1/*
2 * Copyright (c) 2003, 2013, 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 3
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/***************************************************************/
134typedef 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#ifndef __sparc
167
168#if IMG_TYPE == 1
169
170/* Test for the presence of any "1" bit in bits
171 8 to 31 of val. If present, then val is either
172 negative or >255. If over/underflows of 8 bits
173 are uncommon, then this technique can be a win,
174 since only a single test, rather than two, is
175 necessary to determine if clamping is needed.
176 On the other hand, if over/underflows are common,
177 it adds an extra test.
178*/
179#define CLAMP_STORE(dst, val) \
180 if (val & 0xffffff00) { \
181 if (val < MLIB_U8_MIN) \
182 dst = MLIB_U8_MIN; \
183 else \
184 dst = MLIB_U8_MAX; \
185 } else { \
186 dst = (mlib_u8)val; \
187 }
188
189#elif IMG_TYPE == 2
190
191#define CLAMP_STORE(dst, val) \
192 if (val >= MLIB_S16_MAX) \
193 dst = MLIB_S16_MAX; \
194 else if (val <= MLIB_S16_MIN) \
195 dst = MLIB_S16_MIN; \
196 else \
197 dst = (mlib_s16)val
198
199#elif IMG_TYPE == 3
200
201#define CLAMP_STORE(dst, val) \
202 if (val >= MLIB_U16_MAX) \
203 dst = MLIB_U16_MAX; \
204 else if (val <= MLIB_U16_MIN) \
205 dst = MLIB_U16_MIN; \
206 else \
207 dst = (mlib_u16)val
208
209#endif /* IMG_TYPE == 1 */
210#endif /* __sparc */
211
212/***************************************************************/
213#define MAX_KER 7
214#define MAX_N 15
215
216static mlib_status mlib_ImageConv1xN(mlib_image *dst,
217 const mlib_image *src,
218 const mlib_d64 *k,
219 mlib_s32 n,
220 mlib_s32 dn,
221 mlib_s32 cmask)
222{
223 FTYPE buff[BUFF_SIZE];
224 mlib_s32 off, kh;
225 mlib_s32 d0, d1;
226 const FTYPE *pk;
227 FTYPE k0, k1, k2, k3;
228 FTYPE p0, p1, p2, p3, p4;
229 DEF_VARS(DTYPE);
230 DTYPE *sl_c, *dl_c, *sl0;
231 mlib_s32 l, hsize, max_hsize;
232 GET_SRC_DST_PARAMETERS(DTYPE);
233
234 hgt -= (n - 1);
235 adr_dst += dn*dll;
236
237 max_hsize = (CACHE_SIZE/sizeof(DTYPE))/sll;
238
239 if (!max_hsize) max_hsize = 1;
240
241 if (max_hsize > BUFF_SIZE) {
242 pbuff = mlib_malloc(sizeof(FTYPE)*max_hsize);
243 }
244
245 chan1 = nchannel;
246
247 sl_c = adr_src;
248 dl_c = adr_dst;
249
250 for (l = 0; l < hgt; l += hsize) {
251 hsize = hgt - l;
252
253 if (hsize > max_hsize) hsize = max_hsize;
254
255 for (c = 0; c < nchannel; c++) {
256 if (!(cmask & (1 << (chan1 - 1 - c)))) continue;
257
258 sl = sl_c + c;
259 dl = dl_c + c;
260
261#ifdef __SUNPRO_C
262#pragma pipeloop(0)
263#endif /* __SUNPRO_C */
264 for (j = 0; j < hsize; j++) pbuff[j] = 0.0;
265
266 for (i = 0; i < wid; i++) {
267 sl0 = sl;
268
269 for (off = 0; off < (n - 4); off += 4) {
270 pk = k + off;
271 sp = sl0;
272
273 k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
274 p2 = sp[0]; p3 = sp[sll]; p4 = sp[2*sll];
275 sp += 3*sll;
276
277#ifdef __SUNPRO_C
278#pragma pipeloop(0)
279#endif /* __SUNPRO_C */
280 for (j = 0; j < hsize; j += 2) {
281 p0 = p2; p1 = p3; p2 = p4;
282 p3 = sp[0];
283 p4 = sp[sll];
284
285 pbuff[j ] += p0*k0 + p1*k1 + p2*k2 + p3*k3;
286 pbuff[j + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3;
287
288 sp += 2*sll;
289 }
290
291 sl0 += 4*sll;
292 }
293
294 pk = k + off;
295 sp = sl0;
296
297 k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
298 p2 = sp[0]; p3 = sp[sll]; p4 = sp[2*sll];
299
300 dp = dl;
301 kh = n - off;
302
303 if (kh == 4) {
304 sp += 3*sll;
305
306#ifdef __SUNPRO_C
307#pragma pipeloop(0)
308#endif /* __SUNPRO_C */
309 for (j = 0; j <= (hsize - 2); j += 2) {
310 p0 = p2; p1 = p3; p2 = p4;
311 p3 = sp[0];
312 p4 = sp[sll];
313
314 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + pbuff[j]);
315 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + pbuff[j + 1]);
316
317 dp[0 ] = FROM_S32(d0);
318 dp[dll] = FROM_S32(d1);
319
320 pbuff[j] = 0;
321 pbuff[j + 1] = 0;
322
323 sp += 2*sll;
324 dp += 2*dll;
325 }
326
327 if (j < hsize) {
328 p0 = p2; p1 = p3; p2 = p4;
329 p3 = sp[0];
330
331 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + pbuff[j]);
332
333 pbuff[j] = 0;
334
335 dp[0] = FROM_S32(d0);
336 }
337
338 } else if (kh == 3) {
339 sp += 2*sll;
340
341#ifdef __SUNPRO_C
342#pragma pipeloop(0)
343#endif /* __SUNPRO_C */
344 for (j = 0; j <= (hsize - 2); j += 2) {
345 p0 = p2; p1 = p3;
346 p2 = sp[0];
347 p3 = sp[sll];
348
349 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + pbuff[j]);
350 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + pbuff[j + 1]);
351
352 dp[0 ] = FROM_S32(d0);
353 dp[dll] = FROM_S32(d1);
354
355 pbuff[j] = 0;
356 pbuff[j + 1] = 0;
357
358 sp += 2*sll;
359 dp += 2*dll;
360 }
361
362 if (j < hsize) {
363 p0 = p2; p1 = p3;
364 p2 = sp[0];
365
366 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + pbuff[j]);
367
368 pbuff[j] = 0;
369
370 dp[0] = FROM_S32(d0);
371 }
372
373 } else if (kh == 2) {
374 sp += sll;
375
376#ifdef __SUNPRO_C
377#pragma pipeloop(0)
378#endif /* __SUNPRO_C */
379 for (j = 0; j <= (hsize - 2); j += 2) {
380 p0 = p2;
381 p1 = sp[0];
382 p2 = sp[sll];
383
384 d0 = D2I(p0*k0 + p1*k1 + pbuff[j]);
385 d1 = D2I(p1*k0 + p2*k1 + pbuff[j + 1]);
386
387 dp[0 ] = FROM_S32(d0);
388 dp[dll] = FROM_S32(d1);
389
390 pbuff[j] = 0;
391 pbuff[j + 1] = 0;
392
393 sp += 2*sll;
394 dp += 2*dll;
395 }
396
397 if (j < hsize) {
398 p0 = p2;
399 p1 = sp[0];
400
401 d0 = D2I(p0*k0 + p1*k1 + pbuff[j]);
402
403 pbuff[j] = 0;
404
405 dp[0] = FROM_S32(d0);
406 }
407
408 } else /* if (kh == 1) */ {
409#ifdef __SUNPRO_C
410#pragma pipeloop(0)
411#endif /* __SUNPRO_C */
412 for (j = 0; j < hsize; j++) {
413 p0 = sp[0];
414
415 d0 = D2I(p0*k0 + pbuff[j]);
416
417 dp[0] = FROM_S32(d0);
418
419 pbuff[j] = 0;
420
421 sp += sll;
422 dp += dll;
423 }
424 }
425
426 sl += chan1;
427 dl += chan1;
428 }
429 }
430
431 sl_c += max_hsize*sll;
432 dl_c += max_hsize*dll;
433 }
434
435 if (pbuff != buff) mlib_free(pbuff);
436
437 return MLIB_SUCCESS;
438}
439
440/***************************************************************/
441mlib_status CONV_FUNC(MxN)(mlib_image *dst,
442 const mlib_image *src,
443 const mlib_s32 *kernel,
444 mlib_s32 m,
445 mlib_s32 n,
446 mlib_s32 dm,
447 mlib_s32 dn,
448 mlib_s32 scale,
449 mlib_s32 cmask)
450{
451 FTYPE buff[BUFF_SIZE], *buffs_arr[2*(MAX_N + 1)];
452 FTYPE **buffs = buffs_arr, *buffd;
453 FTYPE akernel[256], *k = akernel, fscale = DSCALE;
454 mlib_s32 mn, l, off, kw, bsize, buff_ind;
455 mlib_s32 d0, d1;
456 FTYPE k0, k1, k2, k3, k4, k5, k6;
457 FTYPE p0, p1, p2, p3, p4, p5, p6, p7;
458 d64_2x32 dd;
459 DEF_VARS(DTYPE);
460 mlib_s32 chan2;
461 mlib_s32 *buffo, *buffi;
462 mlib_status status = MLIB_SUCCESS;
463
464 GET_SRC_DST_PARAMETERS(DTYPE);
465
466 if (scale > 30) {
467 fscale *= 1.0/(1 << 30);
468 scale -= 30;
469 }
470
471 fscale /= (1 << scale);
472
473 mn = m*n;
474
475 if (mn > 256) {
476 k = mlib_malloc(mn*sizeof(mlib_d64));
477
478 if (k == NULL) return MLIB_FAILURE;
479 }
480
481 for (i = 0; i < mn; i++) {
482 k[i] = kernel[i]*fscale;
483 }
484
485 if (m == 1) {
486 status = mlib_ImageConv1xN(dst, src, k, n, dn, cmask);
487 FREE_AND_RETURN_STATUS;
488 }
489
490 bsize = (n + 3)*wid;
491
492 if ((bsize > BUFF_SIZE) || (n > MAX_N)) {
493 pbuff = mlib_malloc(sizeof(FTYPE)*bsize + sizeof(FTYPE *)*2*(n + 1));
494
495 if (pbuff == NULL) {
496 status = MLIB_FAILURE;
497 FREE_AND_RETURN_STATUS;
498 }
499 buffs = (FTYPE **)(pbuff + bsize);
500 }
501
502 for (l = 0; l < (n + 1); l++) buffs[l] = pbuff + l*wid;
503 for (l = 0; l < (n + 1); l++) buffs[l + (n + 1)] = buffs[l];
504 buffd = buffs[n] + wid;
505 buffo = (mlib_s32*)(buffd + wid);
506 buffi = buffo + (wid &~ 1);
507
508 chan1 = nchannel;
509 chan2 = chan1 + chan1;
510
511 wid -= (m - 1);
512 hgt -= (n - 1);
513 adr_dst += dn*dll + dm*nchannel;
514
515 for (c = 0; c < nchannel; c++) {
516 if (!(cmask & (1 << (chan1 - 1 - c)))) continue;
517
518 sl = adr_src + c;
519 dl = adr_dst + c;
520
521 for (l = 0; l < n; l++) {
522 FTYPE *buff = buffs[l];
523
524#ifdef __SUNPRO_C
525#pragma pipeloop(0)
526#endif /* __SUNPRO_C */
527 for (i = 0; i < wid + (m - 1); i++) {
528 buff[i] = (FTYPE)sl[i*chan1];
529 }
530
531 sl += sll;
532 }
533
534 buff_ind = 0;
535
536#ifdef __SUNPRO_C
537#pragma pipeloop(0)
538#endif /* __SUNPRO_C */
539 for (i = 0; i < wid; i++) buffd[i] = 0.0;
540
541 for (j = 0; j < hgt; j++) {
542 FTYPE **buffc = buffs + buff_ind;
543 FTYPE *buffn = buffc[n];
544 FTYPE *pk = k;
545
546 for (l = 0; l < n; l++) {
547 FTYPE *buff_l = buffc[l];
548
549 for (off = 0; off < m;) {
550 FTYPE *buff = buff_l + off;
551
552 kw = m - off;
553
554 if (kw > 2*MAX_KER) kw = MAX_KER; else
555 if (kw > MAX_KER) kw = kw/2;
556 off += kw;
557
558 sp = sl;
559 dp = dl;
560
561 p2 = buff[0]; p3 = buff[1]; p4 = buff[2];
562 p5 = buff[3]; p6 = buff[4]; p7 = buff[5];
563
564 k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
565 k4 = pk[4]; k5 = pk[5]; k6 = pk[6];
566 pk += kw;
567
568 if (kw == 7) {
569
570 if (l < (n - 1) || off < m) {
571#ifdef __SUNPRO_C
572#pragma pipeloop(0)
573#endif /* __SUNPRO_C */
574 for (i = 0; i <= (wid - 2); i += 2) {
575 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
576
577 p6 = buff[i + 6]; p7 = buff[i + 7];
578
579 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6;
580 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6;
581 }
582
583 } else {
584#ifdef __SUNPRO_C
585#pragma pipeloop(0)
586#endif /* __SUNPRO_C */
587 for (i = 0; i <= (wid - 2); i += 2) {
588 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
589
590 p6 = buff[i + 6]; p7 = buff[i + 7];
591
592 LOAD_BUFF(buffi);
593
594 dd.d64 = *(FTYPE *)(buffi + i);
595 buffn[i ] = (FTYPE)dd.i32s.i0;
596 buffn[i + 1] = (FTYPE)dd.i32s.i1;
597
598 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6 + buffd[i ]);
599 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6 + buffd[i + 1]);
600
601 dp[0 ] = FROM_S32(d0);
602 dp[chan1] = FROM_S32(d1);
603
604 buffd[i ] = 0.0;
605 buffd[i + 1] = 0.0;
606
607 sp += chan2;
608 dp += chan2;
609 }
610 }
611
612 } else if (kw == 6) {
613
614 if (l < (n - 1) || off < m) {
615#ifdef __SUNPRO_C
616#pragma pipeloop(0)
617#endif /* __SUNPRO_C */
618 for (i = 0; i <= (wid - 2); i += 2) {
619 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
620
621 p5 = buff[i + 5]; p6 = buff[i + 6];
622
623 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5;
624 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5;
625 }
626
627 } else {
628#ifdef __SUNPRO_C
629#pragma pipeloop(0)
630#endif /* __SUNPRO_C */
631 for (i = 0; i <= (wid - 2); i += 2) {
632 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
633
634 p5 = buff[i + 5]; p6 = buff[i + 6];
635
636 buffn[i ] = (FTYPE)sp[0];
637 buffn[i + 1] = (FTYPE)sp[chan1];
638
639 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + buffd[i ]);
640 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + buffd[i + 1]);
641
642 dp[0 ] = FROM_S32(d0);
643 dp[chan1] = FROM_S32(d1);
644
645 buffd[i ] = 0.0;
646 buffd[i + 1] = 0.0;
647
648 sp += chan2;
649 dp += chan2;
650 }
651 }
652
653 } else if (kw == 5) {
654
655 if (l < (n - 1) || off < m) {
656#ifdef __SUNPRO_C
657#pragma pipeloop(0)
658#endif /* __SUNPRO_C */
659 for (i = 0; i <= (wid - 2); i += 2) {
660 p0 = p2; p1 = p3; p2 = p4; p3 = p5;
661
662 p4 = buff[i + 4]; p5 = buff[i + 5];
663
664 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4;
665 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4;
666 }
667
668 } else {
669#ifdef __SUNPRO_C
670#pragma pipeloop(0)
671#endif /* __SUNPRO_C */
672 for (i = 0; i <= (wid - 2); i += 2) {
673 p0 = p2; p1 = p3; p2 = p4; p3 = p5;
674
675 p4 = buff[i + 4]; p5 = buff[i + 5];
676
677 buffn[i ] = (FTYPE)sp[0];
678 buffn[i + 1] = (FTYPE)sp[chan1];
679
680 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + buffd[i ]);
681 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + buffd[i + 1]);
682
683 dp[0 ] = FROM_S32(d0);
684 dp[chan1] = FROM_S32(d1);
685
686 buffd[i ] = 0.0;
687 buffd[i + 1] = 0.0;
688
689 sp += chan2;
690 dp += chan2;
691 }
692 }
693
694 } else if (kw == 4) {
695
696 if (l < (n - 1) || off < m) {
697#ifdef __SUNPRO_C
698#pragma pipeloop(0)
699#endif /* __SUNPRO_C */
700 for (i = 0; i <= (wid - 2); i += 2) {
701 p0 = p2; p1 = p3; p2 = p4;
702
703 p3 = buff[i + 3]; p4 = buff[i + 4];
704
705 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3;
706 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3;
707 }
708
709 } else {
710#ifdef __SUNPRO_C
711#pragma pipeloop(0)
712#endif /* __SUNPRO_C */
713 for (i = 0; i <= (wid - 2); i += 2) {
714 p0 = p2; p1 = p3; p2 = p4;
715
716 p3 = buff[i + 3]; p4 = buff[i + 4];
717
718 buffn[i ] = (FTYPE)sp[0];
719 buffn[i + 1] = (FTYPE)sp[chan1];
720
721 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + buffd[i ]);
722 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + buffd[i + 1]);
723
724 dp[0 ] = FROM_S32(d0);
725 dp[chan1] = FROM_S32(d1);
726
727 buffd[i ] = 0.0;
728 buffd[i + 1] = 0.0;
729
730 sp += chan2;
731 dp += chan2;
732 }
733 }
734
735 } else if (kw == 3) {
736
737 if (l < (n - 1) || off < m) {
738#ifdef __SUNPRO_C
739#pragma pipeloop(0)
740#endif /* __SUNPRO_C */
741 for (i = 0; i <= (wid - 2); i += 2) {
742 p0 = p2; p1 = p3;
743
744 p2 = buff[i + 2]; p3 = buff[i + 3];
745
746 buffd[i ] += p0*k0 + p1*k1 + p2*k2;
747 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2;
748 }
749
750 } else {
751#ifdef __SUNPRO_C
752#pragma pipeloop(0)
753#endif /* __SUNPRO_C */
754 for (i = 0; i <= (wid - 2); i += 2) {
755 p0 = p2; p1 = p3;
756
757 p2 = buff[i + 2]; p3 = buff[i + 3];
758
759 buffn[i ] = (FTYPE)sp[0];
760 buffn[i + 1] = (FTYPE)sp[chan1];
761
762 d0 = D2I(p0*k0 + p1*k1 + p2*k2 + buffd[i ]);
763 d1 = D2I(p1*k0 + p2*k1 + p3*k2 + buffd[i + 1]);
764
765 dp[0 ] = FROM_S32(d0);
766 dp[chan1] = FROM_S32(d1);
767
768 buffd[i ] = 0.0;
769 buffd[i + 1] = 0.0;
770
771 sp += chan2;
772 dp += chan2;
773 }
774 }
775
776 } else /*if (kw == 2)*/ {
777
778 if (l < (n - 1) || off < m) {
779#ifdef __SUNPRO_C
780#pragma pipeloop(0)
781#endif /* __SUNPRO_C */
782 for (i = 0; i <= (wid - 2); i += 2) {
783 p0 = p2;
784
785 p1 = buff[i + 1]; p2 = buff[i + 2];
786
787 buffd[i ] += p0*k0 + p1*k1;
788 buffd[i + 1] += p1*k0 + p2*k1;
789 }
790
791 } else {
792#ifdef __SUNPRO_C
793#pragma pipeloop(0)
794#endif /* __SUNPRO_C */
795 for (i = 0; i <= (wid - 2); i += 2) {
796 p0 = p2;
797
798 p1 = buff[i + 1]; p2 = buff[i + 2];
799
800 buffn[i ] = (FTYPE)sp[0];
801 buffn[i + 1] = (FTYPE)sp[chan1];
802
803 d0 = D2I(p0*k0 + p1*k1 + buffd[i ]);
804 d1 = D2I(p1*k0 + p2*k1 + buffd[i + 1]);
805
806 dp[0 ] = FROM_S32(d0);
807 dp[chan1] = FROM_S32(d1);
808
809 buffd[i ] = 0.0;
810 buffd[i + 1] = 0.0;
811
812 sp += chan2;
813 dp += chan2;
814 }
815 }
816 }
817 }
818 }
819
820 /* last pixels */
821 for (; i < wid; i++) {
822 FTYPE *pk = k, s = 0;
823 mlib_s32 x, d0;
824
825 for (l = 0; l < n; l++) {
826 FTYPE *buff = buffc[l] + i;
827
828 for (x = 0; x < m; x++) s += buff[x] * (*pk++);
829 }
830
831 d0 = D2I(s);
832 dp[0] = FROM_S32(d0);
833
834 buffn[i] = (FTYPE)sp[0];
835
836 sp += chan1;
837 dp += chan1;
838 }
839
840 for (l = 0; l < (m - 1); l++) buffn[wid + l] = sp[l*chan1];
841
842 /* next line */
843 sl += sll;
844 dl += dll;
845
846 buff_ind++;
847
848 if (buff_ind >= n + 1) buff_ind = 0;
849 }
850 }
851
852 FREE_AND_RETURN_STATUS;
853}
854
855/***************************************************************/
856#ifndef __sparc /* for x86, using integer multiplies is faster */
857
858#define STORE_RES(res, x) \
859 x >>= shift2; \
860 CLAMP_STORE(res, x)
861
862mlib_status CONV_FUNC_I(MxN)(mlib_image *dst,
863 const mlib_image *src,
864 const mlib_s32 *kernel,
865 mlib_s32 m,
866 mlib_s32 n,
867 mlib_s32 dm,
868 mlib_s32 dn,
869 mlib_s32 scale,
870 mlib_s32 cmask)
871{
872 mlib_s32 buff[BUFF_SIZE], *buffd = buff;
873 mlib_s32 l, off, kw;
874 mlib_s32 d0, d1, shift1, shift2;
875 mlib_s32 k0, k1, k2, k3, k4, k5, k6;
876 mlib_s32 p0, p1, p2, p3, p4, p5, p6, p7;
877 DTYPE *adr_src, *sl, *sp = NULL;
878 DTYPE *adr_dst, *dl, *dp = NULL;
879 mlib_s32 wid, hgt, sll, dll;
880 mlib_s32 nchannel, chan1;
881 mlib_s32 i, j, c;
882 mlib_s32 chan2;
883 mlib_s32 k_locl[MAX_N*MAX_N], *k = k_locl;
884 GET_SRC_DST_PARAMETERS(DTYPE);
885
886#if IMG_TYPE != 1
887 shift1 = 16;
888#else
889 shift1 = 8;
890#endif /* IMG_TYPE != 1 */
891 shift2 = scale - shift1;
892
893 chan1 = nchannel;
894 chan2 = chan1 + chan1;
895
896 wid -= (m - 1);
897 hgt -= (n - 1);
898 adr_dst += dn*dll + dm*nchannel;
899
900 if (wid > BUFF_SIZE) {
901 buffd = mlib_malloc(sizeof(mlib_s32)*wid);
902
903 if (buffd == NULL) return MLIB_FAILURE;
904 }
905
906 if (m*n > MAX_N*MAX_N) {
907 k = mlib_malloc(sizeof(mlib_s32)*(m*n));
908
909 if (k == NULL) {
910 if (buffd != buff) mlib_free(buffd);
911 return MLIB_FAILURE;
912 }
913 }
914
915 for (i = 0; i < m*n; i++) {
916 k[i] = kernel[i] >> shift1;
917 }
918
919 for (c = 0; c < nchannel; c++) {
920 if (!(cmask & (1 << (nchannel - 1 - c)))) continue;
921
922 sl = adr_src + c;
923 dl = adr_dst + c;
924
925#ifdef __SUNPRO_C
926#pragma pipeloop(0)
927#endif /* __SUNPRO_C */
928 for (i = 0; i < wid; i++) buffd[i] = 0;
929
930 for (j = 0; j < hgt; j++) {
931 mlib_s32 *pk = k;
932
933 for (l = 0; l < n; l++) {
934 DTYPE *sp0 = sl + l*sll;
935
936 for (off = 0; off < m;) {
937 sp = sp0 + off*chan1;
938 dp = dl;
939
940 kw = m - off;
941
942 if (kw > 2*MAX_KER) kw = MAX_KER; else
943 if (kw > MAX_KER) kw = kw/2;
944 off += kw;
945
946 p2 = sp[0]; p3 = sp[chan1]; p4 = sp[chan2];
947 p5 = sp[chan2 + chan1]; p6 = sp[chan2 + chan2]; p7 = sp[5*chan1];
948
949 k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
950 k4 = pk[4]; k5 = pk[5]; k6 = pk[6];
951 pk += kw;
952
953 sp += (kw - 1)*chan1;
954
955 if (kw == 7) {
956
957 if (l < (n - 1) || off < m) {
958#ifdef __SUNPRO_C
959#pragma pipeloop(0)
960#endif /* __SUNPRO_C */
961 for (i = 0; i <= (wid - 2); i += 2) {
962 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
963 p6 = sp[0];
964 p7 = sp[chan1];
965
966 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6;
967 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6;
968
969 sp += chan2;
970 }
971
972 } else {
973#ifdef __SUNPRO_C
974#pragma pipeloop(0)
975#endif /* __SUNPRO_C */
976 for (i = 0; i <= (wid - 2); i += 2) {
977 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
978 p6 = sp[0];
979 p7 = sp[chan1];
980
981 d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6 + buffd[i ]);
982 d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6 + buffd[i + 1]);
983
984 STORE_RES(dp[0 ], d0);
985 STORE_RES(dp[chan1], d1);
986
987 buffd[i ] = 0;
988 buffd[i + 1] = 0;
989
990 sp += chan2;
991 dp += chan2;
992 }
993 }
994
995 } else if (kw == 6) {
996
997 if (l < (n - 1) || off < m) {
998#ifdef __SUNPRO_C
999#pragma pipeloop(0)
1000#endif /* __SUNPRO_C */
1001 for (i = 0; i <= (wid - 2); i += 2) {
1002 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
1003 p5 = sp[0];
1004 p6 = sp[chan1];
1005
1006 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5;
1007 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5;
1008
1009 sp += chan2;
1010 }
1011
1012 } else {
1013#ifdef __SUNPRO_C
1014#pragma pipeloop(0)
1015#endif /* __SUNPRO_C */
1016 for (i = 0; i <= (wid - 2); i += 2) {
1017 p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
1018 p5 = sp[0];
1019 p6 = sp[chan1];
1020
1021 d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + buffd[i ]);
1022 d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + buffd[i + 1]);
1023
1024 STORE_RES(dp[0 ], d0);
1025 STORE_RES(dp[chan1], d1);
1026
1027 buffd[i ] = 0;
1028 buffd[i + 1] = 0;
1029
1030 sp += chan2;
1031 dp += chan2;
1032 }
1033 }
1034
1035 } else if (kw == 5) {
1036
1037 if (l < (n - 1) || off < m) {
1038#ifdef __SUNPRO_C
1039#pragma pipeloop(0)
1040#endif /* __SUNPRO_C */
1041 for (i = 0; i <= (wid - 2); i += 2) {
1042 p0 = p2; p1 = p3; p2 = p4; p3 = p5;
1043 p4 = sp[0];
1044 p5 = sp[chan1];
1045
1046 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4;
1047 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4;
1048
1049 sp += chan2;
1050 }
1051
1052 } else {
1053#ifdef __SUNPRO_C
1054#pragma pipeloop(0)
1055#endif /* __SUNPRO_C */
1056 for (i = 0; i <= (wid - 2); i += 2) {
1057 p0 = p2; p1 = p3; p2 = p4; p3 = p5;
1058 p4 = sp[0];
1059 p5 = sp[chan1];
1060
1061 d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + buffd[i ]);
1062 d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + buffd[i + 1]);
1063
1064 STORE_RES(dp[0 ], d0);
1065 STORE_RES(dp[chan1], d1);
1066
1067 buffd[i ] = 0;
1068 buffd[i + 1] = 0;
1069
1070 sp += chan2;
1071 dp += chan2;
1072 }
1073 }
1074
1075 } else if (kw == 4) {
1076
1077 if (l < (n - 1) || off < m) {
1078#ifdef __SUNPRO_C
1079#pragma pipeloop(0)
1080#endif /* __SUNPRO_C */
1081 for (i = 0; i <= (wid - 2); i += 2) {
1082 p0 = p2; p1 = p3; p2 = p4;
1083 p3 = sp[0];
1084 p4 = sp[chan1];
1085
1086 buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3;
1087 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3;
1088
1089 sp += chan2;
1090 }
1091
1092 } else {
1093#ifdef __SUNPRO_C
1094#pragma pipeloop(0)
1095#endif /* __SUNPRO_C */
1096 for (i = 0; i <= (wid - 2); i += 2) {
1097 p0 = p2; p1 = p3; p2 = p4;
1098 p3 = sp[0];
1099 p4 = sp[chan1];
1100
1101 d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + buffd[i ]);
1102 d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + buffd[i + 1]);
1103
1104 STORE_RES(dp[0 ], d0);
1105 STORE_RES(dp[chan1], d1);
1106
1107 buffd[i ] = 0;
1108 buffd[i + 1] = 0;
1109
1110 sp += chan2;
1111 dp += chan2;
1112 }
1113 }
1114
1115 } else if (kw == 3) {
1116
1117 if (l < (n - 1) || off < m) {
1118#ifdef __SUNPRO_C
1119#pragma pipeloop(0)
1120#endif /* __SUNPRO_C */
1121 for (i = 0; i <= (wid - 2); i += 2) {
1122 p0 = p2; p1 = p3;
1123 p2 = sp[0];
1124 p3 = sp[chan1];
1125
1126 buffd[i ] += p0*k0 + p1*k1 + p2*k2;
1127 buffd[i + 1] += p1*k0 + p2*k1 + p3*k2;
1128
1129 sp += chan2;
1130 }
1131
1132 } else {
1133#ifdef __SUNPRO_C
1134#pragma pipeloop(0)
1135#endif /* __SUNPRO_C */
1136 for (i = 0; i <= (wid - 2); i += 2) {
1137 p0 = p2; p1 = p3;
1138 p2 = sp[0];
1139 p3 = sp[chan1];
1140
1141 d0 = (p0*k0 + p1*k1 + p2*k2 + buffd[i ]);
1142 d1 = (p1*k0 + p2*k1 + p3*k2 + buffd[i + 1]);
1143
1144 STORE_RES(dp[0 ], d0);
1145 STORE_RES(dp[chan1], d1);
1146
1147 buffd[i ] = 0;
1148 buffd[i + 1] = 0;
1149
1150 sp += chan2;
1151 dp += chan2;
1152 }
1153 }
1154
1155 } else if (kw == 2) {
1156
1157 if (l < (n - 1) || off < m) {
1158#ifdef __SUNPRO_C
1159#pragma pipeloop(0)
1160#endif /* __SUNPRO_C */
1161 for (i = 0; i <= (wid - 2); i += 2) {
1162 p0 = p2;
1163 p1 = sp[0];
1164 p2 = sp[chan1];
1165
1166 buffd[i ] += p0*k0 + p1*k1;
1167 buffd[i + 1] += p1*k0 + p2*k1;
1168
1169 sp += chan2;
1170 }
1171
1172 } else {
1173#ifdef __SUNPRO_C
1174#pragma pipeloop(0)
1175#endif /* __SUNPRO_C */
1176 for (i = 0; i <= (wid - 2); i += 2) {
1177 p0 = p2;
1178 p1 = sp[0];
1179 p2 = sp[chan1];
1180
1181 d0 = (p0*k0 + p1*k1 + buffd[i ]);
1182 d1 = (p1*k0 + p2*k1 + buffd[i + 1]);
1183
1184 STORE_RES(dp[0 ], d0);
1185 STORE_RES(dp[chan1], d1);
1186
1187 buffd[i ] = 0;
1188 buffd[i + 1] = 0;
1189
1190 sp += chan2;
1191 dp += chan2;
1192 }
1193 }
1194
1195 } else /*if (kw == 1)*/ {
1196
1197 if (l < (n - 1) || off < m) {
1198#ifdef __SUNPRO_C
1199#pragma pipeloop(0)
1200#endif /* __SUNPRO_C */
1201 for (i = 0; i <= (wid - 2); i += 2) {
1202 p0 = sp[0];
1203 p1 = sp[chan1];
1204
1205 buffd[i ] += p0*k0;
1206 buffd[i + 1] += p1*k0;
1207
1208 sp += chan2;
1209 }
1210
1211 } else {
1212#ifdef __SUNPRO_C
1213#pragma pipeloop(0)
1214#endif /* __SUNPRO_C */
1215 for (i = 0; i <= (wid - 2); i += 2) {
1216 p0 = sp[0];
1217 p1 = sp[chan1];
1218
1219 d0 = (p0*k0 + buffd[i ]);
1220 d1 = (p1*k0 + buffd[i + 1]);
1221
1222 STORE_RES(dp[0 ], d0);
1223 STORE_RES(dp[chan1], d1);
1224
1225 buffd[i ] = 0;
1226 buffd[i + 1] = 0;
1227
1228 sp += chan2;
1229 dp += chan2;
1230 }
1231 }
1232 }
1233 }
1234 }
1235
1236 /* last pixels */
1237 for (; i < wid; i++) {
1238 mlib_s32 *pk = k, s = 0;
1239 mlib_s32 x;
1240
1241 for (l = 0; l < n; l++) {
1242 sp = sl + l*sll + i*chan1;
1243
1244 for (x = 0; x < m; x++) {
1245 s += sp[0] * pk[0];
1246 sp += chan1;
1247 pk ++;
1248 }
1249 }
1250
1251 STORE_RES(dp[0], s);
1252
1253 sp += chan1;
1254 dp += chan1;
1255 }
1256
1257 sl += sll;
1258 dl += dll;
1259 }
1260 }
1261
1262 if (buffd != buff) mlib_free(buffd);
1263 if (k != k_locl) mlib_free(k);
1264
1265 return MLIB_SUCCESS;
1266}
1267
1268/***************************************************************/
1269#endif /* __sparc ( for x86, using integer multiplies is faster ) */
1270
1271/***************************************************************/
1272