[Bf-blender-cvs] SVN commit: /data/svn/bf-blender [59203] branches/ soc-2013-meshdata_transfer/source/blender/bmesh/tools/bmesh_data_transfer.c : multi layer UV transfer: making the l_grp be accessed per_vertex instead of per_mesh which speeds up the UV transfer in an obvious way ...
Walid Shouman
eng.walidshouman at gmail.com
Sat Aug 17 10:55:41 CEST 2013
Revision: 59203
http://projects.blender.org/scm/viewvc.php?view=rev&root=bf-blender&revision=59203
Author: walid
Date: 2013-08-17 08:55:41 +0000 (Sat, 17 Aug 2013)
Log Message:
-----------
multi layer UV transfer: making the l_grp be accessed per_vertex instead of per_mesh which speeds up the UV transfer in an obvious way ... speed difference could be easily noticed compared to transfering a single layer which isn't updated yet
Modified Paths:
--------------
branches/soc-2013-meshdata_transfer/source/blender/bmesh/tools/bmesh_data_transfer.c
Modified: branches/soc-2013-meshdata_transfer/source/blender/bmesh/tools/bmesh_data_transfer.c
===================================================================
--- branches/soc-2013-meshdata_transfer/source/blender/bmesh/tools/bmesh_data_transfer.c 2013-08-17 08:21:40 UTC (rev 59202)
+++ branches/soc-2013-meshdata_transfer/source/blender/bmesh/tools/bmesh_data_transfer.c 2013-08-17 08:55:41 UTC (rev 59203)
@@ -1450,6 +1450,11 @@
int count;
} BM_loop_pool;
+typedef struct BM_loop_group_pool {
+ struct BM_loop_pool *l_grp;
+ int count;
+} BM_loop_group_pool;
+
bool BM_loop_in_loops(BMLoop **l_grp, int len, BMLoop *l)
{
int i;
@@ -2333,7 +2338,7 @@
case CD_MLOOPUV:
{
- int c, d, g, h, i, l_dst_iter;
+ int c, d, g, h, i;
// int const exp_loop_per_face = 4;
// int const exp_tolerance = 2; //expectation tolerance of 2 would allocate double the ideal memory for each face
@@ -2341,9 +2346,11 @@
// int const dst_src_faces_ratio = ceil((float)bm_dst->totface /(float)bm_src->totface); //we should ensure that
//bm_src->totface != 0
// int const exp_dst_loops_per_src_face = dst_src_faces_ratio * exp_loop_per_face * exp_tolerance;
- int const exp_loop_per_vert_double = 10; //it should be 8 for a vertex surrounded by 4 faces (common scenario)
+ int const exp_loop_per_vert = 4 + 1; //4 => the exp nmbr, 1 => tolerance
+ int const exp_loop_per_vert_double = exp_loop_per_vert * 2; //it should be 8 for a vertex surrounded by 4 faces (common scenario)
//as the time penalty in realloc is considered severe, we're adding a
//tolerence of 5 faces per vertex ... more than that we would realloc
+ int const exp_vert_splits = exp_loop_per_vert;
BMFace *f_src;
BMFace **f_src_table = MEM_mallocN(sizeof(*f_src_table) * bm_src->totface, "f_src_table bmesh_data_transfer.c");
@@ -2353,8 +2360,8 @@
BMEdge *e;
BMIter eiter;
// BM_UV_per_face_mapping *fuv_table = MEM_mallocN(sizeof(*fuv_table) * bm_dst->totface, "fuv_table bmesh_data_transfer.c");
- BM_loop_pool *l_grp = MEM_mallocN(sizeof(*l_grp) * bm_dst->totloop, "l_grp bmesh_data_transfer.c");
- int l_grp_count; //count of actual loop groups
+ //store the loop groups per vertex ... for a single island mesh each vertex would have a single l_grp
+ BM_loop_group_pool *v_l_grp = MEM_mallocN(sizeof(*v_l_grp) * bm_dst->totvert, "v_l_grp bmesh_data_transfer.c");
float (*uv_buffer)[2] = MEM_mallocN(sizeof(*uv_buffer) * exp_loop_per_vert_double, "uv_buffer bmesh_data_transfer.c");
float mid_uv[2];
float weight_accu[2];
@@ -2362,17 +2369,20 @@
//Alloc loops
- l_dst_iter = 0;
- BM_ITER_MESH (f_dst, &fiter, bm_dst, BM_FACES_OF_MESH) {
-// BM_ITER_MESH_INDEX (f_dst, &fiter, bm_dst, BM_FACES_OF_MESH, b) {
-// fuv_table[b].uv = MEM_mallocN(sizeof(*(fuv_table->uv)) * (f_dst->len), "fuv_table->uv bmesh_data_transfer.c");
-// fuv_table[b].f = f_dst;
- BM_ITER_ELEM (l, &liter, f_dst, BM_LOOPS_OF_FACE) {
- l_grp[l_dst_iter].l = MEM_mallocN(sizeof(*(l_grp->l)) * exp_loop_per_vert_double, "l_grp[].l bmesh_data_transfer");
+ ///would it be faster/better if we used BM_ITER_MESH_INDEX instead of v->head.index?
+ ///no need to recalculate the sizeof and multiplication everytime!!
+ //the upcoming allocation is so protective ... as the exp_vert_splits assumes having a totally disconnected
+ //vertex and the exp_loop_per_vert_double assumes having a totally connected vertex
+ //this is considered a cost for minimilazing the usage of realloc
+ BM_ITER_MESH (v, &iter, bm_dst, BM_VERTS_OF_MESH) {
+ v_l_grp[v->head.index].l_grp = MEM_mallocN(sizeof(*(v_l_grp[v->head.index].l_grp)) * exp_vert_splits, "v_l_grp[].l_grp bmesh_data_transfer.c");
- l_dst_iter++;
+ for (i = 0; i < exp_vert_splits; i++) {
+ v_l_grp[v->head.index].l_grp[i].l = MEM_mallocN(sizeof(*(v_l_grp[v->head.index].l_grp[i].l)) * exp_loop_per_vert_double, "v_l_grp[].l_grp.l bmesh_data_transfer.c");
}
+
+ v_l_grp[v->head.index].count = 0;
}
BM_ITER_MESH (f_src, &fiter, bm_src, BM_FACES_OF_MESH) {
@@ -2393,8 +2403,6 @@
CD_src = CustomData_get_n_offset(&bm_src->ldata, CD_MLOOPUV, src_lay_iter); //get the offset of the
CD_dst = CustomData_get_n_offset(&bm_dst->ldata, CD_MLOOPUV, dst_lay_iter); //lay_iter(th)CD_SHAPEKEY layer
- l_grp_count = 0;
-// f_src_count = 0;
b = 0;
//the way we do it is by looping over each face!!
BM_ITER_MESH (f_dst, &fiter, bm_dst, BM_FACES_OF_MESH) {
@@ -2442,10 +2450,16 @@
//we finally found loops that shall be averaged from different faces!
//we now shall average them into a buffer!
+ //----start rewrite----
+ int v_ind = fl_table[f_src->head.index].l[d]->v->head.index;
+
+ int *l_grp_count = &v_l_grp[v_ind].count;
//we've got 2 loops ... search for them in the loop groups ... if found
//add a new entry to l_grp and increment l_grp_count
//else append the other loop
- for (h = 0; h < l_grp_count; h++) {
+ for (h = 0; h < *l_grp_count; h++) {
+ BM_loop_pool *l_grp = v_l_grp[v_ind].l_grp;
+
if (BM_loop_in_loops(l_grp[h].l, l_grp[h].count, fl_table[f_n->head.index].l[g])) {
//found the neighboring face's loop in the group
if (!BM_loop_in_loops(l_grp[h].l, l_grp[h].count, fl_table[f_src->head.index].l[d])) {
@@ -2465,14 +2479,14 @@
break;
}
- else if (BM_loop_in_loops(l_grp[h].l,l_grp[h].count, fl_table[f_src->head.index].l[d])) {
+ else if (BM_loop_in_loops(l_grp[h].l, l_grp[h].count, fl_table[f_src->head.index].l[d])) {
//found the source face's loop in the group
- if (!BM_loop_in_loops(l_grp[h].l,l_grp[h].count, fl_table[f_n->head.index].l[g])) {
+ if (!BM_loop_in_loops(l_grp[h].l, l_grp[h].count, fl_table[f_n->head.index].l[g])) {
//now reallocate memory for a new loop
(l_grp[h].count)++;
if (l_grp[h].count < exp_loop_per_vert_double) {
- l_grp[h].l = MEM_reallocN(l_grp[h].l, sizeof(*(l_grp->l)) * (l_grp[h].count));
+ l_grp[h].l = MEM_reallocN(l_grp[h].l, sizeof(*(l_grp[h].l)) * (l_grp[h].count));
}
//and append it
l_grp[h].l[l_grp[h].count - 1] = fl_table[f_n->head.index].l[g];
@@ -2482,16 +2496,18 @@
}
}
- if ( h == l_grp_count) {
+ if ( h == *l_grp_count) {
+ BM_loop_pool *l_grp = v_l_grp[v_ind].l_grp;
+
//the loops weren't found in any group
//make a new group entry and append it
- l_grp[l_grp_count].count = 2;
+ l_grp[h].count = 2;
//adding a place for 2 loops
- // l_grp[l_grp_count].l = MEM_mallocN(sizeof(*(l_grp->l)) * 2, "l_grp[].l bmesh_data_transfer");
- l_grp[l_grp_count].l[0] = fl_table[f_src->head.index].l[d];
- l_grp[l_grp_count].l[1] = fl_table[f_n->head.index].l[g];
- l_grp_count++;
+ l_grp[h].l[0] = fl_table[f_src->head.index].l[d];
+ l_grp[h].l[1] = fl_table[f_n->head.index].l[g];
+ (*l_grp_count)++;
}
+ //====end rewrite====
}
}
}
@@ -2515,51 +2531,56 @@
//get the mid of UVs into a buffer
//get the UV coords after looping over the src faces!!
- for (h = 0; h < l_grp_count; h++) {
- //average the loops' uvs
+ BM_ITER_MESH (v, &iter, bm_dst, BM_VERTS_OF_MESH) {
+ int v_ind = v->head.index;
+ int *l_grp_count = &v_l_grp[v_ind].count;
- if (l_grp[h].count > exp_loop_per_vert_double) { //expand the buffer size when needed
- if (l_grp[h].count > l_grp_max_count) {
- uv_buffer = MEM_reallocN(uv_buffer, sizeof(*uv_buffer) * l_grp[h].count);
- l_grp_max_count = l_grp[h].count;
+ for (h = 0; h < *l_grp_count; h++) {
+ BM_loop_pool *l_grp = v_l_grp[v_ind].l_grp;
+ //average the loops' uvs
+
+ if (l_grp[h].count > exp_loop_per_vert_double) { //expand the buffer size when needed
+ if (l_grp[h].count > l_grp_max_count) {
+ uv_buffer = MEM_reallocN(uv_buffer, sizeof(*uv_buffer) * l_grp[h].count);
+ l_grp_max_count = l_grp[h].count;
+ }
}
- }
- //prepare the uvs to be averaged
- for (i = 0; i < l_grp[h].count; i++) {
- //copying each element is really inefficient it -at least- doubles the time when we are already accessing
- //the pointer for read only! .. we should copy the pointer itself
- copy_v2_v2(uv_buffer[i], BM_ELEM_CD_GET_VOID_P(l_grp[h].l[i], CD_dst));
- }
+ //prepare the uvs to be averaged
+ for (i = 0; i < l_grp[h].count; i++) {
+ //copying each element is really inefficient it -at least- doubles the time when we are already accessing
+ //the pointer for read only! .. we should copy the pointer itself
+ copy_v2_v2(uv_buffer[i], BM_ELEM_CD_GET_VOID_P(l_grp[h].l[i], CD_dst));
+ }
- //get the mid value
- zero_v2(mid_uv);
- mid_poly_v2(mid_uv, uv_buffer, l_grp[h].count);
+ //get the mid value
+ zero_v2(mid_uv);
+ mid_poly_v2(mid_uv, uv_buffer, l_grp[h].count);
- //copy the value to each of them
- for (i = 0; i < l_grp[h].count; i++) {
- //commenting this would leave us with the output of interpolation for each face ^_^
- copy_v2_v2(BM_ELEM_CD_GET_VOID_P(l_grp[h].l[i], CD_dst), mid_uv);
+ //copy the value to each of them
+ for (i = 0; i < l_grp[h].count; i++) {
+ //commenting this would leave us with the output of interpolation for each face ^_^
+ copy_v2_v2(BM_ELEM_CD_GET_VOID_P(l_grp[h].l[i], CD_dst), mid_uv);
+ }
+ //this would make the l_grp[h].l hidden for the function; to avoid unnessecery realloc/memsetting
+ l_grp[h].count = 0;
}
- //this would make the l_grp[h].l hidden for the function; to avoid unnessecery realloc/memsetting
- l_grp[h].count = 0;
}
}
//Free loops
- l_dst_iter = 0;
- BM_ITER_MESH_INDEX (f_dst, &fiter, bm_dst, BM_FACES_OF_MESH, b) {
-// MEM_freeN(fuv_table[b].uv);
+ BM_ITER_MESH (v, &iter, bm_dst, BM_VERTS_OF_MESH) {
- BM_ITER_ELEM (l, &liter, f_dst, BM_LOOPS_OF_FACE) {
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list