512 fscale *= 1.0/(1 << 30);
513 scale -= 30;
514 }
515
516 fscale /= (1 << scale);
517
518 mn = m*n;
519
520 if (mn > 256) {
521 k = mlib_malloc(mn*sizeof(mlib_d64));
522
523 if (k == NULL) return MLIB_FAILURE;
524 }
525
526 for (i = 0; i < mn; i++) {
527 k[i] = kernel[i]*fscale;
528 }
529
530 if (m == 1) {
531 status = mlib_ImageConv1xN_ext(dst, src, k, n, dy_t, dy_b, cmask);
532 FREE_AND_RETURN_STATUS;
533 }
534
535 swid = wid + (m - 1);
536
537 bsize = (n + 3)*swid;
538
539 if ((bsize > BUFF_SIZE) || (n > MAX_N)) {
540 pbuff = mlib_malloc(sizeof(FTYPE)*bsize + sizeof(FTYPE *)*2*(n + 1));
541
542 if (pbuff == NULL) {
543 status = MLIB_FAILURE;
544 FREE_AND_RETURN_STATUS;
545 }
546 buffs = (FTYPE **)(pbuff + bsize);
547 }
548
549 for (l = 0; l < (n + 1); l++) buffs[l] = pbuff + l*swid;
550 for (l = 0; l < (n + 1); l++) buffs[l + (n + 1)] = buffs[l];
551 buffd = buffs[n] + swid;
552 buffi = (mlib_s32*)(buffd + swid);
553
554 chan1 = nchannel;
555 chan2 = chan1 + chan1;
556
557 swid -= (dx_l + dx_r);
558
559 for (c = 0; c < nchannel; c++) {
560 if (!(cmask & (1 << (chan1 - 1 - c)))) continue;
561
562 sl = adr_src + c;
563 dl = adr_dst + c;
564
929
930 for (; i < swid; i++) {
931 buffn[i + dx_l] = (FTYPE)sp[0];
932 sp += chan1;
933 }
934
935 for (i = 0; i < dx_l; i++) buffn[i] = buffn[dx_l];
936 for (i = 0; i < dx_r; i++) buffn[swid + dx_l + i] = buffn[swid + dx_l - 1];
937
938 /* next line */
939
940 if (j < hgt - dy_b - 2) sl += sll;
941 dl += dll;
942
943 buff_ind++;
944
945 if (buff_ind >= n + 1) buff_ind = 0;
946 }
947 }
948
949 FREE_AND_RETURN_STATUS;
950 }
951
952 /***************************************************************/
953 #ifndef __sparc /* for x86, using integer multiplies is faster */
954
955 #define STORE_RES(res, x) \
956 x >>= shift2; \
957 CLAMP_STORE(res, x)
958
959 mlib_status CONV_FUNC_MxN_I
960 {
961 DTYPE *adr_src, *sl, *sp = NULL;
962 DTYPE *adr_dst, *dl, *dp = NULL;
963 mlib_s32 buff[BUFF_SIZE], *buffs_arr[2*(MAX_N + 1)];
964 mlib_s32 *pbuff = buff;
965 mlib_s32 **buffs = buffs_arr, *buffd;
966 mlib_s32 l, off, kw, bsize, buff_ind;
967 mlib_s32 d0, d1, shift1, shift2;
968 mlib_s32 k0, k1, k2, k3, k4, k5, k6;
969 mlib_s32 p0, p1, p2, p3, p4, p5, p6, p7;
|
512 fscale *= 1.0/(1 << 30);
513 scale -= 30;
514 }
515
516 fscale /= (1 << scale);
517
518 mn = m*n;
519
520 if (mn > 256) {
521 k = mlib_malloc(mn*sizeof(mlib_d64));
522
523 if (k == NULL) return MLIB_FAILURE;
524 }
525
526 for (i = 0; i < mn; i++) {
527 k[i] = kernel[i]*fscale;
528 }
529
530 if (m == 1) {
531 status = mlib_ImageConv1xN_ext(dst, src, k, n, dy_t, dy_b, cmask);
532 FREE_AND_RETURN_STATUS
533 }
534
535 swid = wid + (m - 1);
536
537 bsize = (n + 3)*swid;
538
539 if ((bsize > BUFF_SIZE) || (n > MAX_N)) {
540 pbuff = mlib_malloc(sizeof(FTYPE)*bsize + sizeof(FTYPE *)*2*(n + 1));
541
542 if (pbuff == NULL) {
543 status = MLIB_FAILURE;
544 FREE_AND_RETURN_STATUS
545 }
546 buffs = (FTYPE **)(pbuff + bsize);
547 }
548
549 for (l = 0; l < (n + 1); l++) buffs[l] = pbuff + l*swid;
550 for (l = 0; l < (n + 1); l++) buffs[l + (n + 1)] = buffs[l];
551 buffd = buffs[n] + swid;
552 buffi = (mlib_s32*)(buffd + swid);
553
554 chan1 = nchannel;
555 chan2 = chan1 + chan1;
556
557 swid -= (dx_l + dx_r);
558
559 for (c = 0; c < nchannel; c++) {
560 if (!(cmask & (1 << (chan1 - 1 - c)))) continue;
561
562 sl = adr_src + c;
563 dl = adr_dst + c;
564
929
930 for (; i < swid; i++) {
931 buffn[i + dx_l] = (FTYPE)sp[0];
932 sp += chan1;
933 }
934
935 for (i = 0; i < dx_l; i++) buffn[i] = buffn[dx_l];
936 for (i = 0; i < dx_r; i++) buffn[swid + dx_l + i] = buffn[swid + dx_l - 1];
937
938 /* next line */
939
940 if (j < hgt - dy_b - 2) sl += sll;
941 dl += dll;
942
943 buff_ind++;
944
945 if (buff_ind >= n + 1) buff_ind = 0;
946 }
947 }
948
949 FREE_AND_RETURN_STATUS
950 }
951
952 /***************************************************************/
953 #ifndef __sparc /* for x86, using integer multiplies is faster */
954
955 #define STORE_RES(res, x) \
956 x >>= shift2; \
957 CLAMP_STORE(res, x)
958
959 mlib_status CONV_FUNC_MxN_I
960 {
961 DTYPE *adr_src, *sl, *sp = NULL;
962 DTYPE *adr_dst, *dl, *dp = NULL;
963 mlib_s32 buff[BUFF_SIZE], *buffs_arr[2*(MAX_N + 1)];
964 mlib_s32 *pbuff = buff;
965 mlib_s32 **buffs = buffs_arr, *buffd;
966 mlib_s32 l, off, kw, bsize, buff_ind;
967 mlib_s32 d0, d1, shift1, shift2;
968 mlib_s32 k0, k1, k2, k3, k4, k5, k6;
969 mlib_s32 p0, p1, p2, p3, p4, p5, p6, p7;
|