sl@0
|
1 |
/*
|
sl@0
|
2 |
* Copyright (c) 2005
|
sl@0
|
3 |
* Eric Anholt. All rights reserved.
|
sl@0
|
4 |
*
|
sl@0
|
5 |
* Redistribution and use in source and binary forms, with or without
|
sl@0
|
6 |
* modification, are permitted provided that the following conditions
|
sl@0
|
7 |
* are met:
|
sl@0
|
8 |
* 1. Redistributions of source code must retain the above copyright
|
sl@0
|
9 |
* notice, this list of conditions and the following disclaimer.
|
sl@0
|
10 |
* 2. Redistributions in binary form must reproduce the above copyright
|
sl@0
|
11 |
* notice, this list of conditions and the following disclaimer in the
|
sl@0
|
12 |
* documentation and/or other materials provided with the distribution.
|
sl@0
|
13 |
*
|
sl@0
|
14 |
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND
|
sl@0
|
15 |
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
sl@0
|
16 |
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
|
sl@0
|
17 |
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE
|
sl@0
|
18 |
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
sl@0
|
19 |
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
|
sl@0
|
20 |
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
|
sl@0
|
21 |
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
|
sl@0
|
22 |
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
|
sl@0
|
23 |
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
|
sl@0
|
24 |
* SUCH DAMAGE.
|
sl@0
|
25 |
*/
|
sl@0
|
26 |
//Portions Copyright (c) 2008-2009 Nokia Corporation and/or its subsidiary(-ies). All rights reserved.
|
sl@0
|
27 |
|
sl@0
|
28 |
#ifdef HAVE_CONFIG_H
|
sl@0
|
29 |
#include "config.h"
|
sl@0
|
30 |
#endif
|
sl@0
|
31 |
#include "liboil/liboilclasses.h"
|
sl@0
|
32 |
#include "liboil/liboilfunction.h"
|
sl@0
|
33 |
#include <emmintrin.h>
|
sl@0
|
34 |
#include <xmmintrin.h>
|
sl@0
|
35 |
|
sl@0
|
36 |
/* TODO: If we have gcc 4.2 or above, do this. Otherwise, disable all SSE use */
|
sl@0
|
37 |
#define SSE_FUNCTION __attribute__((force_align_arg_pointer))
|
sl@0
|
38 |
|
sl@0
|
39 |
SSE_FUNCTION static void
|
sl@0
|
40 |
clamp_u8_sse (uint8_t *dest, uint8_t *src1, int n, uint8_t *src2_1,
|
sl@0
|
41 |
uint8_t *src3_1)
|
sl@0
|
42 |
{
|
sl@0
|
43 |
__m128i xmm1, xmm2;
|
sl@0
|
44 |
uint8_t min = *src2_1;
|
sl@0
|
45 |
uint8_t max = *src3_1;
|
sl@0
|
46 |
|
sl@0
|
47 |
/* Initial operations to align the destination pointer */
|
sl@0
|
48 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
49 |
uint8_t x = *src1++;
|
sl@0
|
50 |
if (x < min)
|
sl@0
|
51 |
x = min;
|
sl@0
|
52 |
if (x > max)
|
sl@0
|
53 |
x = max;
|
sl@0
|
54 |
*dest++ = x;
|
sl@0
|
55 |
}
|
sl@0
|
56 |
xmm1 = _mm_set1_epi8(min);
|
sl@0
|
57 |
xmm2 = _mm_set1_epi8(max);
|
sl@0
|
58 |
for (; n >= 16; n -= 16) {
|
sl@0
|
59 |
__m128i xmm0;
|
sl@0
|
60 |
xmm0 = _mm_loadu_si128((__m128i *)src1);
|
sl@0
|
61 |
xmm0 = _mm_max_epu8(xmm0, xmm1);
|
sl@0
|
62 |
xmm0 = _mm_min_epu8(xmm0, xmm2);
|
sl@0
|
63 |
_mm_store_si128((__m128i *)dest, xmm0);
|
sl@0
|
64 |
dest += 16;
|
sl@0
|
65 |
src1 += 16;
|
sl@0
|
66 |
}
|
sl@0
|
67 |
for (; n > 0; n--) {
|
sl@0
|
68 |
uint8_t x = *src1++;
|
sl@0
|
69 |
if (x < min)
|
sl@0
|
70 |
x = min;
|
sl@0
|
71 |
if (x > max)
|
sl@0
|
72 |
x = max;
|
sl@0
|
73 |
*dest++ = x;
|
sl@0
|
74 |
}
|
sl@0
|
75 |
}
|
sl@0
|
76 |
OIL_DEFINE_IMPL_FULL (clamp_u8_sse, clamp_u8, OIL_IMPL_FLAG_SSE2);
|
sl@0
|
77 |
|
sl@0
|
78 |
SSE_FUNCTION static void
|
sl@0
|
79 |
clamp_s16_sse (int16_t *dest, int16_t *src1, int n, int16_t *src2_1,
|
sl@0
|
80 |
int16_t *src3_1)
|
sl@0
|
81 |
{
|
sl@0
|
82 |
__m128i xmm1, xmm2;
|
sl@0
|
83 |
int16_t min = *src2_1;
|
sl@0
|
84 |
int16_t max = *src3_1;
|
sl@0
|
85 |
|
sl@0
|
86 |
/* Initial operations to align the destination pointer */
|
sl@0
|
87 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
88 |
int16_t x = *src1++;
|
sl@0
|
89 |
if (x < min)
|
sl@0
|
90 |
x = min;
|
sl@0
|
91 |
if (x > max)
|
sl@0
|
92 |
x = max;
|
sl@0
|
93 |
*dest++ = x;
|
sl@0
|
94 |
}
|
sl@0
|
95 |
xmm1 = _mm_set1_epi16(min);
|
sl@0
|
96 |
xmm2 = _mm_set1_epi16(max);
|
sl@0
|
97 |
for (; n >= 8; n -= 8) {
|
sl@0
|
98 |
__m128i xmm0;
|
sl@0
|
99 |
xmm0 = _mm_loadu_si128((__m128i *)src1);
|
sl@0
|
100 |
xmm0 = _mm_max_epi16(xmm0, xmm1);
|
sl@0
|
101 |
xmm0 = _mm_min_epi16(xmm0, xmm2);
|
sl@0
|
102 |
_mm_store_si128((__m128i *)dest, xmm0);
|
sl@0
|
103 |
dest += 8;
|
sl@0
|
104 |
src1 += 8;
|
sl@0
|
105 |
}
|
sl@0
|
106 |
for (; n > 0; n--) {
|
sl@0
|
107 |
int16_t x = *src1++;
|
sl@0
|
108 |
if (x < min)
|
sl@0
|
109 |
x = min;
|
sl@0
|
110 |
if (x > max)
|
sl@0
|
111 |
x = max;
|
sl@0
|
112 |
*dest++ = x;
|
sl@0
|
113 |
}
|
sl@0
|
114 |
}
|
sl@0
|
115 |
OIL_DEFINE_IMPL_FULL (clamp_s16_sse, clamp_s16, OIL_IMPL_FLAG_SSE2);
|
sl@0
|
116 |
|
sl@0
|
117 |
SSE_FUNCTION static void
|
sl@0
|
118 |
clamp_f32_sse (float *dest, const float *src1, int n, const float *src2_1,
|
sl@0
|
119 |
const float *src3_1)
|
sl@0
|
120 |
{
|
sl@0
|
121 |
__m128 xmm1, xmm2;
|
sl@0
|
122 |
float min = *src2_1;
|
sl@0
|
123 |
float max = *src3_1;
|
sl@0
|
124 |
|
sl@0
|
125 |
/* Initial operations to align the destination pointer */
|
sl@0
|
126 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
127 |
float x = *src1++;
|
sl@0
|
128 |
if (x < min)
|
sl@0
|
129 |
x = min;
|
sl@0
|
130 |
if (x > max)
|
sl@0
|
131 |
x = max;
|
sl@0
|
132 |
*dest++ = x;
|
sl@0
|
133 |
}
|
sl@0
|
134 |
xmm1 = _mm_set_ps1(min);
|
sl@0
|
135 |
xmm2 = _mm_set_ps1(max);
|
sl@0
|
136 |
for (; n >= 4; n -= 4) {
|
sl@0
|
137 |
__m128 xmm0;
|
sl@0
|
138 |
xmm0 = _mm_loadu_ps(src1);
|
sl@0
|
139 |
xmm0 = _mm_max_ps(xmm0, xmm1);
|
sl@0
|
140 |
xmm0 = _mm_min_ps(xmm0, xmm2);
|
sl@0
|
141 |
_mm_store_ps(dest, xmm0);
|
sl@0
|
142 |
dest += 4;
|
sl@0
|
143 |
src1 += 4;
|
sl@0
|
144 |
}
|
sl@0
|
145 |
for (; n > 0; n--) {
|
sl@0
|
146 |
float x = *src1++;
|
sl@0
|
147 |
if (x < min)
|
sl@0
|
148 |
x = min;
|
sl@0
|
149 |
if (x > max)
|
sl@0
|
150 |
x = max;
|
sl@0
|
151 |
*dest++ = x;
|
sl@0
|
152 |
}
|
sl@0
|
153 |
}
|
sl@0
|
154 |
OIL_DEFINE_IMPL_FULL (clamp_f32_sse, clamp_f32, OIL_IMPL_FLAG_SSE);
|
sl@0
|
155 |
|
sl@0
|
156 |
SSE_FUNCTION static void
|
sl@0
|
157 |
clamp_f64_sse (double *dest, const double *src1, int n, const double *src2_1,
|
sl@0
|
158 |
const double *src3_1)
|
sl@0
|
159 |
{
|
sl@0
|
160 |
__m128d xmm1, xmm2;
|
sl@0
|
161 |
double min = *src2_1;
|
sl@0
|
162 |
double max = *src3_1;
|
sl@0
|
163 |
|
sl@0
|
164 |
/* Initial operations to align the destination pointer */
|
sl@0
|
165 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
166 |
double x = *src1++;
|
sl@0
|
167 |
if (x < min)
|
sl@0
|
168 |
x = min;
|
sl@0
|
169 |
if (x > max)
|
sl@0
|
170 |
x = max;
|
sl@0
|
171 |
*dest++ = x;
|
sl@0
|
172 |
}
|
sl@0
|
173 |
xmm1 = _mm_set1_pd(min);
|
sl@0
|
174 |
xmm2 = _mm_set1_pd(max);
|
sl@0
|
175 |
for (; n >= 2; n -= 2) {
|
sl@0
|
176 |
__m128d xmm0;
|
sl@0
|
177 |
xmm0 = _mm_loadu_pd(src1);
|
sl@0
|
178 |
xmm0 = _mm_max_pd(xmm0, xmm1);
|
sl@0
|
179 |
xmm0 = _mm_min_pd(xmm0, xmm2);
|
sl@0
|
180 |
_mm_store_pd(dest, xmm0);
|
sl@0
|
181 |
dest += 2;
|
sl@0
|
182 |
src1 += 2;
|
sl@0
|
183 |
}
|
sl@0
|
184 |
for (; n > 0; n--) {
|
sl@0
|
185 |
double x = *src1++;
|
sl@0
|
186 |
if (x < min)
|
sl@0
|
187 |
x = min;
|
sl@0
|
188 |
if (x > max)
|
sl@0
|
189 |
x = max;
|
sl@0
|
190 |
*dest++ = x;
|
sl@0
|
191 |
}
|
sl@0
|
192 |
}
|
sl@0
|
193 |
OIL_DEFINE_IMPL_FULL (clamp_f64_sse, clamp_f64,
|
sl@0
|
194 |
OIL_IMPL_FLAG_SSE | OIL_IMPL_FLAG_SSE2);
|
sl@0
|
195 |
|
sl@0
|
196 |
SSE_FUNCTION static void
|
sl@0
|
197 |
clamplow_u8_sse (uint8_t *dest, const uint8_t *src1, int n,
|
sl@0
|
198 |
const uint8_t *src2_1)
|
sl@0
|
199 |
{
|
sl@0
|
200 |
__m128i xmm1;
|
sl@0
|
201 |
uint8_t min = *src2_1;
|
sl@0
|
202 |
|
sl@0
|
203 |
/* Initial operations to align the destination pointer */
|
sl@0
|
204 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
205 |
uint8_t x = *src1++;
|
sl@0
|
206 |
if (x < min)
|
sl@0
|
207 |
x = min;
|
sl@0
|
208 |
*dest++ = x;
|
sl@0
|
209 |
}
|
sl@0
|
210 |
xmm1 = _mm_set1_epi8(min);
|
sl@0
|
211 |
for (; n >= 16; n -= 16) {
|
sl@0
|
212 |
__m128i xmm0;
|
sl@0
|
213 |
xmm0 = _mm_loadu_si128((__m128i *)src1);
|
sl@0
|
214 |
xmm0 = _mm_max_epu8(xmm0, xmm1);
|
sl@0
|
215 |
_mm_store_si128((__m128i *)dest, xmm0);
|
sl@0
|
216 |
dest += 16;
|
sl@0
|
217 |
src1 += 16;
|
sl@0
|
218 |
}
|
sl@0
|
219 |
for (; n > 0; n--) {
|
sl@0
|
220 |
uint8_t x = *src1++;
|
sl@0
|
221 |
if (x < min)
|
sl@0
|
222 |
x = min;
|
sl@0
|
223 |
*dest++ = x;
|
sl@0
|
224 |
}
|
sl@0
|
225 |
}
|
sl@0
|
226 |
OIL_DEFINE_IMPL_FULL (clamplow_u8_sse, clamplow_u8, OIL_IMPL_FLAG_SSE2);
|
sl@0
|
227 |
|
sl@0
|
228 |
SSE_FUNCTION static void
|
sl@0
|
229 |
clamplow_s16_sse (int16_t *dest, const int16_t *src1, int n,
|
sl@0
|
230 |
const int16_t *src2_1)
|
sl@0
|
231 |
{
|
sl@0
|
232 |
__m128i xmm1;
|
sl@0
|
233 |
int16_t min = *src2_1;
|
sl@0
|
234 |
|
sl@0
|
235 |
/* Initial operations to align the destination pointer */
|
sl@0
|
236 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
237 |
int16_t x = *src1++;
|
sl@0
|
238 |
if (x < min)
|
sl@0
|
239 |
x = min;
|
sl@0
|
240 |
*dest++ = x;
|
sl@0
|
241 |
}
|
sl@0
|
242 |
xmm1 = _mm_set1_epi16(min);
|
sl@0
|
243 |
for (; n >= 8; n -= 8) {
|
sl@0
|
244 |
__m128i xmm0;
|
sl@0
|
245 |
xmm0 = _mm_loadu_si128((__m128i *)src1);
|
sl@0
|
246 |
xmm0 = _mm_max_epi16(xmm0, xmm1);
|
sl@0
|
247 |
_mm_store_si128((__m128i *)dest, xmm0);
|
sl@0
|
248 |
dest += 8;
|
sl@0
|
249 |
src1 += 8;
|
sl@0
|
250 |
}
|
sl@0
|
251 |
for (; n > 0; n--) {
|
sl@0
|
252 |
int16_t x = *src1++;
|
sl@0
|
253 |
if (x < min)
|
sl@0
|
254 |
x = min;
|
sl@0
|
255 |
*dest++ = x;
|
sl@0
|
256 |
}
|
sl@0
|
257 |
}
|
sl@0
|
258 |
OIL_DEFINE_IMPL_FULL (clamplow_s16_sse, clamplow_s16, OIL_IMPL_FLAG_SSE2);
|
sl@0
|
259 |
|
sl@0
|
260 |
SSE_FUNCTION static void
|
sl@0
|
261 |
clamplow_f32_sse (float *dest, const float *src1, int n, const float *src2_1)
|
sl@0
|
262 |
{
|
sl@0
|
263 |
__m128 xmm1;
|
sl@0
|
264 |
float min = *src2_1;
|
sl@0
|
265 |
|
sl@0
|
266 |
/* Initial operations to align the destination pointer */
|
sl@0
|
267 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
268 |
float x = *src1++;
|
sl@0
|
269 |
if (x < min)
|
sl@0
|
270 |
x = min;
|
sl@0
|
271 |
*dest++ = x;
|
sl@0
|
272 |
}
|
sl@0
|
273 |
xmm1 = _mm_set_ps1(min);
|
sl@0
|
274 |
for (; n >= 4; n -= 4) {
|
sl@0
|
275 |
__m128 xmm0;
|
sl@0
|
276 |
xmm0 = _mm_loadu_ps(src1);
|
sl@0
|
277 |
xmm0 = _mm_max_ps(xmm0, xmm1);
|
sl@0
|
278 |
_mm_store_ps(dest, xmm0);
|
sl@0
|
279 |
dest += 4;
|
sl@0
|
280 |
src1 += 4;
|
sl@0
|
281 |
}
|
sl@0
|
282 |
for (; n > 0; n--) {
|
sl@0
|
283 |
float x = *src1++;
|
sl@0
|
284 |
if (x < min)
|
sl@0
|
285 |
x = min;
|
sl@0
|
286 |
*dest++ = x;
|
sl@0
|
287 |
}
|
sl@0
|
288 |
}
|
sl@0
|
289 |
OIL_DEFINE_IMPL_FULL (clamplow_f32_sse, clamplow_f32, OIL_IMPL_FLAG_SSE);
|
sl@0
|
290 |
|
sl@0
|
291 |
SSE_FUNCTION static void
|
sl@0
|
292 |
clamplow_f64_sse (double *dest, const double *src1, int n, const double *src2_1)
|
sl@0
|
293 |
{
|
sl@0
|
294 |
__m128d xmm1;
|
sl@0
|
295 |
double min = *src2_1;
|
sl@0
|
296 |
|
sl@0
|
297 |
/* Initial operations to align the destination pointer */
|
sl@0
|
298 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
299 |
double x = *src1++;
|
sl@0
|
300 |
if (x < min)
|
sl@0
|
301 |
x = min;
|
sl@0
|
302 |
*dest++ = x;
|
sl@0
|
303 |
}
|
sl@0
|
304 |
xmm1 = _mm_set1_pd(min);
|
sl@0
|
305 |
for (; n >= 2; n -= 2) {
|
sl@0
|
306 |
__m128d xmm0;
|
sl@0
|
307 |
xmm0 = _mm_loadu_pd(src1);
|
sl@0
|
308 |
xmm0 = _mm_max_pd(xmm0, xmm1);
|
sl@0
|
309 |
_mm_store_pd(dest, xmm0);
|
sl@0
|
310 |
dest += 2;
|
sl@0
|
311 |
src1 += 2;
|
sl@0
|
312 |
}
|
sl@0
|
313 |
for (; n > 0; n--) {
|
sl@0
|
314 |
double x = *src1++;
|
sl@0
|
315 |
if (x < min)
|
sl@0
|
316 |
x = min;
|
sl@0
|
317 |
*dest++ = x;
|
sl@0
|
318 |
}
|
sl@0
|
319 |
}
|
sl@0
|
320 |
OIL_DEFINE_IMPL_FULL (clamplow_f64_sse, clamplow_f64,
|
sl@0
|
321 |
OIL_IMPL_FLAG_SSE | OIL_IMPL_FLAG_SSE2);
|
sl@0
|
322 |
|
sl@0
|
323 |
SSE_FUNCTION static void
|
sl@0
|
324 |
clamphigh_u8_sse (uint8_t *dest, const uint8_t *src1, int n,
|
sl@0
|
325 |
const uint8_t *src2_1)
|
sl@0
|
326 |
{
|
sl@0
|
327 |
__m128i xmm1;
|
sl@0
|
328 |
uint8_t max = *src2_1;
|
sl@0
|
329 |
|
sl@0
|
330 |
/* Initial operations to align the destination pointer */
|
sl@0
|
331 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
332 |
uint8_t x = *src1++;
|
sl@0
|
333 |
if (x > max)
|
sl@0
|
334 |
x = max;
|
sl@0
|
335 |
*dest++ = x;
|
sl@0
|
336 |
}
|
sl@0
|
337 |
xmm1 = _mm_set1_epi8(max);
|
sl@0
|
338 |
for (; n >= 16; n -= 16) {
|
sl@0
|
339 |
__m128i xmm0;
|
sl@0
|
340 |
xmm0 = _mm_loadu_si128((__m128i *)src1);
|
sl@0
|
341 |
xmm0 = _mm_min_epu8(xmm0, xmm1);
|
sl@0
|
342 |
_mm_store_si128((__m128i *)dest, xmm0);
|
sl@0
|
343 |
dest += 16;
|
sl@0
|
344 |
src1 += 16;
|
sl@0
|
345 |
}
|
sl@0
|
346 |
for (; n > 0; n--) {
|
sl@0
|
347 |
uint8_t x = *src1++;
|
sl@0
|
348 |
if (x > max)
|
sl@0
|
349 |
x = max;
|
sl@0
|
350 |
*dest++ = x;
|
sl@0
|
351 |
}
|
sl@0
|
352 |
}
|
sl@0
|
353 |
OIL_DEFINE_IMPL_FULL (clamphigh_u8_sse, clamphigh_u8, OIL_IMPL_FLAG_SSE2);
|
sl@0
|
354 |
|
sl@0
|
355 |
SSE_FUNCTION static void
|
sl@0
|
356 |
clamphigh_s16_sse (int16_t *dest, const int16_t *src1, int n,
|
sl@0
|
357 |
const int16_t *src2_1)
|
sl@0
|
358 |
{
|
sl@0
|
359 |
__m128i xmm1;
|
sl@0
|
360 |
int16_t max = *src2_1;
|
sl@0
|
361 |
|
sl@0
|
362 |
/* Initial operations to align the destination pointer */
|
sl@0
|
363 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
364 |
int16_t x = *src1++;
|
sl@0
|
365 |
if (x > max)
|
sl@0
|
366 |
x = max;
|
sl@0
|
367 |
*dest++ = x;
|
sl@0
|
368 |
}
|
sl@0
|
369 |
xmm1 = _mm_set1_epi16(max);
|
sl@0
|
370 |
for (; n >= 8; n -= 8) {
|
sl@0
|
371 |
__m128i xmm0;
|
sl@0
|
372 |
xmm0 = _mm_loadu_si128((__m128i *)src1);
|
sl@0
|
373 |
xmm0 = _mm_min_epi16(xmm0, xmm1);
|
sl@0
|
374 |
_mm_store_si128((__m128i *)dest, xmm0);
|
sl@0
|
375 |
dest += 8;
|
sl@0
|
376 |
src1 += 8;
|
sl@0
|
377 |
}
|
sl@0
|
378 |
for (; n > 0; n--) {
|
sl@0
|
379 |
int16_t x = *src1++;
|
sl@0
|
380 |
if (x > max)
|
sl@0
|
381 |
x = max;
|
sl@0
|
382 |
*dest++ = x;
|
sl@0
|
383 |
}
|
sl@0
|
384 |
}
|
sl@0
|
385 |
OIL_DEFINE_IMPL_FULL (clamphigh_s16_sse, clamphigh_s16, OIL_IMPL_FLAG_SSE2);
|
sl@0
|
386 |
|
sl@0
|
387 |
SSE_FUNCTION static void
|
sl@0
|
388 |
clamphigh_f32_sse (float *dest, const float *src1, int n, const float *src2_1)
|
sl@0
|
389 |
{
|
sl@0
|
390 |
__m128 xmm1;
|
sl@0
|
391 |
float max = *src2_1;
|
sl@0
|
392 |
|
sl@0
|
393 |
/* Initial operations to align the destination pointer */
|
sl@0
|
394 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
395 |
float x = *src1++;
|
sl@0
|
396 |
if (x > max)
|
sl@0
|
397 |
x = max;
|
sl@0
|
398 |
*dest++ = x;
|
sl@0
|
399 |
}
|
sl@0
|
400 |
xmm1 = _mm_set_ps1(max);
|
sl@0
|
401 |
for (; n >= 4; n -= 4) {
|
sl@0
|
402 |
__m128 xmm0;
|
sl@0
|
403 |
xmm0 = _mm_loadu_ps(src1);
|
sl@0
|
404 |
xmm0 = _mm_min_ps(xmm0, xmm1);
|
sl@0
|
405 |
_mm_store_ps(dest, xmm0);
|
sl@0
|
406 |
dest += 4;
|
sl@0
|
407 |
src1 += 4;
|
sl@0
|
408 |
}
|
sl@0
|
409 |
for (; n > 0; n--) {
|
sl@0
|
410 |
float x = *src1++;
|
sl@0
|
411 |
if (x > max)
|
sl@0
|
412 |
x = max;
|
sl@0
|
413 |
*dest++ = x;
|
sl@0
|
414 |
}
|
sl@0
|
415 |
}
|
sl@0
|
416 |
OIL_DEFINE_IMPL_FULL (clamphigh_f32_sse, clamphigh_f32, OIL_IMPL_FLAG_SSE);
|
sl@0
|
417 |
|
sl@0
|
418 |
SSE_FUNCTION static void
|
sl@0
|
419 |
clamphigh_f64_sse (double *dest, const double *src1, int n, const double *src2_1)
|
sl@0
|
420 |
{
|
sl@0
|
421 |
__m128d xmm1;
|
sl@0
|
422 |
double max = *src2_1;
|
sl@0
|
423 |
|
sl@0
|
424 |
/* Initial operations to align the destination pointer */
|
sl@0
|
425 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
426 |
double x = *src1++;
|
sl@0
|
427 |
if (x > max)
|
sl@0
|
428 |
x = max;
|
sl@0
|
429 |
*dest++ = x;
|
sl@0
|
430 |
}
|
sl@0
|
431 |
xmm1 = _mm_set1_pd(max);
|
sl@0
|
432 |
for (; n >= 2; n -= 2) {
|
sl@0
|
433 |
__m128d xmm0;
|
sl@0
|
434 |
xmm0 = _mm_loadu_pd(src1);
|
sl@0
|
435 |
xmm0 = _mm_min_pd(xmm0, xmm1);
|
sl@0
|
436 |
_mm_store_pd(dest, xmm0);
|
sl@0
|
437 |
dest += 2;
|
sl@0
|
438 |
src1 += 2;
|
sl@0
|
439 |
}
|
sl@0
|
440 |
for (; n > 0; n--) {
|
sl@0
|
441 |
double x = *src1++;
|
sl@0
|
442 |
if (x > max)
|
sl@0
|
443 |
x = max;
|
sl@0
|
444 |
*dest++ = x;
|
sl@0
|
445 |
}
|
sl@0
|
446 |
}
|
sl@0
|
447 |
OIL_DEFINE_IMPL_FULL (clamphigh_f64_sse, clamphigh_f64,
|
sl@0
|
448 |
OIL_IMPL_FLAG_SSE | OIL_IMPL_FLAG_SSE2);
|
sl@0
|
449 |
|
sl@0
|
450 |
|
sl@0
|
451 |
#ifdef __SYMBIAN32__
|
sl@0
|
452 |
|
sl@0
|
453 |
OilFunctionImpl* __oil_function_impl_clamp_u8_sse, clamp_u8() {
|
sl@0
|
454 |
return &_oil_function_impl_clamp_u8_sse, clamp_u8;
|
sl@0
|
455 |
}
|
sl@0
|
456 |
#endif
|
sl@0
|
457 |
|
sl@0
|
458 |
#ifdef __SYMBIAN32__
|
sl@0
|
459 |
|
sl@0
|
460 |
OilFunctionImpl* __oil_function_impl_clamp_s16_sse, clamp_s16() {
|
sl@0
|
461 |
return &_oil_function_impl_clamp_s16_sse, clamp_s16;
|
sl@0
|
462 |
}
|
sl@0
|
463 |
#endif
|
sl@0
|
464 |
|
sl@0
|
465 |
#ifdef __SYMBIAN32__
|
sl@0
|
466 |
|
sl@0
|
467 |
OilFunctionImpl* __oil_function_impl_clamp_f32_sse, clamp_f32() {
|
sl@0
|
468 |
return &_oil_function_impl_clamp_f32_sse, clamp_f32;
|
sl@0
|
469 |
}
|
sl@0
|
470 |
#endif
|
sl@0
|
471 |
|
sl@0
|
472 |
#ifdef __SYMBIAN32__
|
sl@0
|
473 |
|
sl@0
|
474 |
OilFunctionImpl* __oil_function_impl_clamp_f64_sse, clamp_f64() {
|
sl@0
|
475 |
return &_oil_function_impl_clamp_f64_sse, clamp_f64;
|
sl@0
|
476 |
}
|
sl@0
|
477 |
#endif
|
sl@0
|
478 |
|
sl@0
|
479 |
#ifdef __SYMBIAN32__
|
sl@0
|
480 |
|
sl@0
|
481 |
OilFunctionImpl* __oil_function_impl_clamplow_u8_sse, clamplow_u8() {
|
sl@0
|
482 |
return &_oil_function_impl_clamplow_u8_sse, clamplow_u8;
|
sl@0
|
483 |
}
|
sl@0
|
484 |
#endif
|
sl@0
|
485 |
|
sl@0
|
486 |
#ifdef __SYMBIAN32__
|
sl@0
|
487 |
|
sl@0
|
488 |
OilFunctionImpl* __oil_function_impl_clamplow_s16_sse, clamplow_s16() {
|
sl@0
|
489 |
return &_oil_function_impl_clamplow_s16_sse, clamplow_s16;
|
sl@0
|
490 |
}
|
sl@0
|
491 |
#endif
|
sl@0
|
492 |
|
sl@0
|
493 |
#ifdef __SYMBIAN32__
|
sl@0
|
494 |
|
sl@0
|
495 |
OilFunctionImpl* __oil_function_impl_clamplow_f32_sse, clamplow_f32() {
|
sl@0
|
496 |
return &_oil_function_impl_clamplow_f32_sse, clamplow_f32;
|
sl@0
|
497 |
}
|
sl@0
|
498 |
#endif
|
sl@0
|
499 |
|
sl@0
|
500 |
#ifdef __SYMBIAN32__
|
sl@0
|
501 |
|
sl@0
|
502 |
OilFunctionImpl* __oil_function_impl_clamplow_f64_sse, clamplow_f64() {
|
sl@0
|
503 |
return &_oil_function_impl_clamplow_f64_sse, clamplow_f64;
|
sl@0
|
504 |
}
|
sl@0
|
505 |
#endif
|
sl@0
|
506 |
|
sl@0
|
507 |
#ifdef __SYMBIAN32__
|
sl@0
|
508 |
|
sl@0
|
509 |
OilFunctionImpl* __oil_function_impl_clamphigh_u8_sse, clamphigh_u8() {
|
sl@0
|
510 |
return &_oil_function_impl_clamphigh_u8_sse, clamphigh_u8;
|
sl@0
|
511 |
}
|
sl@0
|
512 |
#endif
|
sl@0
|
513 |
|
sl@0
|
514 |
#ifdef __SYMBIAN32__
|
sl@0
|
515 |
|
sl@0
|
516 |
OilFunctionImpl* __oil_function_impl_clamphigh_s16_sse, clamphigh_s16() {
|
sl@0
|
517 |
return &_oil_function_impl_clamphigh_s16_sse, clamphigh_s16;
|
sl@0
|
518 |
}
|
sl@0
|
519 |
#endif
|
sl@0
|
520 |
|
sl@0
|
521 |
#ifdef __SYMBIAN32__
|
sl@0
|
522 |
|
sl@0
|
523 |
OilFunctionImpl* __oil_function_impl_clamphigh_f32_sse, clamphigh_f32() {
|
sl@0
|
524 |
return &_oil_function_impl_clamphigh_f32_sse, clamphigh_f32;
|
sl@0
|
525 |
}
|
sl@0
|
526 |
#endif
|
sl@0
|
527 |
|
sl@0
|
528 |
#ifdef __SYMBIAN32__
|
sl@0
|
529 |
|
sl@0
|
530 |
OilFunctionImpl* __oil_function_impl_clamphigh_f64_sse, clamphigh_f64() {
|
sl@0
|
531 |
return &_oil_function_impl_clamphigh_f64_sse, clamphigh_f64;
|
sl@0
|
532 |
}
|
sl@0
|
533 |
#endif
|
sl@0
|
534 |
|