Statistics
| Branch: | Tag: | Revision:

root / volk / include / volk / volk_16sc_deinterleave_real_8s_aligned16.h @ 15ad4b53

History | View | Annotate | Download (3 kB)

1
#ifndef INCLUDED_VOLK_16sc_DEINTERLEAVE_REAL_8s_ALIGNED16_H
2
#define INCLUDED_VOLK_16sc_DEINTERLEAVE_REAL_8s_ALIGNED16_H
3


4
#include <inttypes.h>



5
#include <stdio.h>



6


7
#if LV_HAVE_SSSE3



8
#include <tmmintrin.h>



9
/*!



10
  \brief Deinterleaves the complex 16 bit vector into 8 bit I vector data



11
  \param complexVector The complex input vector



12
  \param iBuffer The I buffer output data



13
  \param num_points The number of complex data values to be deinterleaved



14
*/



15
static inline void volk_16sc_deinterleave_real_8s_aligned16_ssse3(int8_t* iBuffer, const lv_16sc_t* complexVector, unsigned int num_points){



16
  unsigned int number = 0;



17
  const int8_t* complexVectorPtr = (int8_t*)complexVector;



18
  int8_t* iBufferPtr = iBuffer;



19
  __m128i iMoveMask1 = _mm_set_epi8(0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 13, 12, 9, 8, 5, 4, 1, 0);



20
  __m128i iMoveMask2 = _mm_set_epi8(13, 12, 9, 8, 5, 4, 1, 0, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80);



21
  __m128i complexVal1, complexVal2, complexVal3, complexVal4, iOutputVal;



22


23
  unsigned int sixteenthPoints = num_points / 16;



24


25
  for(number = 0; number < sixteenthPoints; number++){



26
    complexVal1 = _mm_load_si128((__m128i*)complexVectorPtr);  complexVectorPtr += 16;



27
    complexVal2 = _mm_load_si128((__m128i*)complexVectorPtr);  complexVectorPtr += 16;



28


29
    complexVal3 = _mm_load_si128((__m128i*)complexVectorPtr);  complexVectorPtr += 16;



30
    complexVal4 = _mm_load_si128((__m128i*)complexVectorPtr);  complexVectorPtr += 16;



31


32
    complexVal1 = _mm_shuffle_epi8(complexVal1, iMoveMask1);



33
    complexVal2 = _mm_shuffle_epi8(complexVal2, iMoveMask2);



34


35
    complexVal1 = _mm_or_si128(complexVal1, complexVal2);



36


37
    complexVal3 = _mm_shuffle_epi8(complexVal3, iMoveMask1);



38
    complexVal4 = _mm_shuffle_epi8(complexVal4, iMoveMask2);



39


40
    complexVal3 = _mm_or_si128(complexVal3, complexVal4);



41


42


43
    complexVal1 = _mm_srai_epi16(complexVal1, 8);



44
    complexVal3 = _mm_srai_epi16(complexVal3, 8);



45


46
    iOutputVal = _mm_packs_epi16(complexVal1, complexVal3);



47


48
    _mm_store_si128((__m128i*)iBufferPtr, iOutputVal);



49


50
    iBufferPtr += 16;



51
  }



52


53
  number = sixteenthPoints * 16;



54
  int16_t* int16ComplexVectorPtr = (int16_t*)complexVectorPtr;



55
  for(; number < num_points; number++){



56
    *iBufferPtr++ = ((int8_t)(*int16ComplexVectorPtr++ / 256));



57
    int16ComplexVectorPtr++;



58
  }



59
}



60
#endif /* LV_HAVE_SSSE3 */



61


62
#if LV_HAVE_GENERIC



63
/*!



64
  \brief Deinterleaves the complex 16 bit vector into 8 bit I vector data



65
  \param complexVector The complex input vector



66
  \param iBuffer The I buffer output data



67
  \param num_points The number of complex data values to be deinterleaved



68
*/



69
static inline void volk_16sc_deinterleave_real_8s_aligned16_generic(int8_t* iBuffer, const lv_16sc_t* complexVector, unsigned int num_points){



70
  unsigned int number = 0;



71
  const int16_t* complexVectorPtr = (int16_t*)complexVector;



72
  int8_t* iBufferPtr = iBuffer;



73
  for(number = 0; number < num_points; number++){



74
    *iBufferPtr++ = (int8_t)(*complexVectorPtr++ / 256);



75
    complexVectorPtr++;



76
  }



77
}



78
#endif /* LV_HAVE_GENERIC */



79


80


81


82


83
#endif /* INCLUDED_VOLK_16sc_DEINTERLEAVE_REAL_8s_ALIGNED16_H */