Vector Optimized Library of Kernels 3.1.0
Architecture-tuned implementations of math kernels
volk_16u_byteswap.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
40#ifndef INCLUDED_volk_16u_byteswap_u_H
41#define INCLUDED_volk_16u_byteswap_u_H
42
43#include <inttypes.h>
44#include <stdio.h>
45
46#ifdef LV_HAVE_GENERIC
47
48static inline void volk_16u_byteswap_generic(uint16_t* intsToSwap,
49 unsigned int num_points)
50{
51 uint16_t* inputPtr = intsToSwap;
52 for (unsigned int point = 0; point < num_points; point++) {
53 uint16_t output = *inputPtr;
54 output = (((output >> 8) & 0xff) | ((output << 8) & 0xff00));
55 *inputPtr = output;
56 inputPtr++;
57 }
58}
59#endif /* LV_HAVE_GENERIC */
60
61
62#if LV_HAVE_AVX2
63#include <immintrin.h>
64static inline void volk_16u_byteswap_a_avx2(uint16_t* intsToSwap, unsigned int num_points)
65{
66 unsigned int number;
67
68 const unsigned int nPerSet = 16;
69 const uint64_t nSets = num_points / nPerSet;
70
71 uint16_t* inputPtr = (uint16_t*)intsToSwap;
72
73 const uint8_t shuffleVector[32] = { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11,
74 10, 13, 12, 15, 14, 17, 16, 19, 18, 21, 20,
75 23, 22, 25, 24, 27, 26, 29, 28, 31, 30 };
76
77 const __m256i myShuffle = _mm256_loadu_si256((__m256i*)&shuffleVector[0]);
78
79 for (number = 0; number < nSets; number++) {
80 // Load the 32t values, increment inputPtr later since we're doing it in-place.
81 const __m256i input = _mm256_load_si256((__m256i*)inputPtr);
82 const __m256i output = _mm256_shuffle_epi8(input, myShuffle);
83
84 // Store the results
85 _mm256_store_si256((__m256i*)inputPtr, output);
86 inputPtr += nPerSet;
87 }
88
89 // Byteswap any remaining points:
90 for (number = nPerSet * nSets; number < num_points; number++) {
91 uint16_t outputVal = *inputPtr;
92 outputVal = (((outputVal >> 8) & 0xff) | ((outputVal << 8) & 0xff00));
93 *inputPtr = outputVal;
94 inputPtr++;
95 }
96}
97#endif /* LV_HAVE_AVX2 */
98
99
100#if LV_HAVE_AVX2
101#include <immintrin.h>
102static inline void volk_16u_byteswap_u_avx2(uint16_t* intsToSwap, unsigned int num_points)
103{
104 unsigned int number;
105
106 const unsigned int nPerSet = 16;
107 const uint64_t nSets = num_points / nPerSet;
108
109 uint16_t* inputPtr = (uint16_t*)intsToSwap;
110
111 const uint8_t shuffleVector[32] = { 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11,
112 10, 13, 12, 15, 14, 17, 16, 19, 18, 21, 20,
113 23, 22, 25, 24, 27, 26, 29, 28, 31, 30 };
114
115 const __m256i myShuffle = _mm256_loadu_si256((__m256i*)&shuffleVector[0]);
116
117 for (number = 0; number < nSets; number++) {
118 // Load the 32t values, increment inputPtr later since we're doing it in-place.
119 const __m256i input = _mm256_loadu_si256((__m256i*)inputPtr);
120 const __m256i output = _mm256_shuffle_epi8(input, myShuffle);
121
122 // Store the results
123 _mm256_storeu_si256((__m256i*)inputPtr, output);
124 inputPtr += nPerSet;
125 }
126
127 // Byteswap any remaining points:
128 for (number = nPerSet * nSets; number < num_points; number++) {
129 uint16_t outputVal = *inputPtr;
130 outputVal = (((outputVal >> 8) & 0xff) | ((outputVal << 8) & 0xff00));
131 *inputPtr = outputVal;
132 inputPtr++;
133 }
134}
135#endif /* LV_HAVE_AVX2 */
136
137
138#ifdef LV_HAVE_SSE2
139#include <emmintrin.h>
140
141static inline void volk_16u_byteswap_u_sse2(uint16_t* intsToSwap, unsigned int num_points)
142{
143 unsigned int number = 0;
144 uint16_t* inputPtr = intsToSwap;
145 __m128i input, left, right, output;
146
147 const unsigned int eighthPoints = num_points / 8;
148 for (; number < eighthPoints; number++) {
149 // Load the 16t values, increment inputPtr later since we're doing it in-place.
150 input = _mm_loadu_si128((__m128i*)inputPtr);
151 // Do the two shifts
152 left = _mm_slli_epi16(input, 8);
153 right = _mm_srli_epi16(input, 8);
154 // Or the left and right halves together
155 output = _mm_or_si128(left, right);
156 // Store the results
157 _mm_storeu_si128((__m128i*)inputPtr, output);
158 inputPtr += 8;
159 }
160
161 // Byteswap any remaining points:
162 number = eighthPoints * 8;
163 for (; number < num_points; number++) {
164 uint16_t outputVal = *inputPtr;
165 outputVal = (((outputVal >> 8) & 0xff) | ((outputVal << 8) & 0xff00));
166 *inputPtr = outputVal;
167 inputPtr++;
168 }
169}
170#endif /* LV_HAVE_SSE2 */
171
172
173#endif /* INCLUDED_volk_16u_byteswap_u_H */
174#ifndef INCLUDED_volk_16u_byteswap_a_H
175#define INCLUDED_volk_16u_byteswap_a_H
176
177#include <inttypes.h>
178#include <stdio.h>
179
180#ifdef LV_HAVE_SSE2
181#include <emmintrin.h>
182
183static inline void volk_16u_byteswap_a_sse2(uint16_t* intsToSwap, unsigned int num_points)
184{
185 uint16_t* inputPtr = intsToSwap;
186 __m128i input, left, right, output;
187
188 const unsigned int eighthPoints = num_points / 8;
189 for (unsigned int number = 0; number < eighthPoints; number++) {
190 // Load the 16t values, increment inputPtr later since we're doing it in-place.
191 input = _mm_load_si128((__m128i*)inputPtr);
192 // Do the two shifts
193 left = _mm_slli_epi16(input, 8);
194 right = _mm_srli_epi16(input, 8);
195 // Or the left and right halves together
196 output = _mm_or_si128(left, right);
197 // Store the results
198 _mm_store_si128((__m128i*)inputPtr, output);
199 inputPtr += 8;
200 }
201
202 // Byteswap any remaining points:
203 volk_16u_byteswap_generic(inputPtr, num_points - eighthPoints * 8);
204}
205#endif /* LV_HAVE_SSE2 */
206
207#ifdef LV_HAVE_NEON
208#include <arm_neon.h>
209
210static inline void volk_16u_byteswap_neon(uint16_t* intsToSwap, unsigned int num_points)
211{
212 unsigned int number;
213 unsigned int eighth_points = num_points / 8;
214 uint16x8_t input, output;
215 uint16_t* inputPtr = intsToSwap;
216
217 for (number = 0; number < eighth_points; number++) {
218 input = vld1q_u16(inputPtr);
219 output = vsriq_n_u16(output, input, 8);
220 output = vsliq_n_u16(output, input, 8);
221 vst1q_u16(inputPtr, output);
222 inputPtr += 8;
223 }
224
225 volk_16u_byteswap_generic(inputPtr, num_points - eighth_points * 8);
226}
227#endif /* LV_HAVE_NEON */
228
229#ifdef LV_HAVE_NEON
230#include <arm_neon.h>
231
232static inline void volk_16u_byteswap_neon_table(uint16_t* intsToSwap,
233 unsigned int num_points)
234{
235 uint16_t* inputPtr = intsToSwap;
236 unsigned int number = 0;
237 unsigned int n16points = num_points / 16;
238
239 uint8x8x4_t input_table;
240 uint8x8_t int_lookup01, int_lookup23, int_lookup45, int_lookup67;
241 uint8x8_t swapped_int01, swapped_int23, swapped_int45, swapped_int67;
242
243 /* these magic numbers are used as byte-indices in the LUT.
244 they are pre-computed to save time. A simple C program
245 can calculate them; for example for lookup01:
246 uint8_t chars[8] = {24, 16, 8, 0, 25, 17, 9, 1};
247 for(ii=0; ii < 8; ++ii) {
248 index += ((uint64_t)(*(chars+ii))) << (ii*8);
249 }
250 */
251 int_lookup01 = vcreate_u8(1232017111498883080);
252 int_lookup23 = vcreate_u8(1376697457175036426);
253 int_lookup45 = vcreate_u8(1521377802851189772);
254 int_lookup67 = vcreate_u8(1666058148527343118);
255
256 for (number = 0; number < n16points; ++number) {
257 input_table = vld4_u8((uint8_t*)inputPtr);
258 swapped_int01 = vtbl4_u8(input_table, int_lookup01);
259 swapped_int23 = vtbl4_u8(input_table, int_lookup23);
260 swapped_int45 = vtbl4_u8(input_table, int_lookup45);
261 swapped_int67 = vtbl4_u8(input_table, int_lookup67);
262 vst1_u8((uint8_t*)inputPtr, swapped_int01);
263 vst1_u8((uint8_t*)(inputPtr + 4), swapped_int23);
264 vst1_u8((uint8_t*)(inputPtr + 8), swapped_int45);
265 vst1_u8((uint8_t*)(inputPtr + 12), swapped_int67);
266
267 inputPtr += 16;
268 }
269
270 volk_16u_byteswap_generic(inputPtr, num_points - n16points * 16);
271}
272#endif /* LV_HAVE_NEON */
273
274#ifdef LV_HAVE_ORC
275
276extern void volk_16u_byteswap_a_orc_impl(uint16_t* intsToSwap, unsigned int num_points);
277static inline void volk_16u_byteswap_u_orc(uint16_t* intsToSwap, unsigned int num_points)
278{
279 volk_16u_byteswap_a_orc_impl(intsToSwap, num_points);
280}
281#endif /* LV_HAVE_ORC */
282
283
284#endif /* INCLUDED_volk_16u_byteswap_a_H */