GNU Radio 3.7.2 C++ API
volk_32f_x2_divide_32f.h
Go to the documentation of this file.
1 #ifndef INCLUDED_volk_32f_x2_divide_32f_a_H
2 #define INCLUDED_volk_32f_x2_divide_32f_a_H
3 
4 #include <inttypes.h>
5 #include <stdio.h>
6 
7 #ifdef LV_HAVE_SSE
8 #include <xmmintrin.h>
9 /*!
10  \brief Divides 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 The vector to be divideed
13  \param bVector The divisor vector
14  \param num_points The number of values in aVector and bVector to be divideed together and stored into cVector
15 */
16 static inline void volk_32f_x2_divide_32f_a_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_load_ps(aPtr);
28  bVal = _mm_load_ps(bPtr);
29 
30  cVal = _mm_div_ps(aVal, bVal);
31 
32  _mm_store_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 Divides 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 The vector to be divideed
51  \param bVector The divisor vector
52  \param num_points The number of values in aVector and bVector to be divideed together and stored into cVector
53 */
54 static inline void volk_32f_x2_divide_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 #ifdef LV_HAVE_ORC
67 /*!
68  \brief Divides the two input vectors and store their results in the third vector
69  \param cVector The vector where the results will be stored
70  \param aVector The vector to be divideed
71  \param bVector The divisor vector
72  \param num_points The number of values in aVector and bVector to be divideed together and stored into cVector
73 */
74 extern void volk_32f_x2_divide_32f_a_orc_impl(float* cVector, const float* aVector, const float* bVector, unsigned int num_points);
75 static inline void volk_32f_x2_divide_32f_u_orc(float* cVector, const float* aVector, const float* bVector, unsigned int num_points){
76  volk_32f_x2_divide_32f_a_orc_impl(cVector, aVector, bVector, num_points);
77 }
78 #endif /* LV_HAVE_ORC */
79 
80 
81 
82 #endif /* INCLUDED_volk_32f_x2_divide_32f_a_H */