OpenJPH
Open-source implementation of JPEG2000 Part-15
Loading...
Searching...
No Matches
ojph_codestream_sse2.cpp
Go to the documentation of this file.
1//***************************************************************************/
2// This software is released under the 2-Clause BSD license, included
3// below.
4//
5// Copyright (c) 2022, Aous Naman
6// Copyright (c) 2022, Kakadu Software Pty Ltd, Australia
7// Copyright (c) 2022, The University of New South Wales, Australia
8//
9// Redistribution and use in source and binary forms, with or without
10// modification, are permitted provided that the following conditions are
11// met:
12//
13// 1. Redistributions of source code must retain the above copyright
14// notice, this list of conditions and the following disclaimer.
15//
16// 2. Redistributions in binary form must reproduce the above copyright
17// notice, this list of conditions and the following disclaimer in the
18// documentation and/or other materials provided with the distribution.
19//
20// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
21// IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22// TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
23// PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
24// HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
25// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
26// TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
27// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
28// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
29// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
30// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31//***************************************************************************/
32// This file is part of the OpenJPH software implementation.
33// File: ojph_codestream_sse2.cpp
34// Author: Aous Naman
35// Date: 15 May 2022
36//***************************************************************************/
37
38#include <climits>
39#include <immintrin.h>
40#include "ojph_defs.h"
41
42namespace ojph {
43 namespace local {
44
47 {
48 __m128i x1, x0 = _mm_loadu_si128((__m128i*)address);
49 x1 = _mm_shuffle_epi32(x0, 0xEE); // x1 = x0[2,3,2,3]
50 x0 = _mm_or_si128(x0, x1);
51 x1 = _mm_shuffle_epi32(x0, 0x55); // x1 = x0[1,1,1,1]
52 x0 = _mm_or_si128(x0, x1);
53 _mm_storeu_si128((__m128i*)address, x0);
54 return *address;
55 // A single movd t, xmm0 can do the trick, but it is not available
56 // in SSE2 intrinsics. extract_epi32 is available in sse4.1
57 // ui32 t = (ui32)_mm_extract_epi16(x0, 0);
58 // t |= (ui32)_mm_extract_epi16(x0, 1) << 16;
59 // return t;
60 }
61
64 {
65 __m128i x1, x0 = _mm_loadu_si128((__m128i*)address);
66 x1 = _mm_shuffle_epi32(x0, 0xEE); // x1 = x0[2,3,2,3]
67 x0 = _mm_or_si128(x0, x1);
68 _mm_storeu_si128((__m128i*)address, x0);
69 return *address;
70 // A single movd t, xmm0 can do the trick, but it is not available
71 // in SSE2 intrinsics. extract_epi32 is available in sse4.1
72 // ui32 t = (ui32)_mm_extract_epi16(x0, 0);
73 // t |= (ui32)_mm_extract_epi16(x0, 1) << 16;
74 // return t;
75 }
76
78 void sse2_rev_tx_to_cb32(const void *sp, ui32 *dp, ui32 K_max,
79 float delta_inv, ui32 count, ui32* max_val)
80 {
81 ojph_unused(delta_inv);
82
83 // convert to sign and magnitude and keep max_val
84 ui32 shift = 31 - K_max;
85 __m128i m0 = _mm_set1_epi32(INT_MIN);
86 __m128i zero = _mm_setzero_si128();
87 __m128i one = _mm_set1_epi32(1);
88 __m128i tmax = _mm_loadu_si128((__m128i*)max_val);
89 __m128i *p = (__m128i*)sp;
90 for ( ; count >= 4; count -= 4, p += 1, dp += 4)
91 {
92 __m128i v = _mm_loadu_si128(p);
93 __m128i sign = _mm_cmplt_epi32(v, zero);
94 __m128i val = _mm_xor_si128(v, sign); // negate 1's complement
95 __m128i ones = _mm_and_si128(sign, one);
96 val = _mm_add_epi32(val, ones); // 2's complement
97 sign = _mm_and_si128(sign, m0);
98 val = _mm_slli_epi32(val, (int)shift);
99 tmax = _mm_or_si128(tmax, val);
100 val = _mm_or_si128(val, sign);
101 _mm_storeu_si128((__m128i*)dp, val);
102 }
103 if (count)
104 {
105 __m128i v = _mm_loadu_si128(p);
106 __m128i sign = _mm_cmplt_epi32(v, zero);
107 __m128i val = _mm_xor_si128(v, sign); // negate 1's complement
108 __m128i ones = _mm_and_si128(sign, one);
109 val = _mm_add_epi32(val, ones); // 2's complement
110 sign = _mm_and_si128(sign, m0);
111 val = _mm_slli_epi32(val, (int)shift);
112
113 __m128i c = _mm_set1_epi32((si32)count);
114 __m128i idx = _mm_set_epi32(3, 2, 1, 0);
115 __m128i mask = _mm_cmpgt_epi32(c, idx);
116 c = _mm_and_si128(val, mask);
117 tmax = _mm_or_si128(tmax, c);
118
119 val = _mm_or_si128(val, sign);
120 _mm_storeu_si128((__m128i*)dp, val);
121 }
122 _mm_storeu_si128((__m128i*)max_val, tmax);
123 }
124
126 void sse2_irv_tx_to_cb32(const void *sp, ui32 *dp, ui32 K_max,
127 float delta_inv, ui32 count, ui32* max_val)
128 {
129 ojph_unused(K_max);
130
131 //quantize and convert to sign and magnitude and keep max_val
132
133 __m128 d = _mm_set1_ps(delta_inv);
134 __m128i zero = _mm_setzero_si128();
135 __m128i one = _mm_set1_epi32(1);
136 __m128i tmax = _mm_loadu_si128((__m128i*)max_val);
137 float *p = (float*)sp;
138 for ( ; count >= 4; count -= 4, p += 4, dp += 4)
139 {
140 __m128 vf = _mm_loadu_ps(p);
141 vf = _mm_mul_ps(vf, d); // multiply
142 __m128i val = _mm_cvtps_epi32(vf); // convert to int
143 __m128i sign = _mm_cmplt_epi32(val, zero); // get sign
144 val = _mm_xor_si128(val, sign); // negate 1's complement
145 __m128i ones = _mm_and_si128(sign, one);
146 val = _mm_add_epi32(val, ones); // 2's complement
147 tmax = _mm_or_si128(tmax, val);
148 sign = _mm_slli_epi32(sign, 31);
149 val = _mm_or_si128(val, sign);
150 _mm_storeu_si128((__m128i*)dp, val);
151 }
152 if (count)
153 {
154 __m128 vf = _mm_loadu_ps(p);
155 vf = _mm_mul_ps(vf, d); // multiply
156 __m128i val = _mm_cvtps_epi32(vf); // convert to int
157 __m128i sign = _mm_cmplt_epi32(val, zero); // get sign
158 val = _mm_xor_si128(val, sign); // negate 1's complement
159 __m128i ones = _mm_and_si128(sign, one);
160 val = _mm_add_epi32(val, ones); // 2's complement
161
162 __m128i c = _mm_set1_epi32((si32)count);
163 __m128i idx = _mm_set_epi32(3, 2, 1, 0);
164 __m128i mask = _mm_cmpgt_epi32(c, idx);
165 c = _mm_and_si128(val, mask);
166 tmax = _mm_or_si128(tmax, c);
167
168 sign = _mm_slli_epi32(sign, 31);
169 val = _mm_or_si128(val, sign);
170 _mm_storeu_si128((__m128i*)dp, val);
171 }
172 _mm_storeu_si128((__m128i*)max_val, tmax);
173 }
174
176 void sse2_rev_tx_from_cb32(const ui32 *sp, void *dp, ui32 K_max,
177 float delta, ui32 count)
178 {
179 ojph_unused(delta);
180 ui32 shift = 31 - K_max;
181 __m128i m1 = _mm_set1_epi32(INT_MAX);
182 __m128i zero = _mm_setzero_si128();
183 __m128i one = _mm_set1_epi32(1);
184 si32 *p = (si32*)dp;
185 for (ui32 i = 0; i < count; i += 4, sp += 4, p += 4)
186 {
187 __m128i v = _mm_load_si128((__m128i*)sp);
188 __m128i val = _mm_and_si128(v, m1);
189 val = _mm_srli_epi32(val, (int)shift);
190 __m128i sign = _mm_cmplt_epi32(v, zero);
191 val = _mm_xor_si128(val, sign); // negate 1's complement
192 __m128i ones = _mm_and_si128(sign, one);
193 val = _mm_add_epi32(val, ones); // 2's complement
194 _mm_storeu_si128((__m128i*)p, val);
195 }
196 }
197
199 void sse2_irv_tx_from_cb32(const ui32 *sp, void *dp, ui32 K_max,
200 float delta, ui32 count)
201 {
202 ojph_unused(K_max);
203 __m128i m1 = _mm_set1_epi32(INT_MAX);
204 __m128 d = _mm_set1_ps(delta);
205 float *p = (float*)dp;
206 for (ui32 i = 0; i < count; i += 4, sp += 4, p += 4)
207 {
208 __m128i v = _mm_load_si128((__m128i*)sp);
209 __m128i vali = _mm_and_si128(v, m1);
210 __m128 valf = _mm_cvtepi32_ps(vali);
211 valf = _mm_mul_ps(valf, d);
212 __m128i sign = _mm_andnot_si128(m1, v);
213 valf = _mm_or_ps(valf, _mm_castsi128_ps(sign));
214 _mm_storeu_ps(p, valf);
215 }
216 }
217
219 void sse2_rev_tx_to_cb64(const void *sp, ui64 *dp, ui32 K_max,
220 float delta_inv, ui32 count, ui64* max_val)
221 {
222 ojph_unused(delta_inv);
223
224 // convert to sign and magnitude and keep max_val
225 ui32 shift = 63 - K_max;
226 __m128i m0 = _mm_set1_epi64x(LLONG_MIN);
227 __m128i zero = _mm_setzero_si128();
228 __m128i one = _mm_set1_epi64x(1);
229 __m128i tmax = _mm_loadu_si128((__m128i*)max_val);
230 __m128i *p = (__m128i*)sp;
231 for ( ; count >= 2; count -= 2, p += 1, dp += 2)
232 {
233 __m128i v = _mm_loadu_si128(p);
234 __m128i sign = _mm_cmplt_epi32(v, zero);
235 sign = _mm_shuffle_epi32(sign, 0xF5); // sign = sign[1,1,3,3];
236 __m128i val = _mm_xor_si128(v, sign); // negate 1's complement
237 __m128i ones = _mm_and_si128(sign, one);
238 val = _mm_add_epi64(val, ones); // 2's complement
239 sign = _mm_and_si128(sign, m0);
240 val = _mm_slli_epi64(val, (int)shift);
241 tmax = _mm_or_si128(tmax, val);
242 val = _mm_or_si128(val, sign);
243 _mm_storeu_si128((__m128i*)dp, val);
244 }
245 if (count)
246 {
247 __m128i v = _mm_loadu_si128(p);
248 __m128i sign = _mm_cmplt_epi32(v, zero);
249 sign = _mm_shuffle_epi32(sign, 0xF5); // sign = sign[1,1,3,3];
250 __m128i val = _mm_xor_si128(v, sign); // negate 1's complement
251 __m128i ones = _mm_and_si128(sign, one);
252 val = _mm_add_epi64(val, ones); // 2's complement
253 sign = _mm_and_si128(sign, m0);
254 val = _mm_slli_epi64(val, (int)shift);
255
256 __m128i c = _mm_set_epi32(0, 0, (si32)0xFFFFFFFF, (si32)0xFFFFFFFF);
257 c = _mm_and_si128(val, c);
258 tmax = _mm_or_si128(tmax, c);
259
260 val = _mm_or_si128(val, sign);
261 _mm_storeu_si128((__m128i*)dp, val);
262 }
263 _mm_storeu_si128((__m128i*)max_val, tmax);
264 }
265
267 void sse2_rev_tx_from_cb64(const ui64 *sp, void *dp, ui32 K_max,
268 float delta, ui32 count)
269 {
270 ojph_unused(delta);
271 ui32 shift = 63 - K_max;
272 __m128i m1 = _mm_set1_epi64x(LLONG_MAX);
273 __m128i zero = _mm_setzero_si128();
274 __m128i one = _mm_set1_epi64x(1);
275 si64 *p = (si64*)dp;
276 for (ui32 i = 0; i < count; i += 2, sp += 2, p += 2)
277 {
278 __m128i v = _mm_load_si128((__m128i*)sp);
279 __m128i val = _mm_and_si128(v, m1);
280 val = _mm_srli_epi64(val, (int)shift);
281 __m128i sign = _mm_cmplt_epi32(v, zero);
282 sign = _mm_shuffle_epi32(sign, 0xF5); // sign = sign[1,1,3,3];
283 val = _mm_xor_si128(val, sign); // negate 1's complement
284 __m128i ones = _mm_and_si128(sign, one);
285 val = _mm_add_epi64(val, ones); // 2's complement
286 _mm_storeu_si128((__m128i*)p, val);
287 }
288 }
289 }
290}
void sse2_rev_tx_from_cb32(const ui32 *sp, void *dp, ui32 K_max, float delta, ui32 count)
ui32 sse2_find_max_val32(ui32 *address)
ui64 sse2_find_max_val64(ui64 *address)
void sse2_rev_tx_to_cb32(const void *sp, ui32 *dp, ui32 K_max, float delta_inv, ui32 count, ui32 *max_val)
void sse2_rev_tx_to_cb64(const void *sp, ui64 *dp, ui32 K_max, float delta_inv, ui32 count, ui64 *max_val)
void sse2_irv_tx_from_cb32(const ui32 *sp, void *dp, ui32 K_max, float delta, ui32 count)
void sse2_rev_tx_from_cb64(const ui64 *sp, void *dp, ui32 K_max, float delta, ui32 count)
void sse2_irv_tx_to_cb32(const void *sp, ui32 *dp, ui32 K_max, float delta_inv, ui32 count, ui32 *max_val)
int64_t si64
Definition ojph_defs.h:57
uint64_t ui64
Definition ojph_defs.h:56
int32_t si32
Definition ojph_defs.h:55
uint32_t ui32
Definition ojph_defs.h:54
#define ojph_unused(x)
Definition ojph_defs.h:78