GNU Radio Manual and C++ API Reference  3.7.4
The Free & Open Software Radio Ecosystem
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Properties Friends Macros Groups Pages
volk_32f_x2_add_32f.h
Go to the documentation of this file.
1 #ifndef INCLUDED_volk_32f_x2_add_32f_u_H
2 #define INCLUDED_volk_32f_x2_add_32f_u_H
3 
4 #include <inttypes.h>
5 #include <stdio.h>
6 
7 #ifdef LV_HAVE_SSE
8 #include <xmmintrin.h>
9 /*!
10  \brief Adds the two input vectors and store their results in the third vector
11  \param cVector The vector where the results will be stored
12  \param aVector One of the vectors to be added
13  \param bVector One of the vectors to be added
14  \param num_points The number of values in aVector and bVector to be added together and stored into cVector
15 */
16 static inline void volk_32f_x2_add_32f_u_sse(float* cVector, const float* aVector, const float* bVector, unsigned int num_points){
17  unsigned int number = 0;
18  const unsigned int quarterPoints = num_points / 4;
19 
20  float* cPtr = cVector;
21  const float* aPtr = aVector;
22  const float* bPtr= bVector;
23 
24  __m128 aVal, bVal, cVal;
25  for(;number < quarterPoints; number++){
26 
27  aVal = _mm_loadu_ps(aPtr);
28  bVal = _mm_loadu_ps(bPtr);
29 
30  cVal = _mm_add_ps(aVal, bVal);
31 
32  _mm_storeu_ps(cPtr,cVal); // Store the results back into the C container
33 
34  aPtr += 4;
35  bPtr += 4;
36  cPtr += 4;
37  }
38 
39  number = quarterPoints * 4;
40  for(;number < num_points; number++){
41  *cPtr++ = (*aPtr++) + (*bPtr++);
42  }
43 }
44 #endif /* LV_HAVE_SSE */
45 
46 #ifdef LV_HAVE_GENERIC
47 /*!
48  \brief Adds the two input vectors and store their results in the third vector
49  \param cVector The vector where the results will be stored
50  \param aVector One of the vectors to be added
51  \param bVector One of the vectors to be added
52  \param num_points The number of values in aVector and bVector to be added together and stored into cVector
53 */
54 static inline void volk_32f_x2_add_32f_generic(float* cVector, const float* aVector, const float* bVector, unsigned int num_points){
55  float* cPtr = cVector;
56  const float* aPtr = aVector;
57  const float* bPtr= bVector;
58  unsigned int number = 0;
59 
60  for(number = 0; number < num_points; number++){
61  *cPtr++ = (*aPtr++) + (*bPtr++);
62  }
63 }
64 #endif /* LV_HAVE_GENERIC */
65 
66 #endif /* INCLUDED_volk_32f_x2_add_32f_u_H */
67 #ifndef INCLUDED_volk_32f_x2_add_32f_a_H
68 #define INCLUDED_volk_32f_x2_add_32f_a_H
69 
70 #include <inttypes.h>
71 #include <stdio.h>
72 
73 #ifdef LV_HAVE_SSE
74 #include <xmmintrin.h>
75 /*!
76  \brief Adds the two input vectors and store their results in the third vector
77  \param cVector The vector where the results will be stored
78  \param aVector One of the vectors to be added
79  \param bVector One of the vectors to be added
80  \param num_points The number of values in aVector and bVector to be added together and stored into cVector
81 */
82 static inline void volk_32f_x2_add_32f_a_sse(float* cVector, const float* aVector, const float* bVector, unsigned int num_points){
83  unsigned int number = 0;
84  const unsigned int quarterPoints = num_points / 4;
85 
86  float* cPtr = cVector;
87  const float* aPtr = aVector;
88  const float* bPtr= bVector;
89 
90  __m128 aVal, bVal, cVal;
91  for(;number < quarterPoints; number++){
92 
93  aVal = _mm_load_ps(aPtr);
94  bVal = _mm_load_ps(bPtr);
95 
96  cVal = _mm_add_ps(aVal, bVal);
97 
98  _mm_store_ps(cPtr,cVal); // Store the results back into the C container
99 
100  aPtr += 4;
101  bPtr += 4;
102  cPtr += 4;
103  }
104 
105  number = quarterPoints * 4;
106  for(;number < num_points; number++){
107  *cPtr++ = (*aPtr++) + (*bPtr++);
108  }
109 }
110 #endif /* LV_HAVE_SSE */
111 
112 #ifdef LV_HAVE_GENERIC
113 /*!
114  \brief Adds the two input vectors and store their results in the third vector
115  \param cVector The vector where the results will be stored
116  \param aVector One of the vectors to be added
117  \param bVector One of the vectors to be added
118  \param num_points The number of values in aVector and bVector to be added together and stored into cVector
119 */
120 static inline void volk_32f_x2_add_32f_a_generic(float* cVector, const float* aVector, const float* bVector, unsigned int num_points){
121  float* cPtr = cVector;
122  const float* aPtr = aVector;
123  const float* bPtr= bVector;
124  unsigned int number = 0;
125 
126  for(number = 0; number < num_points; number++){
127  *cPtr++ = (*aPtr++) + (*bPtr++);
128  }
129 }
130 #endif /* LV_HAVE_GENERIC */
131 
132 #ifdef LV_HAVE_ORC
133 /*!
134  \brief Adds the two input vectors and store their results in the third vector
135  \param cVector The vector where the results will be stored
136  \param aVector One of the vectors to be added
137  \param bVector One of the vectors to be added
138  \param num_points The number of values in aVector and bVector to be added together and stored into cVector
139 */
140 extern void volk_32f_x2_add_32f_a_orc_impl(float* cVector, const float* aVector, const float* bVector, unsigned int num_points);
141 static inline void volk_32f_x2_add_32f_u_orc(float* cVector, const float* aVector, const float* bVector, unsigned int num_points){
142  volk_32f_x2_add_32f_a_orc_impl(cVector, aVector, bVector, num_points);
143 }
144 #endif /* LV_HAVE_ORC */
145 
146 
147 #endif /* INCLUDED_volk_32f_x2_add_32f_a_H */