[Bf-blender-cvs] [a055395] master: Dynapaint: Serious cleanup and optimization of dynamicPaint_createUVSurface (mostly).
Bastien Montagne
noreply at git.blender.org
Fri May 20 23:18:31 CEST 2016
Commit: a055395487a338ab5e4fe2b749df80cc87da97c5
Author: Bastien Montagne
Date: Fri May 20 23:16:52 2016 +0200
Branches: master
https://developer.blender.org/rBa055395487a338ab5e4fe2b749df80cc87da97c5
Dynapaint: Serious cleanup and optimization of dynamicPaint_createUVSurface (mostly).
Gives 3-4% speedup in pre-bake step (from 112 to 108 seconds with own heavy test file).
Note that here we have a huge potential performance boost if we replace the flat
`Bounds2D *faceBB` array of UV tris bounding boxes by a real 2D AABB tree (right now,
we can loop over all UV triangles times the number of pixels of the surface times 5...).
===================================================================
M source/blender/blenkernel/intern/dynamicpaint.c
===================================================================
diff --git a/source/blender/blenkernel/intern/dynamicpaint.c b/source/blender/blenkernel/intern/dynamicpaint.c
index 9e384eb..daf1e86 100644
--- a/source/blender/blenkernel/intern/dynamicpaint.c
+++ b/source/blender/blenkernel/intern/dynamicpaint.c
@@ -529,7 +529,7 @@ static int surface_getBrushFlags(DynamicPaintSurface *surface, const Scene *scen
return flags;
}
-static int brush_usesMaterial(DynamicPaintBrushSettings *brush, Scene *scene)
+static int brush_usesMaterial(const DynamicPaintBrushSettings *brush, const Scene *scene)
{
return ((brush->flags & MOD_DPAINT_USE_MATERIAL) && (!BKE_scene_use_new_shading_nodes(scene)));
}
@@ -2174,11 +2174,9 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
0.4f, -0.2f,
-0.4f, 0.3f,
};
- int ty;
- int w, h;
- int tottri;
+ const int aa_samples = (surface->flags & MOD_DPAINT_ANTIALIAS) ? 5 : 1;
char uvname[MAX_CUSTOMDATA_LAYER_NAME];
- int active_points = 0;
+ uint32_t active_points = 0;
bool error = false;
PaintSurfaceData *sData;
@@ -2193,7 +2191,6 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
Bounds2D *faceBB = NULL;
int *final_index;
- int aa_samples;
if (!dm)
return setError(canvas, N_("Canvas mesh not updated"));
@@ -2202,7 +2199,7 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
mloop = dm->getLoopArray(dm);
mlooptri = dm->getLoopTriArray(dm);
- tottri = dm->getNumLoopTri(dm);
+ const int tottri = dm->getNumLoopTri(dm);
/* get uv map */
if (CustomData_has_layer(&dm->loopData, CD_MLOOPUV)) {
@@ -2216,7 +2213,8 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
if (surface->image_resolution < 16 || surface->image_resolution > 8192)
return setError(canvas, N_("Invalid resolution"));
- w = h = surface->image_resolution;
+ const int w = surface->image_resolution;
+ const int h = w;
/*
* Start generating the surface
@@ -2230,16 +2228,15 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
if (!surface->data)
return setError(canvas, N_("Not enough free memory"));
- aa_samples = (surface->flags & MOD_DPAINT_ANTIALIAS) ? 5 : 1;
- tempPoints = (struct PaintUVPoint *) MEM_callocN(w * h * sizeof(struct PaintUVPoint), "Temp PaintUVPoint");
+ tempPoints = MEM_callocN(w * h * sizeof(*tempPoints), "Temp PaintUVPoint");
if (!tempPoints)
error = true;
- final_index = (int *) MEM_callocN(w * h * sizeof(int), "Temp UV Final Indexes");
+ final_index = MEM_callocN(w * h * sizeof(*final_index), "Temp UV Final Indexes");
if (!final_index)
error = true;
- tempWeights = (struct Vec3f *) MEM_mallocN(w * h * aa_samples * sizeof(struct Vec3f), "Temp bWeights");
+ tempWeights = MEM_mallocN(w * h * aa_samples * sizeof(*tempWeights), "Temp bWeights");
if (!tempWeights)
error = true;
@@ -2248,31 +2245,27 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
* the pixel-inside-a-face search.
*/
if (!error) {
- faceBB = (struct Bounds2D *) MEM_mallocN(tottri * sizeof(struct Bounds2D), "MPCanvasFaceBB");
+ faceBB = MEM_mallocN(tottri * sizeof(*faceBB), "MPCanvasFaceBB");
if (!faceBB)
error = true;
}
if (!error) {
- for (ty = 0; ty < tottri; ty++) {
- int i;
-
- copy_v2_v2(faceBB[ty].min, mloopuv[mlooptri[ty].tri[0]].uv);
- copy_v2_v2(faceBB[ty].max, mloopuv[mlooptri[ty].tri[0]].uv);
+ for (int i = 0; i < tottri; i++) {
+ copy_v2_v2(faceBB[i].min, mloopuv[mlooptri[i].tri[0]].uv);
+ copy_v2_v2(faceBB[i].max, mloopuv[mlooptri[i].tri[0]].uv);
- for (i = 1; i < 3; i++) {
- minmax_v2v2_v2(faceBB[ty].min, faceBB[ty].max, mloopuv[mlooptri[ty].tri[i]].uv);
+ for (int j = 1; j < 3; j++) {
+ minmax_v2v2_v2(faceBB[i].min, faceBB[i].max, mloopuv[mlooptri[i].tri[j]].uv);
}
}
/* Loop through every pixel and check if pixel is uv-mapped on a canvas face. */
#pragma omp parallel for schedule(static)
- for (ty = 0; ty < h; ty++) {
- int tx;
- for (tx = 0; tx < w; tx++) {
- int i, sample;
- int index = tx + w * ty;
- PaintUVPoint *tPoint = (&tempPoints[index]);
+ for (int ty = 0; ty < h; ty++) {
+ for (int tx = 0; tx < w; tx++) {
+ const int index = tx + w * ty;
+ PaintUVPoint *tPoint = &tempPoints[index];
float point[5][2];
/* Init per pixel settings */
@@ -2302,14 +2295,15 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
/* Loop through samples, starting from middle point */
- for (sample = 0; sample < 5; sample++) {
+ for (int sample = 0; sample < 5; sample++) {
/* Loop through every face in the mesh */
- for (i = 0; i < tottri; i++) {
+ /* XXX TODO This is *horrible* with big meshes, should use a 2D BVHTree over UV tris here! */
+ for (int i = 0; i < tottri; i++) {
/* Check uv bb */
- if ((faceBB[i].min[0] > (point[sample][0])) ||
- (faceBB[i].min[1] > (point[sample][1])) ||
- (faceBB[i].max[0] < (point[sample][0])) ||
- (faceBB[i].max[1] < (point[sample][1])))
+ if ((faceBB[i].min[0] > point[sample][0]) ||
+ (faceBB[i].min[1] > point[sample][1]) ||
+ (faceBB[i].max[0] < point[sample][0]) ||
+ (faceBB[i].max[1] < point[sample][1]))
{
continue;
}
@@ -2353,42 +2347,51 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
* (To avoid seams on uv island edges)
*/
#pragma omp parallel for schedule(static)
- for (ty = 0; ty < h; ty++) {
- int tx;
- for (tx = 0; tx < w; tx++) {
- int index = tx + w * ty;
- PaintUVPoint *tPoint = (&tempPoints[index]);
+ for (int ty = 0; ty < h; ty++) {
+ for (int tx = 0; tx < w; tx++) {
+ const int index = tx + w * ty;
+ PaintUVPoint *tPoint = &tempPoints[index];
- /* If point isn't't on canvas mesh */
+ /* If point isn't on canvas mesh */
if (tPoint->tri_index == -1) {
- int u_min, u_max, v_min, v_max;
- int u, v, ind;
float point[2];
/* get loop area */
- u_min = (tx > 0) ? -1 : 0;
- u_max = (tx < (w - 1)) ? 1 : 0;
- v_min = (ty > 0) ? -1 : 0;
- v_max = (ty < (h - 1)) ? 1 : 0;
+ const int u_min = (tx > 0) ? -1 : 0;
+ const int u_max = (tx < (w - 1)) ? 1 : 0;
+ const int v_min = (ty > 0) ? -1 : 0;
+ const int v_max = (ty < (h - 1)) ? 1 : 0;
point[0] = ((float)tx + 0.5f) / w;
point[1] = ((float)ty + 0.5f) / h;
/* search through defined area for neighbor */
- for (u = u_min; u <= u_max; u++)
- for (v = v_min; v <= v_max; v++) {
+ for (int u = u_min; u <= u_max; u++) {
+ for (int v = v_min; v <= v_max; v++) {
/* if not this pixel itself */
if (u != 0 || v != 0) {
- ind = (tx + u) + w * (ty + v);
+ const int ind = (tx + u) + w * (ty + v);
- /* if neighbor has index */
- if (tempPoints[ind].tri_index != -1) {
+ /* if neighbor has index */
+ if (tempPoints[ind].neighbour_pixel == -1 && tempPoints[ind].tri_index != -1) {
float uv[2];
const int i = tempPoints[ind].tri_index;
const float *uv1 = mloopuv[mlooptri[i].tri[0]].uv;
const float *uv2 = mloopuv[mlooptri[i].tri[1]].uv;
const float *uv3 = mloopuv[mlooptri[i].tri[2]].uv;
+ /* tri index */
+ /* There is a low possibility of actually having a neighbor point which tri is
+ * already set from another neighbor in a separate thread here.
+ * Cheking for both tri_index and neighbour_pixel above reduces that probability
+ * but it remains possible.
+ * That atomic op (and its memory fence) ensures tPoint->neighbour_pixel is set
+ * to non--1 *before* its tri_index is set (i.e. that it cannot be used a neighbour).
+ */
+ tPoint->neighbour_pixel = ind - 1;
+ atomic_add_uint32(&tPoint->neighbour_pixel, 1);
+ tPoint->tri_index = i;
+
/* Now calculate pixel data for this pixel as it was on polygon surface */
/* Add b-weights per anti-aliasing sample */
for (int j = 0; j < aa_samples; j++) {
@@ -2397,9 +2400,6 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
barycentric_weights_v2(uv1, uv2, uv3, uv, tempWeights[index * aa_samples + j].v);
}
- /* Set values */
- tPoint->neighbour_pixel = ind; /* tri index */
-
/* save vertex indexes */
tPoint->v1 = mloop[mlooptri[i].tri[0]].v;
tPoint->v2 = mloop[mlooptri[i].tri[1]].v;
@@ -2410,24 +2410,12 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
}
}
}
+ }
}
- }
- }
- /*
- * When base loop is over convert found neighbor indexes to real ones
- * Also count the final number of active surface points
- */
- for (ty = 0; ty < h; ty++) {
- int tx;
- for (tx = 0; tx < w; tx++) {
- int index = tx + w * ty;
- PaintUVPoint *tPoint = &tempPoints[index];
-
- if (tPoint->tri_index == -1 && tPoint->neighbour_pixel != -1)
- tPoint->tri_index = tempPoints[tPoint->neighbour_pixel].tri_index;
+ /* Increase the final number of active surface points if relevant. */
if (tPoint->tri_index != -1)
- active_points++;
+ atomic_add_uint32(&active_points, 1);
}
}
@@ -2450,20 +2438,17 @@ int dynamicPaint_createUVSurface(Scene *scene, DynamicPaintSurface *surface)
if (sData->adj_data) {
PaintAdjData *ed = sData->adj_data;
unsigned int n_pos = 0;
- for (ty = 0; ty < h; ty++) {
- int tx;
- for (tx = 0; tx < w; tx++) {
- int index = tx + w * ty;
+ for (int ty = 0; ty < h; ty++) {
+ for (int tx = 0; tx < w; tx++) {
+ const int index = tx + w * ty;
if (tempPoints[index].tri_index != -1) {
ed->n_index[final_index[index]] = n_pos;
ed->n_num[final_index[index]] = 0;
for (int i = 0; i < 8; i++) {
-
- /* Try to find a neighboring pixel in defined direction
- * If not found, -1 is returned */
- int n_target = dynamicPaint_findNeighbourPixel(tempPoints, dm, uvname, w, h,
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list