522 fscale *= 1.0/(1 << 30);
523 scale -= 30;
524 }
525
526 fscale /= (1 << scale);
527
528 mn = m*n;
529
530 if (mn > 256) {
531 k = mlib_malloc(mn*sizeof(mlib_d64));
532
533 if (k == NULL) return MLIB_FAILURE;
534 }
535
536 for (i = 0; i < mn; i++) {
537 k[i] = kernel[i]*fscale;
538 }
539
540 if (m == 1) {
541 status = mlib_ImageConv1xN_ext(dst, src, k, n, dy_t, dy_b, cmask);
542 FREE_AND_RETURN_STATUS;
543 }
544
545 swid = wid + (m - 1);
546
547 bsize = (n + 3)*swid;
548
549 if ((bsize > BUFF_SIZE) || (n > MAX_N)) {
550 pbuff = mlib_malloc(sizeof(FTYPE)*bsize + sizeof(FTYPE *)*2*(n + 1));
551
552 if (pbuff == NULL) {
553 status = MLIB_FAILURE;
554 FREE_AND_RETURN_STATUS;
555 }
556 buffs = (FTYPE **)(pbuff + bsize);
557 }
558
559 for (l = 0; l < (n + 1); l++) buffs[l] = pbuff + l*swid;
560 for (l = 0; l < (n + 1); l++) buffs[l + (n + 1)] = buffs[l];
561 buffd = buffs[n] + swid;
562 buffi = (mlib_s32*)(buffd + swid);
563
564 chan1 = nchannel;
565 chan2 = chan1 + chan1;
566
567 swid -= (dx_l + dx_r);
568
569 for (c = 0; c < nchannel; c++) {
570 if (!(cmask & (1 << (chan1 - 1 - c)))) continue;
571
572 sl = adr_src + c;
573 dl = adr_dst + c;
574
939
940 for (; i < swid; i++) {
941 buffn[i + dx_l] = (FTYPE)sp[0];
942 sp += chan1;
943 }
944
945 for (i = 0; i < dx_l; i++) buffn[i] = buffn[dx_l];
946 for (i = 0; i < dx_r; i++) buffn[swid + dx_l + i] = buffn[swid + dx_l - 1];
947
948 /* next line */
949
950 if (j < hgt - dy_b - 2) sl += sll;
951 dl += dll;
952
953 buff_ind++;
954
955 if (buff_ind >= n + 1) buff_ind = 0;
956 }
957 }
958
959 FREE_AND_RETURN_STATUS;
960 }
961
962 /***************************************************************/
963 #ifndef __sparc /* for x86, using integer multiplies is faster */
964
965 #define STORE_RES(res, x) \
966 x >>= shift2; \
967 CLAMP_STORE(res, x)
968
969 mlib_status CONV_FUNC_MxN_I
970 {
971 DTYPE *adr_src, *sl, *sp = NULL;
972 DTYPE *adr_dst, *dl, *dp = NULL;
973 mlib_s32 buff[BUFF_SIZE], *buffs_arr[2*(MAX_N + 1)];
974 mlib_s32 *pbuff = buff;
975 mlib_s32 **buffs = buffs_arr, *buffd;
976 mlib_s32 l, off, kw, bsize, buff_ind;
977 mlib_s32 d0, d1, shift1, shift2;
978 mlib_s32 k0, k1, k2, k3, k4, k5, k6;
979 mlib_s32 p0, p1, p2, p3, p4, p5, p6, p7;
|
522 fscale *= 1.0/(1 << 30);
523 scale -= 30;
524 }
525
526 fscale /= (1 << scale);
527
528 mn = m*n;
529
530 if (mn > 256) {
531 k = mlib_malloc(mn*sizeof(mlib_d64));
532
533 if (k == NULL) return MLIB_FAILURE;
534 }
535
536 for (i = 0; i < mn; i++) {
537 k[i] = kernel[i]*fscale;
538 }
539
540 if (m == 1) {
541 status = mlib_ImageConv1xN_ext(dst, src, k, n, dy_t, dy_b, cmask);
542 FREE_AND_RETURN_STATUS
543 }
544
545 swid = wid + (m - 1);
546
547 bsize = (n + 3)*swid;
548
549 if ((bsize > BUFF_SIZE) || (n > MAX_N)) {
550 pbuff = mlib_malloc(sizeof(FTYPE)*bsize + sizeof(FTYPE *)*2*(n + 1));
551
552 if (pbuff == NULL) {
553 status = MLIB_FAILURE;
554 FREE_AND_RETURN_STATUS
555 }
556 buffs = (FTYPE **)(pbuff + bsize);
557 }
558
559 for (l = 0; l < (n + 1); l++) buffs[l] = pbuff + l*swid;
560 for (l = 0; l < (n + 1); l++) buffs[l + (n + 1)] = buffs[l];
561 buffd = buffs[n] + swid;
562 buffi = (mlib_s32*)(buffd + swid);
563
564 chan1 = nchannel;
565 chan2 = chan1 + chan1;
566
567 swid -= (dx_l + dx_r);
568
569 for (c = 0; c < nchannel; c++) {
570 if (!(cmask & (1 << (chan1 - 1 - c)))) continue;
571
572 sl = adr_src + c;
573 dl = adr_dst + c;
574
939
940 for (; i < swid; i++) {
941 buffn[i + dx_l] = (FTYPE)sp[0];
942 sp += chan1;
943 }
944
945 for (i = 0; i < dx_l; i++) buffn[i] = buffn[dx_l];
946 for (i = 0; i < dx_r; i++) buffn[swid + dx_l + i] = buffn[swid + dx_l - 1];
947
948 /* next line */
949
950 if (j < hgt - dy_b - 2) sl += sll;
951 dl += dll;
952
953 buff_ind++;
954
955 if (buff_ind >= n + 1) buff_ind = 0;
956 }
957 }
958
959 FREE_AND_RETURN_STATUS
960 }
961
962 /***************************************************************/
963 #ifndef __sparc /* for x86, using integer multiplies is faster */
964
965 #define STORE_RES(res, x) \
966 x >>= shift2; \
967 CLAMP_STORE(res, x)
968
969 mlib_status CONV_FUNC_MxN_I
970 {
971 DTYPE *adr_src, *sl, *sp = NULL;
972 DTYPE *adr_dst, *dl, *dp = NULL;
973 mlib_s32 buff[BUFF_SIZE], *buffs_arr[2*(MAX_N + 1)];
974 mlib_s32 *pbuff = buff;
975 mlib_s32 **buffs = buffs_arr, *buffd;
976 mlib_s32 l, off, kw, bsize, buff_ind;
977 mlib_s32 d0, d1, shift1, shift2;
978 mlib_s32 k0, k1, k2, k3, k4, k5, k6;
979 mlib_s32 p0, p1, p2, p3, p4, p5, p6, p7;
|