Vector Optimized Library of Kernels  2.4
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 GNU Radio
6  *
7  * GNU Radio is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License as published by
9  * the Free Software Foundation; either version 3, or (at your option)
10  * any later version.
11  *
12  * GNU Radio is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15  * GNU General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with GNU Radio; see the file COPYING. If not, write to
19  * the Free Software Foundation, Inc., 51 Franklin Street,
20  * Boston, MA 02110-1301, USA.
21  */
22 
23 /*
24  * This file is intended to hold AVX intrinsics of intrinsics.
25  * They should be used in VOLK kernels to avoid copy-pasta.
26  */
27 
28 #ifndef INCLUDE_VOLK_VOLK_AVX_INTRINSICS_H_
29 #define INCLUDE_VOLK_VOLK_AVX_INTRINSICS_H_
30 #include <immintrin.h>
31 
32 static inline __m256 _mm256_complexmul_ps(__m256 x, __m256 y)
33 {
34  __m256 yl, yh, tmp1, tmp2;
35  yl = _mm256_moveldup_ps(y); // Load yl with cr,cr,dr,dr ...
36  yh = _mm256_movehdup_ps(y); // Load yh with ci,ci,di,di ...
37  tmp1 = _mm256_mul_ps(x, yl); // tmp1 = ar*cr,ai*cr,br*dr,bi*dr ...
38  x = _mm256_shuffle_ps(x, x, 0xB1); // Re-arrange x to be ai,ar,bi,br ...
39  tmp2 = _mm256_mul_ps(x, yh); // tmp2 = ai*ci,ar*ci,bi*di,br*di
40 
41  // ar*cr-ai*ci, ai*cr+ar*ci, br*dr-bi*di, bi*dr+br*di
42  return _mm256_addsub_ps(tmp1, tmp2);
43 }
44 
45 static inline __m256 _mm256_conjugate_ps(__m256 x)
46 {
47  const __m256 conjugator = _mm256_setr_ps(0, -0.f, 0, -0.f, 0, -0.f, 0, -0.f);
48  return _mm256_xor_ps(x, conjugator); // conjugate y
49 }
50 
51 static inline __m256 _mm256_complexconjugatemul_ps(const __m256 x, const __m256 y)
52 {
53  const __m256 nswap = _mm256_permute_ps(x, 0xb1);
54  const __m256 dreal = _mm256_moveldup_ps(y);
55  const __m256 dimag = _mm256_movehdup_ps(y);
56 
57  const __m256 conjugator = _mm256_setr_ps(0, -0.f, 0, -0.f, 0, -0.f, 0, -0.f);
58  const __m256 dimagconj = _mm256_xor_ps(dimag, conjugator);
59  const __m256 multreal = _mm256_mul_ps(x, dreal);
60  const __m256 multimag = _mm256_mul_ps(nswap, dimagconj);
61  return _mm256_add_ps(multreal, multimag);
62 }
63 
64 static inline __m256 _mm256_normalize_ps(__m256 val)
65 {
66  __m256 tmp1 = _mm256_mul_ps(val, val);
67  tmp1 = _mm256_hadd_ps(tmp1, tmp1);
68  tmp1 = _mm256_shuffle_ps(tmp1, tmp1, _MM_SHUFFLE(3, 1, 2, 0)); // equals 0xD8
69  tmp1 = _mm256_sqrt_ps(tmp1);
70  return _mm256_div_ps(val, tmp1);
71 }
72 
73 static inline __m256 _mm256_magnitudesquared_ps(__m256 cplxValue1, __m256 cplxValue2)
74 {
75  __m256 complex1, complex2;
76  cplxValue1 = _mm256_mul_ps(cplxValue1, cplxValue1); // Square the values
77  cplxValue2 = _mm256_mul_ps(cplxValue2, cplxValue2); // Square the Values
78  complex1 = _mm256_permute2f128_ps(cplxValue1, cplxValue2, 0x20);
79  complex2 = _mm256_permute2f128_ps(cplxValue1, cplxValue2, 0x31);
80  return _mm256_hadd_ps(complex1, complex2); // Add the I2 and Q2 values
81 }
82 
83 static inline __m256 _mm256_magnitude_ps(__m256 cplxValue1, __m256 cplxValue2)
84 {
85  return _mm256_sqrt_ps(_mm256_magnitudesquared_ps(cplxValue1, cplxValue2));
86 }
87 
88 static inline __m256 _mm256_scaled_norm_dist_ps(const __m256 symbols0,
89  const __m256 symbols1,
90  const __m256 points0,
91  const __m256 points1,
92  const __m256 scalar)
93 {
94  /*
95  * Calculate: |y - x|^2 * SNR_lin
96  * Consider 'symbolsX' and 'pointsX' to be complex float
97  * 'symbolsX' are 'y' and 'pointsX' are 'x'
98  */
99  const __m256 diff0 = _mm256_sub_ps(symbols0, points0);
100  const __m256 diff1 = _mm256_sub_ps(symbols1, points1);
101  const __m256 norms = _mm256_magnitudesquared_ps(diff0, diff1);
102  return _mm256_mul_ps(norms, scalar);
103 }
104 
105 static inline __m256 _mm256_polar_sign_mask(__m128i fbits)
106 {
107  __m256 sign_mask_dummy = _mm256_setzero_ps();
108  const __m128i zeros = _mm_set1_epi8(0x00);
109  const __m128i sign_extract = _mm_set1_epi8(0x80);
110  const __m128i shuffle_mask0 = _mm_setr_epi8(0xff,
111  0xff,
112  0xff,
113  0x00,
114  0xff,
115  0xff,
116  0xff,
117  0x01,
118  0xff,
119  0xff,
120  0xff,
121  0x02,
122  0xff,
123  0xff,
124  0xff,
125  0x03);
126  const __m128i shuffle_mask1 = _mm_setr_epi8(0xff,
127  0xff,
128  0xff,
129  0x04,
130  0xff,
131  0xff,
132  0xff,
133  0x05,
134  0xff,
135  0xff,
136  0xff,
137  0x06,
138  0xff,
139  0xff,
140  0xff,
141  0x07);
142 
143  fbits = _mm_cmpgt_epi8(fbits, zeros);
144  fbits = _mm_and_si128(fbits, sign_extract);
145  __m128i sign_bits0 = _mm_shuffle_epi8(fbits, shuffle_mask0);
146  __m128i sign_bits1 = _mm_shuffle_epi8(fbits, shuffle_mask1);
147 
148  __m256 sign_mask =
149  _mm256_insertf128_ps(sign_mask_dummy, _mm_castsi128_ps(sign_bits0), 0x0);
150  return _mm256_insertf128_ps(sign_mask, _mm_castsi128_ps(sign_bits1), 0x1);
151  // // This is the desired function call. Though it seems to be missing in GCC.
152  // // Compare: https://software.intel.com/sites/landingpage/IntrinsicsGuide/#
153  // return _mm256_set_m128(_mm_castsi128_ps(sign_bits1),
154  // _mm_castsi128_ps(sign_bits0));
155 }
156 
157 static inline void
158 _mm256_polar_deinterleave(__m256* llr0, __m256* llr1, __m256 src0, __m256 src1)
159 {
160  // deinterleave values
161  __m256 part0 = _mm256_permute2f128_ps(src0, src1, 0x20);
162  __m256 part1 = _mm256_permute2f128_ps(src0, src1, 0x31);
163  *llr0 = _mm256_shuffle_ps(part0, part1, 0x88);
164  *llr1 = _mm256_shuffle_ps(part0, part1, 0xdd);
165 }
166 
167 static inline __m256 _mm256_polar_minsum_llrs(__m256 src0, __m256 src1)
168 {
169  const __m256 sign_mask = _mm256_set1_ps(-0.0f);
170  const __m256 abs_mask =
171  _mm256_andnot_ps(sign_mask, _mm256_castsi256_ps(_mm256_set1_epi8(0xff)));
172 
173  __m256 llr0, llr1;
174  _mm256_polar_deinterleave(&llr0, &llr1, src0, src1);
175 
176  // calculate result
177  __m256 sign =
178  _mm256_xor_ps(_mm256_and_ps(llr0, sign_mask), _mm256_and_ps(llr1, sign_mask));
179  __m256 dst =
180  _mm256_min_ps(_mm256_and_ps(llr0, abs_mask), _mm256_and_ps(llr1, abs_mask));
181  return _mm256_or_ps(dst, sign);
182 }
183 
184 static inline __m256 _mm256_polar_fsign_add_llrs(__m256 src0, __m256 src1, __m128i fbits)
185 {
186  // prepare sign mask for correct +-
187  __m256 sign_mask = _mm256_polar_sign_mask(fbits);
188 
189  __m256 llr0, llr1;
190  _mm256_polar_deinterleave(&llr0, &llr1, src0, src1);
191 
192  // calculate result
193  llr0 = _mm256_xor_ps(llr0, sign_mask);
194  __m256 dst = _mm256_add_ps(llr0, llr1);
195  return dst;
196 }
197 
198 #endif /* INCLUDE_VOLK_VOLK_AVX_INTRINSICS_H_ */
_mm256_complexmul_ps
static __m256 _mm256_complexmul_ps(__m256 x, __m256 y)
Definition: volk_avx_intrinsics.h:32
_mm256_conjugate_ps
static __m256 _mm256_conjugate_ps(__m256 x)
Definition: volk_avx_intrinsics.h:45
_mm256_scaled_norm_dist_ps
static __m256 _mm256_scaled_norm_dist_ps(const __m256 symbols0, const __m256 symbols1, const __m256 points0, const __m256 points1, const __m256 scalar)
Definition: volk_avx_intrinsics.h:88
volk_arch_defs.val
val
Definition: volk_arch_defs.py:66
_mm256_normalize_ps
static __m256 _mm256_normalize_ps(__m256 val)
Definition: volk_avx_intrinsics.h:64
_mm256_polar_deinterleave
static void _mm256_polar_deinterleave(__m256 *llr0, __m256 *llr1, __m256 src0, __m256 src1)
Definition: volk_avx_intrinsics.h:158
_mm256_magnitudesquared_ps
static __m256 _mm256_magnitudesquared_ps(__m256 cplxValue1, __m256 cplxValue2)
Definition: volk_avx_intrinsics.h:73
_mm256_magnitude_ps
static __m256 _mm256_magnitude_ps(__m256 cplxValue1, __m256 cplxValue2)
Definition: volk_avx_intrinsics.h:83
_mm256_polar_sign_mask
static __m256 _mm256_polar_sign_mask(__m128i fbits)
Definition: volk_avx_intrinsics.h:105
_mm256_complexconjugatemul_ps
static __m256 _mm256_complexconjugatemul_ps(const __m256 x, const __m256 y)
Definition: volk_avx_intrinsics.h:51
_mm256_polar_fsign_add_llrs
static __m256 _mm256_polar_fsign_add_llrs(__m256 src0, __m256 src1, __m128i fbits)
Definition: volk_avx_intrinsics.h:184
_mm256_polar_minsum_llrs
static __m256 _mm256_polar_minsum_llrs(__m256 src0, __m256 src1)
Definition: volk_avx_intrinsics.h:167