root / volk / include / volk / volk_16ic_deinterleave_real_16i_a16.h @ e3600f59
History | View | Annotate | Download (4.1 kB)
| 1 | #ifndef INCLUDED_volk_16ic_deinterleave_real_16i_a16_H
|
|---|---|
| 2 | #define INCLUDED_volk_16ic_deinterleave_real_16i_a16_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 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_16ic_deinterleave_real_16i_a16_ssse3(int16_t* iBuffer, const lv_16sc_t* complexVector, unsigned int num_points){ |
| 16 | unsigned int number = 0; |
| 17 | const int16_t* complexVectorPtr = (int16_t*)complexVector;
|
| 18 | int16_t* iBufferPtr = iBuffer; |
| 19 | |
| 20 | __m128i iMoveMask1 = _mm_set_epi8(0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 13, 12, 9, 8, 5, 4, 1, 0); |
| 21 | __m128i iMoveMask2 = _mm_set_epi8(13, 12, 9, 8, 5, 4, 1, 0, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80); |
| 22 | |
| 23 | __m128i complexVal1, complexVal2, iOutputVal; |
| 24 | |
| 25 | unsigned int eighthPoints = num_points / 8; |
| 26 | |
| 27 | for(number = 0; number < eighthPoints; number++){ |
| 28 | complexVal1 = _mm_load_si128((__m128i*)complexVectorPtr); complexVectorPtr += 8;
|
| 29 | complexVal2 = _mm_load_si128((__m128i*)complexVectorPtr); complexVectorPtr += 8;
|
| 30 | |
| 31 | complexVal1 = _mm_shuffle_epi8(complexVal1, iMoveMask1); |
| 32 | complexVal2 = _mm_shuffle_epi8(complexVal2, iMoveMask2); |
| 33 | |
| 34 | iOutputVal = _mm_or_si128(complexVal1, complexVal2); |
| 35 | |
| 36 | _mm_store_si128((__m128i*)iBufferPtr, iOutputVal); |
| 37 | |
| 38 | iBufferPtr += 8;
|
| 39 | } |
| 40 | |
| 41 | number = eighthPoints * 8;
|
| 42 | for(; number < num_points; number++){
|
| 43 | *iBufferPtr++ = *complexVectorPtr++; |
| 44 | complexVectorPtr++; |
| 45 | } |
| 46 | } |
| 47 | #endif /* LV_HAVE_SSSE3 */ |
| 48 | |
| 49 | |
| 50 | #if LV_HAVE_SSE2
|
| 51 | #include <emmintrin.h> |
| 52 | /*!
|
| 53 | \brief Deinterleaves the complex 16 bit vector into I vector data |
| 54 | \param complexVector The complex input vector |
| 55 | \param iBuffer The I buffer output data |
| 56 | \param num_points The number of complex data values to be deinterleaved |
| 57 | */ |
| 58 | static inline void volk_16ic_deinterleave_real_16i_a16_sse2(int16_t* iBuffer, const lv_16sc_t* complexVector, unsigned int num_points){ |
| 59 | unsigned int number = 0; |
| 60 | const int16_t* complexVectorPtr = (int16_t*)complexVector;
|
| 61 | int16_t* iBufferPtr = iBuffer; |
| 62 | __m128i complexVal1, complexVal2, iOutputVal; |
| 63 | __m128i lowMask = _mm_set_epi32(0x0, 0x0, 0xFFFFFFFF, 0xFFFFFFFF); |
| 64 | __m128i highMask = _mm_set_epi32(0xFFFFFFFF, 0xFFFFFFFF, 0x0, 0x0); |
| 65 | |
| 66 | unsigned int eighthPoints = num_points / 8; |
| 67 | |
| 68 | for(number = 0; number < eighthPoints; number++){ |
| 69 | complexVal1 = _mm_load_si128((__m128i*)complexVectorPtr); complexVectorPtr += 8;
|
| 70 | complexVal2 = _mm_load_si128((__m128i*)complexVectorPtr); complexVectorPtr += 8;
|
| 71 | |
| 72 | complexVal1 = _mm_shufflelo_epi16(complexVal1, _MM_SHUFFLE(3,1,2,0)); |
| 73 | |
| 74 | complexVal1 = _mm_shufflehi_epi16(complexVal1, _MM_SHUFFLE(3,1,2,0)); |
| 75 | |
| 76 | complexVal1 = _mm_shuffle_epi32(complexVal1, _MM_SHUFFLE(3,1,2,0)); |
| 77 | |
| 78 | complexVal2 = _mm_shufflelo_epi16(complexVal2, _MM_SHUFFLE(3,1,2,0)); |
| 79 | |
| 80 | complexVal2 = _mm_shufflehi_epi16(complexVal2, _MM_SHUFFLE(3,1,2,0)); |
| 81 | |
| 82 | complexVal2 = _mm_shuffle_epi32(complexVal2, _MM_SHUFFLE(2,0,3,1)); |
| 83 | |
| 84 | iOutputVal = _mm_or_si128(_mm_and_si128(complexVal1, lowMask), _mm_and_si128(complexVal2, highMask)); |
| 85 | |
| 86 | _mm_store_si128((__m128i*)iBufferPtr, iOutputVal); |
| 87 | |
| 88 | iBufferPtr += 8;
|
| 89 | } |
| 90 | |
| 91 | number = eighthPoints * 8;
|
| 92 | for(; number < num_points; number++){
|
| 93 | *iBufferPtr++ = *complexVectorPtr++; |
| 94 | complexVectorPtr++; |
| 95 | } |
| 96 | } |
| 97 | #endif /* LV_HAVE_SSE2 */ |
| 98 | |
| 99 | #if LV_HAVE_GENERIC
|
| 100 | /*!
|
| 101 | \brief Deinterleaves the complex 16 bit vector into I vector data |
| 102 | \param complexVector The complex input vector |
| 103 | \param iBuffer The I buffer output data |
| 104 | \param num_points The number of complex data values to be deinterleaved |
| 105 | */ |
| 106 | static inline void volk_16ic_deinterleave_real_16i_a16_generic(int16_t* iBuffer, const lv_16sc_t* complexVector, unsigned int num_points){ |
| 107 | unsigned int number = 0; |
| 108 | const int16_t* complexVectorPtr = (int16_t*)complexVector;
|
| 109 | int16_t* iBufferPtr = iBuffer; |
| 110 | for(number = 0; number < num_points; number++){ |
| 111 | *iBufferPtr++ = *complexVectorPtr++; |
| 112 | complexVectorPtr++; |
| 113 | } |
| 114 | } |
| 115 | #endif /* LV_HAVE_GENERIC */ |
| 116 | |
| 117 | |
| 118 | |
| 119 | |
| 120 | #endif /* INCLUDED_volk_16ic_deinterleave_real_16i_a16_H */ |