1 /*
2 * Copyright (c) 2000, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation. Oracle designates this
8 * particular file as subject to the "Classpath" exception as provided
9 * by Oracle in the LICENSE file that accompanied this code.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 */
25
26
27 /*
28 * FUNCTION
29 * Internal functions for mlib_ImageConv* on U8/S16/U16 types and
30 * MLIB_EDGE_DST_NO_WRITE mask
31 */
32
33 #include "mlib_image.h"
34 #include "mlib_c_ImageConv.h"
35
36 /*
37 This define switches between functions of different data types
38 */
39 #define IMG_TYPE 2
40
41 /***************************************************************/
42 #if IMG_TYPE == 1
43
44 #define DTYPE mlib_u8
45 #define CONV_FUNC(KERN) mlib_c_conv##KERN##nw_u8
46 #define CONV_FUNC_I(KERN) mlib_i_conv##KERN##nw_u8
47 #define DSCALE (1 << 24)
48 #define FROM_S32(x) (((x) >> 24) ^ 128)
49 #define S64TOS32(x) (x)
50 #define SAT_OFF -(1u << 31)
51
52 #elif IMG_TYPE == 2
53
54 #define DTYPE mlib_s16
55 #define CONV_FUNC(KERN) mlib_conv##KERN##nw_s16
56 #define CONV_FUNC_I(KERN) mlib_i_conv##KERN##nw_s16
57 #define DSCALE 65536.0
58 #define FROM_S32(x) ((x) >> 16)
59 #define S64TOS32(x) ((x) & 0xffffffff)
60 #define SAT_OFF
61
62 #elif IMG_TYPE == 3
63
64 #define DTYPE mlib_u16
65 #define CONV_FUNC(KERN) mlib_conv##KERN##nw_u16
66 #define CONV_FUNC_I(KERN) mlib_i_conv##KERN##nw_u16
67 #define DSCALE 65536.0
68 #define FROM_S32(x) (((x) >> 16) ^ 0x8000)
69 #define S64TOS32(x) (x)
70 #define SAT_OFF -(1u << 31)
71
72 #endif /* IMG_TYPE == 1 */
73
74 /***************************************************************/
75 #define BUFF_SIZE 1600
76
77 #define CACHE_SIZE (64*1024)
78
79 /***************************************************************/
80 #define FTYPE mlib_d64
81
82 #ifndef MLIB_USE_FTOI_CLAMPING
83
84 #define CLAMP_S32(x) \
85 (((x) <= MLIB_S32_MIN) ? MLIB_S32_MIN : (((x) >= MLIB_S32_MAX) ? MLIB_S32_MAX : (mlib_s32)(x)))
86
87 #else
88
89 #define CLAMP_S32(x) ((mlib_s32)(x))
90
91 #endif /* MLIB_USE_FTOI_CLAMPING */
92
93 /***************************************************************/
94 #define D2I(x) CLAMP_S32((x) SAT_OFF)
95
96 /***************************************************************/
97 #ifdef _LITTLE_ENDIAN
98
99 #define STORE2(res0, res1) \
100 dp[0 ] = res1; \
101 dp[chan1] = res0
102
103 #else
104
105 #define STORE2(res0, res1) \
106 dp[0 ] = res0; \
107 dp[chan1] = res1
108
109 #endif /* _LITTLE_ENDIAN */
110
111 /***************************************************************/
112 #ifdef _NO_LONGLONG
113
114 #define LOAD_BUFF(buff) \
115 buff[i ] = sp[0]; \
116 buff[i + 1] = sp[chan1]
117
118 #else /* _NO_LONGLONG */
119
120 #ifdef _LITTLE_ENDIAN
121
122 #define LOAD_BUFF(buff) \
123 *(mlib_s64*)(buff + i) = (((mlib_s64)sp[chan1]) << 32) | S64TOS32((mlib_s64)sp[0])
124
125 #else /* _LITTLE_ENDIAN */
126
127 #define LOAD_BUFF(buff) \
128 *(mlib_s64*)(buff + i) = (((mlib_s64)sp[0]) << 32) | S64TOS32((mlib_s64)sp[chan1])
129
130 #endif /* _LITTLE_ENDIAN */
131 #endif /* _NO_LONGLONG */
132
133 /***************************************************************/
134 typedef union {
135 mlib_d64 d64;
136 struct {
137 mlib_s32 i0;
138 mlib_s32 i1;
139 } i32s;
140 struct {
141 mlib_s32 f0;
142 mlib_s32 f1;
143 } f32s;
144 } d64_2x32;
145
146 /***************************************************************/
147 #define DEF_VARS(type) \
148 type *adr_src, *sl, *sp = NULL; \
149 type *adr_dst, *dl, *dp = NULL; \
150 FTYPE *pbuff = buff; \
151 mlib_s32 wid, hgt, sll, dll; \
152 mlib_s32 nchannel, chan1; \
153 mlib_s32 i, j, c
154
155 /***************************************************************/
156 #define GET_SRC_DST_PARAMETERS(type) \
157 hgt = mlib_ImageGetHeight(src); \
158 wid = mlib_ImageGetWidth(src); \
159 nchannel = mlib_ImageGetChannels(src); \
160 sll = mlib_ImageGetStride(src) / sizeof(type); \
161 dll = mlib_ImageGetStride(dst) / sizeof(type); \
162 adr_src = (type *)mlib_ImageGetData(src); \
163 adr_dst = (type *)mlib_ImageGetData(dst)
164
165 /***************************************************************/
166 #if IMG_TYPE == 1
167
168 /* Test for the presence of any "1" bit in bits
169 8 to 31 of val. If present, then val is either
170 negative or >255. If over/underflows of 8 bits
171 are uncommon, then this technique can be a win,
172 since only a single test, rather than two, is
173 necessary to determine if clamping is needed.
174 On the other hand, if over/underflows are common,
175 it adds an extra test.
176 */
177 #define CLAMP_STORE(dst, val) \
178 if (val & 0xffffff00) { \
179 if (val < MLIB_U8_MIN) \
180 dst = MLIB_U8_MIN; \
181 else \
182 dst = MLIB_U8_MAX; \
183 } else { \
184 dst = (mlib_u8)val; \
185 }
186
187 #elif IMG_TYPE == 2
188
189 #define CLAMP_STORE(dst, val) \
190 if (val >= MLIB_S16_MAX) \
191 dst = MLIB_S16_MAX; \
192 else if (val <= MLIB_S16_MIN) \
193 dst = MLIB_S16_MIN; \
194 else \
195 dst = (mlib_s16)val
196
197 #elif IMG_TYPE == 3
198
199 #define CLAMP_STORE(dst, val) \
200 if (val >= MLIB_U16_MAX) \
201 dst = MLIB_U16_MAX; \
202 else if (val <= MLIB_U16_MIN) \
203 dst = MLIB_U16_MIN; \
204 else \
205 dst = (mlib_u16)val
206
207 #endif /* IMG_TYPE == 1 */
208
209 /***************************************************************/
210 #define MAX_KER 7
211 #define MAX_N 15
212
213 static mlib_status mlib_ImageConv1xN(mlib_image *dst,
214 const mlib_image *src,
215 const mlib_d64 *k,
216 mlib_s32 n,
217 mlib_s32 dn,
218 mlib_s32 cmask)
219 {
220 FTYPE buff[BUFF_SIZE];
221 mlib_s32 off, kh;
222 mlib_s32 d0, d1;
223 const FTYPE *pk;
224 FTYPE k0, k1, k2, k3;
225 FTYPE p0, p1, p2, p3, p4;
226 DEF_VARS(DTYPE);
227 DTYPE *sl_c, *dl_c, *sl0;
228 mlib_s32 l, hsize, max_hsize;
229 GET_SRC_DST_PARAMETERS(DTYPE);
230
231 hgt -= (n - 1);
232 adr_dst += dn*dll;
233
234 max_hsize = (CACHE_SIZE/sizeof(DTYPE))/sll;
235
236 if (!max_hsize) max_hsize = 1;
237
238 if (max_hsize > BUFF_SIZE) {
239 pbuff = mlib_malloc(sizeof(FTYPE)*max_hsize);
240 }
241
242 chan1 = nchannel;
243
244 sl_c = adr_src;
245 dl_c = adr_dst;
246
247 for (l = 0; l < hgt; l += hsize) {
248 hsize = hgt - l;
249
250 if (hsize > max_hsize) hsize = max_hsize;
251
252 for (c = 0; c < nchannel; c++) {
253 if (!(cmask & (1 << (chan1 - 1 - c)))) continue;
254
255 sl = sl_c + c;
256 dl = dl_c + c;
257
258 for (j = 0; j < hsize; j++) pbuff[j] = 0.0;
259
260 for (i = 0; i < wid; i++) {
261 sl0 = sl;
262
263 for (off = 0; off < (n - 4); off += 4) {
264 pk = k + off;
265 sp = sl0;
266
267 k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
268 p2 = sp[0]; p3 = sp[sll]; p4 = sp[2*sll];
269 sp += 3*sll;
270
271 for (j = 0; j < hsize; j += 2) {
272 p0 = p2; p1 = p3; p2 = p4;
273 p3 = sp[0];
274 p4 = sp[sll];
275
276 pbuff[j ] += p0*k0 + p1*k1 + p2*k2 + p3*k3;
277 pbuff[j + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3;
278
279 sp += 2*sll;
280 }
281
282 sl0 += 4*sll;
283 }
284
285 pk = k + off;
286 sp = sl0;
287
288 k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
289 p2 = sp[0]; p3 = sp[sll]; p4 = sp[2*sll];
290
291 dp = dl;
292 kh = n - off;
293
294 if (kh == 4) {
295 sp += 3*sll;
296
297 for (j = 0; j <= (hsize - 2); j += 2) {
298 p0 = p2; p1 = p3; p2 = p4;
299 p3 = sp[0];
300 p4 = sp[sll];
301
302 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + pbuff[j]);
303 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + pbuff[j + 1]);
304
305 dp[0 ] = FROM_S32(d0);
306 dp[dll] = FROM_S32(d1);
307
308 pbuff[j] = 0;
309 pbuff[j + 1] = 0;
310
311 sp += 2*sll;
312 dp += 2*dll;
313 }
314
315 if (j < hsize) {
316 p0 = p2; p1 = p3; p2 = p4;
317 p3 = sp[0];
318
319 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + pbuff[j]);
320
321 pbuff[j] = 0;
322
323 dp[0] = FROM_S32(d0);
324 }
325
326 } else if (kh == 3) {
327 sp += 2*sll;
328
329 for (j = 0; j <= (hsize - 2); j += 2) {
330 p0 = p2; p1 = p3;
331 p2 = sp[0];
332 p3 = sp[sll];
333
334 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + pbuff[j]);
335 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + pbuff[j + 1]);
336
337 dp[0 ] = FROM_S32(d0);
338 dp[dll] = FROM_S32(d1);
339
340 pbuff[j] = 0;
341 pbuff[j + 1] = 0;
342
343 sp += 2*sll;
344 dp += 2*dll;
345 }
346
347 if (j < hsize) {
348 p0 = p2; p1 = p3;
349 p2 = sp[0];
350
351 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + pbuff[j]);
352
353 pbuff[j] = 0;
354
355 dp[0] = FROM_S32(d0);
356 }
357
358 } else if (kh == 2) {
359 sp += sll;
360
361 for (j = 0; j <= (hsize - 2); j += 2) {
362 p0 = p2;
363 p1 = sp[0];
364 p2 = sp[sll];
365
366 d0 = D2I(p0*k0 + p1*k1 + pbuff[j]);
367 d1 = D2I(p1*k0 + p2*k1 + pbuff[j + 1]);
368
369 dp[0 ] = FROM_S32(d0);
370 dp[dll] = FROM_S32(d1);
371
372 pbuff[j] = 0;
373 pbuff[j + 1] = 0;
374
375 sp += 2*sll;
376 dp += 2*dll;
377 }
378
379 if (j < hsize) {
380 p0 = p2;
381 p1 = sp[0];
382
383 d0 = D2I(p0*k0 + p1*k1 + pbuff[j]);
384
385 pbuff[j] = 0;
386
387 dp[0] = FROM_S32(d0);
388 }
389
390 } else /* if (kh == 1) */ {
391 for (j = 0; j < hsize; j++) {
392 p0 = sp[0];
393
394 d0 = D2I(p0*k0 + pbuff[j]);
395
396 dp[0] = FROM_S32(d0);
397
398 pbuff[j] = 0;
399
400 sp += sll;
401 dp += dll;
402 }
403 }
404
405 sl += chan1;
406 dl += chan1;
407 }
408 }
409
410 sl_c += max_hsize*sll;
411 dl_c += max_hsize*dll;
412 }
413
414 if (pbuff != buff) mlib_free(pbuff);
415
416 return MLIB_SUCCESS;
417 }
418
419 /***************************************************************/
420 mlib_status CONV_FUNC(MxN)(mlib_image *dst,
421 const mlib_image *src,
422 const mlib_s32 *kernel,
423 mlib_s32 m,
424 mlib_s32 n,
425 mlib_s32 dm,
426 mlib_s32 dn,
427 mlib_s32 scale,
428 mlib_s32 cmask)
429 {
430 FTYPE buff[BUFF_SIZE], *buffs_arr[2*(MAX_N + 1)];
431 FTYPE **buffs = buffs_arr, *buffd;
432 FTYPE akernel[256], *k = akernel, fscale = DSCALE;
433 mlib_s32 mn, l, off, kw, bsize, buff_ind;
434 mlib_s32 d0, d1;
435 FTYPE k0, k1, k2, k3, k4, k5, k6;
436 FTYPE p0, p1, p2, p3, p4, p5, p6, p7;
437 d64_2x32 dd;
438 DEF_VARS(DTYPE);
439 mlib_s32 chan2;
440 mlib_s32 *buffo, *buffi;
441 mlib_status status = MLIB_SUCCESS;
442
443 GET_SRC_DST_PARAMETERS(DTYPE);
444
445 if (scale > 30) {
446 fscale *= 1.0/(1 << 30);
447 scale -= 30;
448 }
449
450 fscale /= (1 << scale);
451
452 mn = m*n;
453
454 if (mn > 256) {
455 k = mlib_malloc(mn*sizeof(mlib_d64));
456
457 if (k == NULL) return MLIB_FAILURE;
458 }
459
460 for (i = 0; i < mn; i++) {
461 k[i] = kernel[i]*fscale;
462 }
463
464 if (m == 1) {
465 status = mlib_ImageConv1xN(dst, src, k, n, dn, cmask);
466 FREE_AND_RETURN_STATUS;
467 }
468
469 bsize = (n + 3)*wid;
470
471 if ((bsize > BUFF_SIZE) || (n > MAX_N)) {
472 pbuff = mlib_malloc(sizeof(FTYPE)*bsize + sizeof(FTYPE *)*2*(n + 1));
473
474 if (pbuff == NULL) {
475 status = MLIB_FAILURE;
476 FREE_AND_RETURN_STATUS;
477 }
478 buffs = (FTYPE **)(pbuff + bsize);
479 }
480
481 for (l = 0; l < (n + 1); l++) buffs[l] = pbuff + l*wid;
482 for (l = 0; l < (n + 1); l++) buffs[l + (n + 1)] = buffs[l];
483 buffd = buffs[n] + wid;
484 buffo = (mlib_s32*)(buffd + wid);
485 buffi = buffo + (wid &~ 1);
486
487 chan1 = nchannel;
488 chan2 = chan1 + chan1;
489
490 wid -= (m - 1);
491 hgt -= (n - 1);
492 adr_dst += dn*dll + dm*nchannel;
493
494 for (c = 0; c < nchannel; c++) {
495 if (!(cmask & (1 << (chan1 - 1 - c)))) continue;
496
497 sl = adr_src + c;
498 dl = adr_dst + c;
499
500 for (l = 0; l < n; l++) {
501 FTYPE *buff = buffs[l];
502
503 for (i = 0; i < wid + (m - 1); i++) {
504 buff[i] = (FTYPE)sl[i*chan1];
505 }
506
507 sl += sll;
508 }
509
510 buff_ind = 0;
511
512 for (i = 0; i < wid; i++) buffd[i] = 0.0;
513
514 for (j = 0; j < hgt; j++) {
515 FTYPE **buffc = buffs + buff_ind;
516 FTYPE *buffn = buffc[n];
517 FTYPE *pk = k;
518
519 for (l = 0; l < n; l++) {
520 FTYPE *buff_l = buffc[l];
521
522 for (off = 0; off < m;) {
523 FTYPE *buff = buff_l + off;
524
525 kw = m - off;
526
527 if (kw > 2*MAX_KER) kw = MAX_KER; else
528 if (kw > MAX_KER) kw = kw/2;
529 off += kw;
530
531 sp = sl;
532 dp = dl;
533
534 p2 = buff[0]; p3 = buff[1]; p4 = buff[2];
535 p5 = buff[3]; p6 = buff[4]; p7 = buff[5];
536
537 k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
538 k4 = pk[4]; k5 = pk[5]; k6 = pk[6];
539 pk += kw;
540
541 if (kw == 7) {
542
543 if (l < (n - 1) || off < m) {
544 for (i = 0; i <= (wid - 2); i += 2) {
545 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
546
547 p6 = buff[i + 6]; p7 = buff[i + 7];
548
549 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6;
550 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6;
551 }
552
553 } else {
554 for (i = 0; i <= (wid - 2); i += 2) {
555 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
556
557 p6 = buff[i + 6]; p7 = buff[i + 7];
558
559 LOAD_BUFF(buffi);
560
561 dd.d64 = *(FTYPE *)(buffi + i);
562 buffn[i ] = (FTYPE)dd.i32s.i0;
563 buffn[i + 1] = (FTYPE)dd.i32s.i1;
564
565 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6 + buffd[i ]);
566 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6 + buffd[i + 1]);
567
568 dp[0 ] = FROM_S32(d0);
569 dp[chan1] = FROM_S32(d1);
570
571 buffd[i ] = 0.0;
572 buffd[i + 1] = 0.0;
573
574 sp += chan2;
575 dp += chan2;
576 }
577 }
578
579 } else if (kw == 6) {
580
581 if (l < (n - 1) || off < m) {
582 for (i = 0; i <= (wid - 2); i += 2) {
583 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
584
585 p5 = buff[i + 5]; p6 = buff[i + 6];
586
587 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5;
588 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5;
589 }
590
591 } else {
592 for (i = 0; i <= (wid - 2); i += 2) {
593 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
594
595 p5 = buff[i + 5]; p6 = buff[i + 6];
596
597 buffn[i ] = (FTYPE)sp[0];
598 buffn[i + 1] = (FTYPE)sp[chan1];
599
600 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + buffd[i ]);
601 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + buffd[i + 1]);
602
603 dp[0 ] = FROM_S32(d0);
604 dp[chan1] = FROM_S32(d1);
605
606 buffd[i ] = 0.0;
607 buffd[i + 1] = 0.0;
608
609 sp += chan2;
610 dp += chan2;
611 }
612 }
613
614 } else if (kw == 5) {
615
616 if (l < (n - 1) || off < m) {
617 for (i = 0; i <= (wid - 2); i += 2) {
618 p0 = p2; p1 = p3; p2 = p4; p3 = p5;
619
620 p4 = buff[i + 4]; p5 = buff[i + 5];
621
622 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4;
623 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4;
624 }
625
626 } else {
627 for (i = 0; i <= (wid - 2); i += 2) {
628 p0 = p2; p1 = p3; p2 = p4; p3 = p5;
629
630 p4 = buff[i + 4]; p5 = buff[i + 5];
631
632 buffn[i ] = (FTYPE)sp[0];
633 buffn[i + 1] = (FTYPE)sp[chan1];
634
635 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + buffd[i ]);
636 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + buffd[i + 1]);
637
638 dp[0 ] = FROM_S32(d0);
639 dp[chan1] = FROM_S32(d1);
640
641 buffd[i ] = 0.0;
642 buffd[i + 1] = 0.0;
643
644 sp += chan2;
645 dp += chan2;
646 }
647 }
648
649 } else if (kw == 4) {
650
651 if (l < (n - 1) || off < m) {
652 for (i = 0; i <= (wid - 2); i += 2) {
653 p0 = p2; p1 = p3; p2 = p4;
654
655 p3 = buff[i + 3]; p4 = buff[i + 4];
656
657 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3;
658 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3;
659 }
660
661 } else {
662 for (i = 0; i <= (wid - 2); i += 2) {
663 p0 = p2; p1 = p3; p2 = p4;
664
665 p3 = buff[i + 3]; p4 = buff[i + 4];
666
667 buffn[i ] = (FTYPE)sp[0];
668 buffn[i + 1] = (FTYPE)sp[chan1];
669
670 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + buffd[i ]);
671 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + buffd[i + 1]);
672
673 dp[0 ] = FROM_S32(d0);
674 dp[chan1] = FROM_S32(d1);
675
676 buffd[i ] = 0.0;
677 buffd[i + 1] = 0.0;
678
679 sp += chan2;
680 dp += chan2;
681 }
682 }
683
684 } else if (kw == 3) {
685
686 if (l < (n - 1) || off < m) {
687 for (i = 0; i <= (wid - 2); i += 2) {
688 p0 = p2; p1 = p3;
689
690 p2 = buff[i + 2]; p3 = buff[i + 3];
691
692 buffd[i ] += p0*k0 + p1*k1 + p2*k2;
693 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2;
694 }
695
696 } else {
697 for (i = 0; i <= (wid - 2); i += 2) {
698 p0 = p2; p1 = p3;
699
700 p2 = buff[i + 2]; p3 = buff[i + 3];
701
702 buffn[i ] = (FTYPE)sp[0];
703 buffn[i + 1] = (FTYPE)sp[chan1];
704
705 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + buffd[i ]);
706 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + buffd[i + 1]);
707
708 dp[0 ] = FROM_S32(d0);
709 dp[chan1] = FROM_S32(d1);
710
711 buffd[i ] = 0.0;
712 buffd[i + 1] = 0.0;
713
714 sp += chan2;
715 dp += chan2;
716 }
717 }
718
719 } else /*if (kw == 2)*/ {
720
721 if (l < (n - 1) || off < m) {
722 for (i = 0; i <= (wid - 2); i += 2) {
723 p0 = p2;
724
725 p1 = buff[i + 1]; p2 = buff[i + 2];
726
727 buffd[i ] += p0*k0 + p1*k1;
728 buffd[i + 1] += p1*k0 + p2*k1;
729 }
730
731 } else {
732 for (i = 0; i <= (wid - 2); i += 2) {
733 p0 = p2;
734
735 p1 = buff[i + 1]; p2 = buff[i + 2];
736
737 buffn[i ] = (FTYPE)sp[0];
738 buffn[i + 1] = (FTYPE)sp[chan1];
739
740 d0 = D2I(p0*k0 + p1*k1 + buffd[i ]);
741 d1 = D2I(p1*k0 + p2*k1 + buffd[i + 1]);
742
743 dp[0 ] = FROM_S32(d0);
744 dp[chan1] = FROM_S32(d1);
745
746 buffd[i ] = 0.0;
747 buffd[i + 1] = 0.0;
748
749 sp += chan2;
750 dp += chan2;
751 }
752 }
753 }
754 }
755 }
756
757 /* last pixels */
758 for (; i < wid; i++) {
759 FTYPE *pk = k, s = 0;
760 mlib_s32 x, d0;
761
762 for (l = 0; l < n; l++) {
763 FTYPE *buff = buffc[l] + i;
764
765 for (x = 0; x < m; x++) s += buff[x] * (*pk++);
766 }
767
768 d0 = D2I(s);
769 dp[0] = FROM_S32(d0);
770
771 buffn[i] = (FTYPE)sp[0];
772
773 sp += chan1;
774 dp += chan1;
775 }
776
777 for (l = 0; l < (m - 1); l++) buffn[wid + l] = sp[l*chan1];
778
779 /* next line */
780 sl += sll;
781 dl += dll;
782
783 buff_ind++;
784
785 if (buff_ind >= n + 1) buff_ind = 0;
786 }
787 }
788
789 FREE_AND_RETURN_STATUS;
790 }
791
792 /***************************************************************/
793 /* for x86, using integer multiplies is faster */
794
795 #define STORE_RES(res, x) \
796 x >>= shift2; \
797 CLAMP_STORE(res, x)
798
799 mlib_status CONV_FUNC_I(MxN)(mlib_image *dst,
800 const mlib_image *src,
801 const mlib_s32 *kernel,
802 mlib_s32 m,
803 mlib_s32 n,
804 mlib_s32 dm,
805 mlib_s32 dn,
806 mlib_s32 scale,
807 mlib_s32 cmask)
808 {
809 mlib_s32 buff[BUFF_SIZE], *buffd = buff;
810 mlib_s32 l, off, kw;
811 mlib_s32 d0, d1, shift1, shift2;
812 mlib_s32 k0, k1, k2, k3, k4, k5, k6;
813 mlib_s32 p0, p1, p2, p3, p4, p5, p6, p7;
814 DTYPE *adr_src, *sl, *sp = NULL;
815 DTYPE *adr_dst, *dl, *dp = NULL;
816 mlib_s32 wid, hgt, sll, dll;
817 mlib_s32 nchannel, chan1;
818 mlib_s32 i, j, c;
819 mlib_s32 chan2;
820 mlib_s32 k_locl[MAX_N*MAX_N], *k = k_locl;
821 GET_SRC_DST_PARAMETERS(DTYPE);
822
823 #if IMG_TYPE != 1
824 shift1 = 16;
825 #else
826 shift1 = 8;
827 #endif /* IMG_TYPE != 1 */
828 shift2 = scale - shift1;
829
830 chan1 = nchannel;
831 chan2 = chan1 + chan1;
832
833 wid -= (m - 1);
834 hgt -= (n - 1);
835 adr_dst += dn*dll + dm*nchannel;
836
837 if (wid > BUFF_SIZE) {
838 buffd = mlib_malloc(sizeof(mlib_s32)*wid);
839
840 if (buffd == NULL) return MLIB_FAILURE;
841 }
842
843 if (m*n > MAX_N*MAX_N) {
844 k = mlib_malloc(sizeof(mlib_s32)*(m*n));
845
846 if (k == NULL) {
847 if (buffd != buff) mlib_free(buffd);
848 return MLIB_FAILURE;
849 }
850 }
851
852 for (i = 0; i < m*n; i++) {
853 k[i] = kernel[i] >> shift1;
854 }
855
856 for (c = 0; c < nchannel; c++) {
857 if (!(cmask & (1 << (nchannel - 1 - c)))) continue;
858
859 sl = adr_src + c;
860 dl = adr_dst + c;
861
862 for (i = 0; i < wid; i++) buffd[i] = 0;
863
864 for (j = 0; j < hgt; j++) {
865 mlib_s32 *pk = k;
866
867 for (l = 0; l < n; l++) {
868 DTYPE *sp0 = sl + l*sll;
869
870 for (off = 0; off < m;) {
871 sp = sp0 + off*chan1;
872 dp = dl;
873
874 kw = m - off;
875
876 if (kw > 2*MAX_KER) kw = MAX_KER; else
877 if (kw > MAX_KER) kw = kw/2;
878 off += kw;
879
880 p2 = sp[0]; p3 = sp[chan1]; p4 = sp[chan2];
881 p5 = sp[chan2 + chan1]; p6 = sp[chan2 + chan2]; p7 = sp[5*chan1];
882
883 k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
884 k4 = pk[4]; k5 = pk[5]; k6 = pk[6];
885 pk += kw;
886
887 sp += (kw - 1)*chan1;
888
889 if (kw == 7) {
890
891 if (l < (n - 1) || off < m) {
892 for (i = 0; i <= (wid - 2); i += 2) {
893 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
894 p6 = sp[0];
895 p7 = sp[chan1];
896
897 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6;
898 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6;
899
900 sp += chan2;
901 }
902
903 } else {
904 for (i = 0; i <= (wid - 2); i += 2) {
905 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
906 p6 = sp[0];
907 p7 = sp[chan1];
908
909 d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6 + buffd[i ]);
910 d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6 + buffd[i + 1]);
911
912 STORE_RES(dp[0 ], d0);
913 STORE_RES(dp[chan1], d1);
914
915 buffd[i ] = 0;
916 buffd[i + 1] = 0;
917
918 sp += chan2;
919 dp += chan2;
920 }
921 }
922
923 } else if (kw == 6) {
924
925 if (l < (n - 1) || off < m) {
926 for (i = 0; i <= (wid - 2); i += 2) {
927 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
928 p5 = sp[0];
929 p6 = sp[chan1];
930
931 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5;
932 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5;
933
934 sp += chan2;
935 }
936
937 } else {
938 for (i = 0; i <= (wid - 2); i += 2) {
939 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
940 p5 = sp[0];
941 p6 = sp[chan1];
942
943 d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + buffd[i ]);
944 d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + buffd[i + 1]);
945
946 STORE_RES(dp[0 ], d0);
947 STORE_RES(dp[chan1], d1);
948
949 buffd[i ] = 0;
950 buffd[i + 1] = 0;
951
952 sp += chan2;
953 dp += chan2;
954 }
955 }
956
957 } else if (kw == 5) {
958
959 if (l < (n - 1) || off < m) {
960 for (i = 0; i <= (wid - 2); i += 2) {
961 p0 = p2; p1 = p3; p2 = p4; p3 = p5;
962 p4 = sp[0];
963 p5 = sp[chan1];
964
965 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4;
966 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4;
967
968 sp += chan2;
969 }
970
971 } else {
972 for (i = 0; i <= (wid - 2); i += 2) {
973 p0 = p2; p1 = p3; p2 = p4; p3 = p5;
974 p4 = sp[0];
975 p5 = sp[chan1];
976
977 d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + buffd[i ]);
978 d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + buffd[i + 1]);
979
980 STORE_RES(dp[0 ], d0);
981 STORE_RES(dp[chan1], d1);
982
983 buffd[i ] = 0;
984 buffd[i + 1] = 0;
985
986 sp += chan2;
987 dp += chan2;
988 }
989 }
990
991 } else if (kw == 4) {
992
993 if (l < (n - 1) || off < m) {
994 for (i = 0; i <= (wid - 2); i += 2) {
995 p0 = p2; p1 = p3; p2 = p4;
996 p3 = sp[0];
997 p4 = sp[chan1];
998
999 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3;
1000 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3;
1001
1002 sp += chan2;
1003 }
1004
1005 } else {
1006 for (i = 0; i <= (wid - 2); i += 2) {
1007 p0 = p2; p1 = p3; p2 = p4;
1008 p3 = sp[0];
1009 p4 = sp[chan1];
1010
1011 d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + buffd[i ]);
1012 d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + buffd[i + 1]);
1013
1014 STORE_RES(dp[0 ], d0);
1015 STORE_RES(dp[chan1], d1);
1016
1017 buffd[i ] = 0;
1018 buffd[i + 1] = 0;
1019
1020 sp += chan2;
1021 dp += chan2;
1022 }
1023 }
1024
1025 } else if (kw == 3) {
1026
1027 if (l < (n - 1) || off < m) {
1028 for (i = 0; i <= (wid - 2); i += 2) {
1029 p0 = p2; p1 = p3;
1030 p2 = sp[0];
1031 p3 = sp[chan1];
1032
1033 buffd[i ] += p0*k0 + p1*k1 + p2*k2;
1034 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2;
1035
1036 sp += chan2;
1037 }
1038
1039 } else {
1040 for (i = 0; i <= (wid - 2); i += 2) {
1041 p0 = p2; p1 = p3;
1042 p2 = sp[0];
1043 p3 = sp[chan1];
1044
1045 d0 = (p0*k0 + p1*k1 + p2*k2 + buffd[i ]);
1046 d1 = (p1*k0 + p2*k1 + p3*k2 + buffd[i + 1]);
1047
1048 STORE_RES(dp[0 ], d0);
1049 STORE_RES(dp[chan1], d1);
1050
1051 buffd[i ] = 0;
1052 buffd[i + 1] = 0;
1053
1054 sp += chan2;
1055 dp += chan2;
1056 }
1057 }
1058
1059 } else if (kw == 2) {
1060
1061 if (l < (n - 1) || off < m) {
1062 for (i = 0; i <= (wid - 2); i += 2) {
1063 p0 = p2;
1064 p1 = sp[0];
1065 p2 = sp[chan1];
1066
1067 buffd[i ] += p0*k0 + p1*k1;
1068 buffd[i + 1] += p1*k0 + p2*k1;
1069
1070 sp += chan2;
1071 }
1072
1073 } else {
1074 for (i = 0; i <= (wid - 2); i += 2) {
1075 p0 = p2;
1076 p1 = sp[0];
1077 p2 = sp[chan1];
1078
1079 d0 = (p0*k0 + p1*k1 + buffd[i ]);
1080 d1 = (p1*k0 + p2*k1 + buffd[i + 1]);
1081
1082 STORE_RES(dp[0 ], d0);
1083 STORE_RES(dp[chan1], d1);
1084
1085 buffd[i ] = 0;
1086 buffd[i + 1] = 0;
1087
1088 sp += chan2;
1089 dp += chan2;
1090 }
1091 }
1092
1093 } else /*if (kw == 1)*/ {
1094
1095 if (l < (n - 1) || off < m) {
1096 for (i = 0; i <= (wid - 2); i += 2) {
1097 p0 = sp[0];
1098 p1 = sp[chan1];
1099
1100 buffd[i ] += p0*k0;
1101 buffd[i + 1] += p1*k0;
1102
1103 sp += chan2;
1104 }
1105
1106 } else {
1107 for (i = 0; i <= (wid - 2); i += 2) {
1108 p0 = sp[0];
1109 p1 = sp[chan1];
1110
1111 d0 = (p0*k0 + buffd[i ]);
1112 d1 = (p1*k0 + buffd[i + 1]);
1113
1114 STORE_RES(dp[0 ], d0);
1115 STORE_RES(dp[chan1], d1);
1116
1117 buffd[i ] = 0;
1118 buffd[i + 1] = 0;
1119
1120 sp += chan2;
1121 dp += chan2;
1122 }
1123 }
1124 }
1125 }
1126 }
1127
1128 /* last pixels */
1129 for (; i < wid; i++) {
1130 mlib_s32 *pk = k, s = 0;
1131 mlib_s32 x;
1132
1133 for (l = 0; l < n; l++) {
1134 sp = sl + l*sll + i*chan1;
1135
1136 for (x = 0; x < m; x++) {
1137 s += sp[0] * pk[0];
1138 sp += chan1;
1139 pk ++;
1140 }
1141 }
1142
1143 STORE_RES(dp[0], s);
1144
1145 sp += chan1;
1146 dp += chan1;
1147 }
1148
1149 sl += sll;
1150 dl += dll;
1151 }
1152 }
1153
1154 if (buffd != buff) mlib_free(buffd);
1155 if (k != k_locl) mlib_free(k);
1156
1157 return MLIB_SUCCESS;
1158 }
1159
1160 /***************************************************************/