Vector Optimized Library of Kernels 3.1.0
Architecture-tuned implementations of math kernels
volk_16i_permute_and_scalar_add.h
Go to the documentation of this file.
1/* -*- c++ -*- */
2/*
3 * Copyright 2012, 2014 Free Software Foundation, Inc.
4 *
5 * This file is part of VOLK
6 *
7 * SPDX-License-Identifier: LGPL-3.0-or-later
8 */
9
46#ifndef INCLUDED_volk_16i_permute_and_scalar_add_a_H
47#define INCLUDED_volk_16i_permute_and_scalar_add_a_H
48
49#include <inttypes.h>
50#include <stdio.h>
51
52#ifdef LV_HAVE_SSE2
53
54#include <emmintrin.h>
55#include <xmmintrin.h>
56
57static inline void volk_16i_permute_and_scalar_add_a_sse2(short* target,
58 short* src0,
59 short* permute_indexes,
60 short* cntl0,
61 short* cntl1,
62 short* cntl2,
63 short* cntl3,
64 short* scalars,
65 unsigned int num_points)
66{
67
68 const unsigned int num_bytes = num_points * 2;
69
70 __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
71
72 __m128i *p_target, *p_cntl0, *p_cntl1, *p_cntl2, *p_cntl3, *p_scalars;
73
74 short* p_permute_indexes = permute_indexes;
75
76 p_target = (__m128i*)target;
77 p_cntl0 = (__m128i*)cntl0;
78 p_cntl1 = (__m128i*)cntl1;
79 p_cntl2 = (__m128i*)cntl2;
80 p_cntl3 = (__m128i*)cntl3;
81 p_scalars = (__m128i*)scalars;
82
83 int i = 0;
84
85 int bound = (num_bytes >> 4);
86 int leftovers = (num_bytes >> 1) & 7;
87
88 xmm0 = _mm_load_si128(p_scalars);
89
90 xmm1 = _mm_shufflelo_epi16(xmm0, 0);
91 xmm2 = _mm_shufflelo_epi16(xmm0, 0x55);
92 xmm3 = _mm_shufflelo_epi16(xmm0, 0xaa);
93 xmm4 = _mm_shufflelo_epi16(xmm0, 0xff);
94
95 xmm1 = _mm_shuffle_epi32(xmm1, 0x00);
96 xmm2 = _mm_shuffle_epi32(xmm2, 0x00);
97 xmm3 = _mm_shuffle_epi32(xmm3, 0x00);
98 xmm4 = _mm_shuffle_epi32(xmm4, 0x00);
99
100
101 for (; i < bound; ++i) {
102 xmm0 = _mm_setzero_si128();
103 xmm5 = _mm_setzero_si128();
104 xmm6 = _mm_setzero_si128();
105 xmm7 = _mm_setzero_si128();
106
107 xmm0 = _mm_insert_epi16(xmm0, src0[p_permute_indexes[0]], 0);
108 xmm5 = _mm_insert_epi16(xmm5, src0[p_permute_indexes[1]], 1);
109 xmm6 = _mm_insert_epi16(xmm6, src0[p_permute_indexes[2]], 2);
110 xmm7 = _mm_insert_epi16(xmm7, src0[p_permute_indexes[3]], 3);
111 xmm0 = _mm_insert_epi16(xmm0, src0[p_permute_indexes[4]], 4);
112 xmm5 = _mm_insert_epi16(xmm5, src0[p_permute_indexes[5]], 5);
113 xmm6 = _mm_insert_epi16(xmm6, src0[p_permute_indexes[6]], 6);
114 xmm7 = _mm_insert_epi16(xmm7, src0[p_permute_indexes[7]], 7);
115
116 xmm0 = _mm_add_epi16(xmm0, xmm5);
117 xmm6 = _mm_add_epi16(xmm6, xmm7);
118
119 p_permute_indexes += 8;
120
121 xmm0 = _mm_add_epi16(xmm0, xmm6);
122
123 xmm5 = _mm_load_si128(p_cntl0);
124 xmm6 = _mm_load_si128(p_cntl1);
125 xmm7 = _mm_load_si128(p_cntl2);
126
127 xmm5 = _mm_and_si128(xmm5, xmm1);
128 xmm6 = _mm_and_si128(xmm6, xmm2);
129 xmm7 = _mm_and_si128(xmm7, xmm3);
130
131 xmm0 = _mm_add_epi16(xmm0, xmm5);
132
133 xmm5 = _mm_load_si128(p_cntl3);
134
135 xmm6 = _mm_add_epi16(xmm6, xmm7);
136
137 p_cntl0 += 1;
138
139 xmm5 = _mm_and_si128(xmm5, xmm4);
140
141 xmm0 = _mm_add_epi16(xmm0, xmm6);
142
143 p_cntl1 += 1;
144 p_cntl2 += 1;
145
146 xmm0 = _mm_add_epi16(xmm0, xmm5);
147
148 p_cntl3 += 1;
149
150 _mm_store_si128(p_target, xmm0);
151
152 p_target += 1;
153 }
154
155 for (i = bound * 8; i < (bound * 8) + leftovers; ++i) {
156 target[i] = src0[permute_indexes[i]] + (cntl0[i] & scalars[0]) +
157 (cntl1[i] & scalars[1]) + (cntl2[i] & scalars[2]) +
158 (cntl3[i] & scalars[3]);
159 }
160}
161#endif /*LV_HAVE_SSE*/
162
163
164#ifdef LV_HAVE_GENERIC
165static inline void volk_16i_permute_and_scalar_add_generic(short* target,
166 short* src0,
167 short* permute_indexes,
168 short* cntl0,
169 short* cntl1,
170 short* cntl2,
171 short* cntl3,
172 short* scalars,
173 unsigned int num_points)
174{
175 const unsigned int num_bytes = num_points * 2;
176
177 int i = 0;
178
179 int bound = num_bytes >> 1;
180
181 for (i = 0; i < bound; ++i) {
182 target[i] = src0[permute_indexes[i]] + (cntl0[i] & scalars[0]) +
183 (cntl1[i] & scalars[1]) + (cntl2[i] & scalars[2]) +
184 (cntl3[i] & scalars[3]);
185 }
186}
187
188#endif /*LV_HAVE_GENERIC*/
189
190#endif /*INCLUDED_volk_16i_permute_and_scalar_add_a_H*/