSum, vecA, vecB, p0);
98 }
99
100 sum = vecAddAcrossF16Mve(vecSum);
101
102 /* Store result in destination buffer */
103 *result = sum;
104
105 }
106
107 #else
108 #if defined(ARM_FLOAT16_SUPPORTED)
arm_dot_prod_f16(const float16_t * pSrcA,const float16_t * pSrcB,uint32_t blockSize,float16_t * result)109 void arm_dot_prod_f16(
110 const float16_t * pSrcA,
111 const float16_t * pSrcB,
112 uint32_t blockSize,
113 float16_t * result)
114 {
115 uint32_t blkCnt; /* Loop counter */
116 _Float16 sum = 0.0f; /* Temporary return variable */
117
118
119 #if defined (ARM_MATH_LOOPUNROLL) && !defined(ARM_MATH_AUTOVECTORIZE)
120
121 /* Loop unrolling: Compute 4 outputs at a time */
122 blkCnt = blockSize >> 2U;
123
124 /* First part of the processing with loop unrolling. Compute 4 outputs at a time.
125 ** a second loop below computes the remaining 1 to 3 samples. */
126 while (blkCnt > 0U)
127 {
128 /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */
129
130 /* Calculate dot product and store result in a temporary buffer. */
131 sum += (_Float16)(*pSrcA++) * (_Float16)(*pSrcB++);
132
133 sum += (_Float16)(*pSrcA++) * (_Float16)(*pSrcB++);
134
135 sum += (_Float16)(*pSrcA++) * (_Float16)(*pSrcB++);
136
137 sum += (_Float16)(*pSrcA++) * (_Float16)(*pSrcB++);
138
139 /* Decrement loop counter */
140 blkCnt--;
141 }
142
143 /* Loop unrolling: Compute remaining outputs */
144 blkCnt = blockSize % 0x4U;
145
146 #else
147
148 /* Initialize blkCnt with number of samples */
149 blkCnt = blockSize;
150
151 #endif /* #if defined (ARM_MATH_LOOPUNROLL) */
152
153 while (blkCnt > 0U)
154 {
155 /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */
156
157 /* Calculate dot product and store result in a temporary buffer. */
158 sum += (_Float16)(*pSrcA++) * (_Float16)(*pSrcB++);
159
160 /* Decrement loop counter */
161 blkCnt--;
162 }
163
164 /* Store result in destination buffer */
165 *result = sum;
166 }
167 #endif
168 #endif /* defined(ARM_MATH_MVEF) && !defined(ARM_MATH_AUTOVECTORIZE) */
169 /**
170 @} end of BasicDotProd group
171 */
172
