| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* -*- c++ -*- */ | ||
| 2 | /* | ||
| 3 | * Copyright 2015 Free Software Foundation, Inc. | ||
| 4 | * | ||
| 5 | * This file is part of VOLK | ||
| 6 | * | ||
| 7 | * SPDX-License-Identifier: LGPL-3.0-or-later | ||
| 8 | */ | ||
| 9 | |||
| 10 | /* For documentation see 'kernels/volk/volk_8u_x3_encodepolar_8u_x2.h' | ||
| 11 | * This file exists for test purposes only. Should not be used directly. | ||
| 12 | */ | ||
| 13 | |||
| 14 | #ifndef VOLK_KERNELS_VOLK_VOLK_8U_X3_ENCODEPOLARPUPPET_8U_H_ | ||
| 15 | #define VOLK_KERNELS_VOLK_VOLK_8U_X3_ENCODEPOLARPUPPET_8U_H_ | ||
| 16 | #include <volk/volk.h> | ||
| 17 | #include <volk/volk_8u_x3_encodepolar_8u_x2.h> | ||
| 18 | |||
| 19 | 22 | static inline unsigned int next_lower_power_of_two(const unsigned int val) | |
| 20 | { | ||
| 21 | // algorithm found and adopted from: | ||
| 22 | // http://acius2.blogspot.de/2007/11/calculating-next-power-of-2.html | ||
| 23 | 22 | unsigned int res = val; | |
| 24 | 22 | res = (res >> 1) | res; | |
| 25 | 22 | res = (res >> 2) | res; | |
| 26 | 22 | res = (res >> 4) | res; | |
| 27 | 22 | res = (res >> 8) | res; | |
| 28 | 22 | res = (res >> 16) | res; | |
| 29 | 22 | res += 1; | |
| 30 | 22 | return res >> 1; | |
| 31 | } | ||
| 32 | |||
| 33 | 10 | static inline void adjust_frozen_mask(unsigned char* mask, const unsigned int frame_size) | |
| 34 | { | ||
| 35 | // just like the rest of the puppet this function exists for test purposes only. | ||
| 36 | unsigned int i; | ||
| 37 |
2/2✓ Branch 0 taken 655360 times.
✓ Branch 1 taken 10 times.
|
655370 | for (i = 0; i < frame_size; ++i) { |
| 38 |
2/2✓ Branch 0 taken 327145 times.
✓ Branch 1 taken 328215 times.
|
655360 | *mask = (*mask & 0x80) ? 0xFF : 0x00; |
| 39 | 655360 | mask++; | |
| 40 | } | ||
| 41 | 10 | } | |
| 42 | |||
| 43 | #ifdef LV_HAVE_GENERIC | ||
| 44 | static inline void | ||
| 45 | 2 | volk_8u_x3_encodepolarpuppet_8u_generic(unsigned char* frame, | |
| 46 | unsigned char* frozen_bit_mask, | ||
| 47 | const unsigned char* frozen_bits, | ||
| 48 | const unsigned char* info_bits, | ||
| 49 | unsigned int frame_size) | ||
| 50 | { | ||
| 51 | 2 | frame_size = next_lower_power_of_two(frame_size); | |
| 52 | 2 | unsigned char* temp = (unsigned char*)volk_malloc(sizeof(unsigned char) * frame_size, | |
| 53 | volk_get_alignment()); | ||
| 54 | 2 | adjust_frozen_mask(frozen_bit_mask, frame_size); | |
| 55 | 2 | volk_8u_x3_encodepolar_8u_x2_generic( | |
| 56 | frame, temp, frozen_bit_mask, frozen_bits, info_bits, frame_size); | ||
| 57 | 2 | volk_free(temp); | |
| 58 | 2 | } | |
| 59 | #endif /* LV_HAVE_GENERIC */ | ||
| 60 | |||
| 61 | |||
| 62 | #ifdef LV_HAVE_SSSE3 | ||
| 63 | static inline void | ||
| 64 | 2 | volk_8u_x3_encodepolarpuppet_8u_u_ssse3(unsigned char* frame, | |
| 65 | unsigned char* frozen_bit_mask, | ||
| 66 | const unsigned char* frozen_bits, | ||
| 67 | const unsigned char* info_bits, | ||
| 68 | unsigned int frame_size) | ||
| 69 | { | ||
| 70 | 2 | frame_size = next_lower_power_of_two(frame_size); | |
| 71 | 2 | unsigned char* temp = (unsigned char*)volk_malloc(sizeof(unsigned char) * frame_size, | |
| 72 | volk_get_alignment()); | ||
| 73 | 2 | adjust_frozen_mask(frozen_bit_mask, frame_size); | |
| 74 | 2 | volk_8u_x3_encodepolar_8u_x2_u_ssse3( | |
| 75 | frame, temp, frozen_bit_mask, frozen_bits, info_bits, frame_size); | ||
| 76 | 2 | volk_free(temp); | |
| 77 | 2 | } | |
| 78 | #endif /* LV_HAVE_SSSE3 */ | ||
| 79 | |||
| 80 | #ifdef LV_HAVE_AVX2 | ||
| 81 | static inline void | ||
| 82 | 2 | volk_8u_x3_encodepolarpuppet_8u_u_avx2(unsigned char* frame, | |
| 83 | unsigned char* frozen_bit_mask, | ||
| 84 | const unsigned char* frozen_bits, | ||
| 85 | const unsigned char* info_bits, | ||
| 86 | unsigned int frame_size) | ||
| 87 | { | ||
| 88 | 2 | frame_size = next_lower_power_of_two(frame_size); | |
| 89 | 2 | unsigned char* temp = (unsigned char*)volk_malloc(sizeof(unsigned char) * frame_size, | |
| 90 | volk_get_alignment()); | ||
| 91 | 2 | adjust_frozen_mask(frozen_bit_mask, frame_size); | |
| 92 | 2 | volk_8u_x3_encodepolar_8u_x2_u_avx2( | |
| 93 | frame, temp, frozen_bit_mask, frozen_bits, info_bits, frame_size); | ||
| 94 | 2 | volk_free(temp); | |
| 95 | 2 | } | |
| 96 | #endif /* LV_HAVE_AVX2 */ | ||
| 97 | |||
| 98 | #endif /* VOLK_KERNELS_VOLK_VOLK_8U_X3_ENCODEPOLARPUPPET_8U_H_ */ | ||
| 99 | |||
| 100 | #ifndef VOLK_KERNELS_VOLK_VOLK_8U_X3_ENCODEPOLARPUPPET_8U_A_H_ | ||
| 101 | #define VOLK_KERNELS_VOLK_VOLK_8U_X3_ENCODEPOLARPUPPET_8U_A_H_ | ||
| 102 | |||
| 103 | #ifdef LV_HAVE_SSSE3 | ||
| 104 | static inline void | ||
| 105 | 2 | volk_8u_x3_encodepolarpuppet_8u_a_ssse3(unsigned char* frame, | |
| 106 | unsigned char* frozen_bit_mask, | ||
| 107 | const unsigned char* frozen_bits, | ||
| 108 | const unsigned char* info_bits, | ||
| 109 | unsigned int frame_size) | ||
| 110 | { | ||
| 111 | 2 | frame_size = next_lower_power_of_two(frame_size); | |
| 112 | 2 | unsigned char* temp = (unsigned char*)volk_malloc(sizeof(unsigned char) * frame_size, | |
| 113 | volk_get_alignment()); | ||
| 114 | 2 | adjust_frozen_mask(frozen_bit_mask, frame_size); | |
| 115 | 2 | volk_8u_x3_encodepolar_8u_x2_a_ssse3( | |
| 116 | frame, temp, frozen_bit_mask, frozen_bits, info_bits, frame_size); | ||
| 117 | 2 | volk_free(temp); | |
| 118 | 2 | } | |
| 119 | #endif /* LV_HAVE_SSSE3 */ | ||
| 120 | |||
| 121 | #ifdef LV_HAVE_AVX2 | ||
| 122 | static inline void | ||
| 123 | 2 | volk_8u_x3_encodepolarpuppet_8u_a_avx2(unsigned char* frame, | |
| 124 | unsigned char* frozen_bit_mask, | ||
| 125 | const unsigned char* frozen_bits, | ||
| 126 | const unsigned char* info_bits, | ||
| 127 | unsigned int frame_size) | ||
| 128 | { | ||
| 129 | 2 | frame_size = next_lower_power_of_two(frame_size); | |
| 130 | 2 | unsigned char* temp = (unsigned char*)volk_malloc(sizeof(unsigned char) * frame_size, | |
| 131 | volk_get_alignment()); | ||
| 132 | 2 | adjust_frozen_mask(frozen_bit_mask, frame_size); | |
| 133 | 2 | volk_8u_x3_encodepolar_8u_x2_a_avx2( | |
| 134 | frame, temp, frozen_bit_mask, frozen_bits, info_bits, frame_size); | ||
| 135 | 2 | volk_free(temp); | |
| 136 | 2 | } | |
| 137 | #endif /* LV_HAVE_AVX2 */ | ||
| 138 | |||
| 139 | |||
| 140 | #endif /* VOLK_KERNELS_VOLK_VOLK_8U_X3_ENCODEPOLARPUPPET_8U_A_H_ */ | ||
| 141 |