Vector Optimized Library of Kernels 2.5.2
Architecture-tuned implementations of math kernels
volk_avx_intrinsics.h
Go to the documentation of this file.
1/* -*- c++ -*- */
2/*
3 * Copyright 2015 Free Software Foundation, Inc.
4 *
5 * This file is part of VOLK
6 *
7 * SPDX-License-Identifier: GPL-3.0-or-later
8 */
9
10/*
11 * This file is intended to hold AVX intrinsics of intrinsics.
12 * They should be used in VOLK kernels to avoid copy-pasta.
13 */
14
15#ifndef INCLUDE_VOLK_VOLK_AVX_INTRINSICS_H_
16#define INCLUDE_VOLK_VOLK_AVX_INTRINSICS_H_
17#include <immintrin.h>
18
19static inline __m256 _mm256_complexmul_ps(__m256 x, __m256 y)
20{
21 __m256 yl, yh, tmp1, tmp2;
22 yl = _mm256_moveldup_ps(y); // Load yl with cr,cr,dr,dr ...
23 yh = _mm256_movehdup_ps(y); // Load yh with ci,ci,di,di ...
24 tmp1 = _mm256_mul_ps(x, yl); // tmp1 = ar*cr,ai*cr,br*dr,bi*dr ...
25 x = _mm256_shuffle_ps(x, x, 0xB1); // Re-arrange x to be ai,ar,bi,br ...
26 tmp2 = _mm256_mul_ps(x, yh); // tmp2 = ai*ci,ar*ci,bi*di,br*di
27
28 // ar*cr-ai*ci, ai*cr+ar*ci, br*dr-bi*di, bi*dr+br*di
29 return _mm256_addsub_ps(tmp1, tmp2);
30}
31
32static inline __m256 _mm256_conjugate_ps(__m256 x)
33{
34 const __m256 conjugator = _mm256_setr_ps(0, -0.f, 0, -0.f, 0, -0.f, 0, -0.f);
35 return _mm256_xor_ps(x, conjugator); // conjugate y
36}
37
38static inline __m256 _mm256_complexconjugatemul_ps(const __m256 x, const __m256 y)
39{
40 const __m256 nswap = _mm256_permute_ps(x, 0xb1);
41 const __m256 dreal = _mm256_moveldup_ps(y);
42 const __m256 dimag = _mm256_movehdup_ps(y);
43
44 const __m256 conjugator = _mm256_setr_ps(0, -0.f, 0, -0.f, 0, -0.f, 0, -0.f);
45 const __m256 dimagconj = _mm256_xor_ps(dimag, conjugator);
46 const __m256 multreal = _mm256_mul_ps(x, dreal);
47 const __m256 multimag = _mm256_mul_ps(nswap, dimagconj);
48 return _mm256_add_ps(multreal, multimag);
49}
50
51static inline __m256 _mm256_normalize_ps(__m256 val)
52{
53 __m256 tmp1 = _mm256_mul_ps(val, val);
54 tmp1 = _mm256_hadd_ps(tmp1, tmp1);
55 tmp1 = _mm256_shuffle_ps(tmp1, tmp1, _MM_SHUFFLE(3, 1, 2, 0)); // equals 0xD8
56 tmp1 = _mm256_sqrt_ps(tmp1);
57 return _mm256_div_ps(val, tmp1);
58}
59
60static inline __m256 _mm256_magnitudesquared_ps(__m256 cplxValue1, __m256 cplxValue2)
61{
62 __m256 complex1, complex2;
63 cplxValue1 = _mm256_mul_ps(cplxValue1, cplxValue1); // Square the values
64 cplxValue2 = _mm256_mul_ps(cplxValue2, cplxValue2); // Square the Values
65 complex1 = _mm256_permute2f128_ps(cplxValue1, cplxValue2, 0x20);
66 complex2 = _mm256_permute2f128_ps(cplxValue1, cplxValue2, 0x31);
67 return _mm256_hadd_ps(complex1, complex2); // Add the I2 and Q2 values
68}
69
70static inline __m256 _mm256_magnitude_ps(__m256 cplxValue1, __m256 cplxValue2)
71{
72 return _mm256_sqrt_ps(_mm256_magnitudesquared_ps(cplxValue1, cplxValue2));
73}
74
75static inline __m256 _mm256_scaled_norm_dist_ps(const __m256 symbols0,
76 const __m256 symbols1,
77 const __m256 points0,
78 const __m256 points1,
79 const __m256 scalar)
80{
81 /*
82 * Calculate: |y - x|^2 * SNR_lin
83 * Consider 'symbolsX' and 'pointsX' to be complex float
84 * 'symbolsX' are 'y' and 'pointsX' are 'x'
85 */
86 const __m256 diff0 = _mm256_sub_ps(symbols0, points0);
87 const __m256 diff1 = _mm256_sub_ps(symbols1, points1);
88 const __m256 norms = _mm256_magnitudesquared_ps(diff0, diff1);
89 return _mm256_mul_ps(norms, scalar);
90}
91
92static inline __m256 _mm256_polar_sign_mask(__m128i fbits)
93{
94 __m256 sign_mask_dummy = _mm256_setzero_ps();
95 const __m128i zeros = _mm_set1_epi8(0x00);
96 const __m128i sign_extract = _mm_set1_epi8(0x80);
97 const __m128i shuffle_mask0 = _mm_setr_epi8(0xff,
98 0xff,
99 0xff,
100 0x00,
101 0xff,
102 0xff,
103 0xff,
104 0x01,
105 0xff,
106 0xff,
107 0xff,
108 0x02,
109 0xff,
110 0xff,
111 0xff,
112 0x03);
113 const __m128i shuffle_mask1 = _mm_setr_epi8(0xff,
114 0xff,
115 0xff,
116 0x04,
117 0xff,
118 0xff,
119 0xff,
120 0x05,
121 0xff,
122 0xff,
123 0xff,
124 0x06,
125 0xff,
126 0xff,
127 0xff,
128 0x07);
129
130 fbits = _mm_cmpgt_epi8(fbits, zeros);
131 fbits = _mm_and_si128(fbits, sign_extract);
132 __m128i sign_bits0 = _mm_shuffle_epi8(fbits, shuffle_mask0);
133 __m128i sign_bits1 = _mm_shuffle_epi8(fbits, shuffle_mask1);
134
135 __m256 sign_mask =
136 _mm256_insertf128_ps(sign_mask_dummy, _mm_castsi128_ps(sign_bits0), 0x0);
137 return _mm256_insertf128_ps(sign_mask, _mm_castsi128_ps(sign_bits1), 0x1);
138 // // This is the desired function call. Though it seems to be missing in GCC.
139 // // Compare: https://software.intel.com/sites/landingpage/IntrinsicsGuide/#
140 // return _mm256_set_m128(_mm_castsi128_ps(sign_bits1),
141 // _mm_castsi128_ps(sign_bits0));
142}
143
144static inline void
145_mm256_polar_deinterleave(__m256* llr0, __m256* llr1, __m256 src0, __m256 src1)
146{
147 // deinterleave values
148 __m256 part0 = _mm256_permute2f128_ps(src0, src1, 0x20);
149 __m256 part1 = _mm256_permute2f128_ps(src0, src1, 0x31);
150 *llr0 = _mm256_shuffle_ps(part0, part1, 0x88);
151 *llr1 = _mm256_shuffle_ps(part0, part1, 0xdd);
152}
153
154static inline __m256 _mm256_polar_minsum_llrs(__m256 src0, __m256 src1)
155{
156 const __m256 sign_mask = _mm256_set1_ps(-0.0f);
157 const __m256 abs_mask =
158 _mm256_andnot_ps(sign_mask, _mm256_castsi256_ps(_mm256_set1_epi8(0xff)));
159
160 __m256 llr0, llr1;
161 _mm256_polar_deinterleave(&llr0, &llr1, src0, src1);
162
163 // calculate result
164 __m256 sign =
165 _mm256_xor_ps(_mm256_and_ps(llr0, sign_mask), _mm256_and_ps(llr1, sign_mask));
166 __m256 dst =
167 _mm256_min_ps(_mm256_and_ps(llr0, abs_mask), _mm256_and_ps(llr1, abs_mask));
168 return _mm256_or_ps(dst, sign);
169}
170
171static inline __m256 _mm256_polar_fsign_add_llrs(__m256 src0, __m256 src1, __m128i fbits)
172{
173 // prepare sign mask for correct +-
174 __m256 sign_mask = _mm256_polar_sign_mask(fbits);
175
176 __m256 llr0, llr1;
177 _mm256_polar_deinterleave(&llr0, &llr1, src0, src1);
178
179 // calculate result
180 llr0 = _mm256_xor_ps(llr0, sign_mask);
181 __m256 dst = _mm256_add_ps(llr0, llr1);
182 return dst;
183}
184
186 __m256 sq_acc, __m256 acc, __m256 val, __m256 rec, __m256 aux)
187{
188 aux = _mm256_mul_ps(aux, val);
189 aux = _mm256_sub_ps(aux, acc);
190 aux = _mm256_mul_ps(aux, aux);
191 aux = _mm256_mul_ps(aux, rec);
192 return _mm256_add_ps(sq_acc, aux);
193}
194
195#endif /* INCLUDE_VOLK_VOLK_AVX_INTRINSICS_H_ */