Welcome to mirror list, hosted at ThFree Co, Russian Federation.

git.blender.org/blender.git - Unnamed repository; edit this file 'description' to name the repository.
summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat (limited to 'source/blender/editors/object/object_lattice.c')
-rw-r--r--source/blender/editors/object/object_lattice.c278
1 files changed, 264 insertions, 14 deletions
diff --git a/source/blender/editors/object/object_lattice.c b/source/blender/editors/object/object_lattice.c
index 1f7be0bf9a6..4aa2e825954 100644
--- a/source/blender/editors/object/object_lattice.c
+++ b/source/blender/editors/object/object_lattice.c
@@ -46,6 +46,7 @@
#include "DNA_scene_types.h"
#include "RNA_access.h"
+#include "RNA_define.h"
#include "BKE_context.h"
#include "BKE_depsgraph.h"
@@ -167,7 +168,7 @@ void load_editLatt(Object *obedit)
}
}
-/************************** Operators *************************/
+/************************** Select All Operator *************************/
void ED_setflagsLatt(Object *obedit, int flag)
{
@@ -254,6 +255,8 @@ void LATTICE_OT_select_all(wmOperatorType *ot)
WM_operator_properties_select_all(ot);
}
+/************************** Make Regular Operator *************************/
+
static int make_regular_poll(bContext *C)
{
Object *ob;
@@ -300,18 +303,265 @@ void LATTICE_OT_make_regular(wmOperatorType *ot)
ot->flag = OPTYPE_REGISTER | OPTYPE_UNDO;
}
+/************************** Flip Verts Operator *************************/
+
+/* flipping options */
+typedef enum eLattice_FlipAxes {
+ LATTICE_FLIP_U = 0,
+ LATTICE_FLIP_V = 1,
+ LATTICE_FLIP_W = 2
+} eLattice_FlipAxes;
+
+/* Helper macro for accessing item at index (u, v, w)
+ * < lt: (Lattice)
+ * < U: (int) u-axis coordinate of point
+ * < V: (int) v-axis coordinate of point
+ * < W: (int) w-axis coordinate of point
+ * < dimU: (int) number of points per row or number of columns (U-Axis)
+ * < dimV: (int) number of rows (V-Axis)
+ * > returns: (BPoint *) pointer to BPoint at this index
+ */
+#define LATTICE_PT(lt, U, V, W, dimU, dimV) \
+ ( (lt)->def + \
+ ((dimU) * (dimV)) * (W) + \
+ (dimU) * (V) + \
+ (U) \
+ )
+
+/* Flip midpoint value so that relative distances between midpoint and neighbour-pair is maintained
+ * ! Assumes that uvw <=> xyz (i.e. axis-aligned index-axes with coordinate-axes)
+ * - Helper for lattice_flip_exec()
+ */
+static void lattice_flip_point_value(Lattice *lt, int u, int v, int w, float mid, eLattice_FlipAxes axis)
+{
+ BPoint *bp;
+ float diff;
+
+ /* just the point in the middle (unpaired) */
+ bp = LATTICE_PT(lt, u, v, w, lt->pntsu, lt->pntsv);
+
+ /* flip over axis */
+ diff = mid - bp->vec[axis];
+ bp->vec[axis] = mid + diff;
+}
+
+/* Swap pairs of lattice points along a specified axis
+ * - Helper for lattice_flip_exec()
+ */
+static void lattice_swap_point_pairs(Lattice *lt, int u, int v, int w, float mid, eLattice_FlipAxes axis)
+{
+ BPoint *bpA, *bpB;
+
+ int numU = lt->pntsu;
+ int numV = lt->pntsv;
+ int numW = lt->pntsw;
+
+ int u0 = u, u1 = u;
+ int v0 = v, v1 = v;
+ int w0 = w, w1 = w;
+
+ /* get pair index by just overriding the relevant pair-value
+ * - "-1" else buffer overflow
+ */
+ switch (axis) {
+ case LATTICE_FLIP_U:
+ u1 = numU - u - 1;
+ break;
+ case LATTICE_FLIP_V:
+ v1 = numV - v - 1;
+ break;
+ case LATTICE_FLIP_W:
+ w1 = numW - w - 1;
+ break;
+ }
+
+ /* get points to operate on */
+ bpA = LATTICE_PT(lt, u0, v0, w0, numU, numV);
+ bpB = LATTICE_PT(lt, u1, v1, w1, numU, numV);
+
+ /* Swap all coordinates, so that flipped coordinates belong to
+ * the indices on the correct side of the lattice.
+ *
+ * Coords: (-2 4) |0| (3 4) --> (3 4) |0| (-2 4)
+ * Indices: (0,L) (1,R) --> (0,L) (1,R)
+ */
+ swap_v3_v3(bpA->vec, bpB->vec);
+
+ /* However, we need to mirror the coordinate values on the axis we're dealing with,
+ * otherwise we'd have effectively only rotated the points around. If we don't do this,
+ * we'd just be reimplementing the naive mirroring algorithm, which causes unwanted deforms
+ * such as flipped normals, etc.
+ *
+ * Coords: (3 4) |0| (-2 4) --\
+ * \-> (-3 4) |0| (2 4)
+ * Indices: (0,L) (1,R) --> (0,L) (1,R)
+ */
+ lattice_flip_point_value(lt, u0, v0, w0, mid, axis);
+ lattice_flip_point_value(lt, u1, v1, w1, mid, axis);
+}
+
+static int lattice_flip_exec(bContext *C, wmOperator *op)
+{
+ Object *obedit = CTX_data_edit_object(C);
+ Lattice *lt;
+
+ eLattice_FlipAxes axis = RNA_enum_get(op->ptr, "axis");
+ int numU, numV, numW;
+ int totP;
+
+ float mid = 0.0f;
+ short isOdd = 0;
+
+ /* get lattice - we need the "edit lattice" from the lattice... confusing... */
+ lt = (Lattice *)obedit->data;
+ lt = lt->editlatt->latt;
+
+ numU = lt->pntsu;
+ numV = lt->pntsv;
+ numW = lt->pntsw;
+ totP = numU * numV * numW;
+
+ /* First Pass: determine midpoint - used for flipping center verts if there are odd number of points on axis */
+ switch (axis) {
+ case LATTICE_FLIP_U:
+ isOdd = numU & 1;
+ break;
+ case LATTICE_FLIP_V:
+ isOdd = numV & 1;
+ break;
+ case LATTICE_FLIP_W:
+ isOdd = numW & 1;
+ break;
+
+ default:
+ printf("lattice_flip(): Unknown flipping axis (%d)\n", axis);
+ return OPERATOR_CANCELLED;
+ }
+
+ if (isOdd) {
+ BPoint *bp;
+ float avgInv = 1.0f / (float)totP;
+ int i;
+
+ /* midpoint calculation - assuming that u/v/w are axis-aligned */
+ for (i = 0, bp = lt->def; i < totP; i++, bp++) {
+ mid += bp->vec[axis] * avgInv;
+ }
+ }
+
+ /* Second Pass: swap pairs of vertices per axis, assuming they are all sorted */
+ switch (axis) {
+ case LATTICE_FLIP_U:
+ {
+ int u, v, w;
+
+ /* v/w strips - front to back, top to bottom */
+ for (w = 0; w < numW; w++) {
+ for (v = 0; v < numV; v++) {
+ /* swap coordinates of pairs of vertices on u */
+ for (u = 0; u < (numU / 2); u++) {
+ lattice_swap_point_pairs(lt, u, v, w, mid, axis);
+ }
+
+ /* flip u-coordinate of midpoint (i.e. unpaired point on u) */
+ if (isOdd) {
+ u = (numU / 2);
+ lattice_flip_point_value(lt, u, v, w, mid, axis);
+ }
+ }
+ }
+ }
+ break;
+ case LATTICE_FLIP_V:
+ {
+ int u, v, w;
+
+ /* u/w strips - front to back, left to right */
+ for (w = 0; w < numW; w++) {
+ for (u = 0; u < numU; u++) {
+ /* swap coordinates of pairs of vertices on v */
+ for (v = 0; v < (numV / 2); v++) {
+ lattice_swap_point_pairs(lt, u, v, w, mid, axis);
+ }
+
+ /* flip v-coordinate of midpoint (i.e. unpaired point on v) */
+ if (isOdd) {
+ v = (numV / 2);
+ lattice_flip_point_value(lt, u, v, w, mid, axis);
+ }
+ }
+ }
+ }
+ break;
+ case LATTICE_FLIP_W:
+ {
+ int u, v, w;
+
+ for (v = 0; v < numV; v++) {
+ for (u = 0; u < numU; u++) {
+ /* swap coordinates of pairs of vertices on w */
+ for (w = 0; w < (numW / 2); w++) {
+ lattice_swap_point_pairs(lt, u, v, w, mid, axis);
+ }
+
+ /* flip w-coordinate of midpoint (i.e. unpaired point on w) */
+ if (isOdd) {
+ w = (numW / 2);
+ lattice_flip_point_value(lt, u, v, w, mid, axis);
+ }
+ }
+ }
+ }
+ break;
+
+ default: /* shouldn't happen, but just in case */
+ break;
+ }
+
+ /* updates */
+ DAG_id_tag_update(&obedit->id, OB_RECALC_DATA);
+ WM_event_add_notifier(C, NC_GEOM | ND_DATA, obedit->data);
+
+ return OPERATOR_FINISHED;
+}
+
+void LATTICE_OT_flip(wmOperatorType *ot)
+{
+ static EnumPropertyItem flip_items[] = {
+ {LATTICE_FLIP_U, "U", 0, "U (X) Axis", ""},
+ {LATTICE_FLIP_V, "V", 0, "V (Y) Axis", ""},
+ {LATTICE_FLIP_W, "W", 0, "W (Z) Axis", ""},
+ {0, NULL, 0, NULL, NULL}};
+
+ /* identifiers */
+ ot->name = "Flip (Distortion Free)";
+ ot->description = "Mirror all control points without inverting the lattice deform";
+ ot->idname = "LATTICE_OT_flip";
+
+ /* api callbacks */
+ ot->poll = ED_operator_editlattice;
+ ot->invoke = WM_menu_invoke;
+ ot->exec = lattice_flip_exec;
+
+ /* flags */
+ ot->flag = OPTYPE_REGISTER | OPTYPE_UNDO;
+
+ /* properties */
+ ot->prop = RNA_def_enum(ot->srna, "axis", flip_items, LATTICE_FLIP_U, "Flip Axis", "Coordinates along this axis get flipped");
+}
+
/****************************** Mouse Selection *************************/
-static void findnearestLattvert__doClosest(void *userData, BPoint *bp, int x, int y)
+static void findnearestLattvert__doClosest(void *userData, BPoint *bp, const float screen_co[2])
{
- struct { BPoint *bp; short dist, select; int mval[2]; } *data = userData;
- float temp = abs(data->mval[0] - x) + abs(data->mval[1] - y);
+ struct { BPoint *bp; float dist; int select; float mval_fl[2]; } *data = userData;
+ float dist_test = len_manhattan_v2v2(data->mval_fl, screen_co);
- if ((bp->f1 & SELECT) == data->select)
- temp += 5;
+ if ((bp->f1 & SELECT) && data->select)
+ dist_test += 5.0f;
- if (temp < data->dist) {
- data->dist = temp;
+ if (dist_test < data->dist) {
+ data->dist = dist_test;
data->bp = bp;
}
@@ -319,18 +569,18 @@ static void findnearestLattvert__doClosest(void *userData, BPoint *bp, int x, in
static BPoint *findnearestLattvert(ViewContext *vc, const int mval[2], int sel)
{
- /* sel==1: selected gets a disadvantage */
+ /* (sel == 1): selected gets a disadvantage */
/* in nurb and bezt or bp the nearest is written */
/* return 0 1 2: handlepunt */
- struct { BPoint *bp; short dist, select; int mval[2]; } data = {NULL};
+ struct { BPoint *bp; float dist; int select; float mval_fl[2]; } data = {NULL};
data.dist = 100;
data.select = sel;
- data.mval[0] = mval[0];
- data.mval[1] = mval[1];
+ data.mval_fl[0] = mval[0];
+ data.mval_fl[1] = mval[1];
ED_view3d_init_mats_rv3d(vc->obedit, vc->rv3d);
- lattice_foreachScreenVert(vc, findnearestLattvert__doClosest, &data);
+ lattice_foreachScreenVert(vc, findnearestLattvert__doClosest, &data, V3D_PROJ_TEST_CLIP_DEFAULT);
return data.bp;
}
@@ -341,7 +591,7 @@ int mouse_lattice(bContext *C, const int mval[2], int extend, int deselect, int
BPoint *bp = NULL;
view3d_set_viewcontext(C, &vc);
- bp = findnearestLattvert(&vc, mval, 1);
+ bp = findnearestLattvert(&vc, mval, TRUE);
if (bp) {
if (extend) {