Make 3d projpaint iteration lockless.

While speedup is hard to detect (highly fluctuent), it seems to be
around 5% on average on my 8 threads machine...

It also remove usage of a 'global' thread lock, which is always good.

Note that I also tried to use proper foreach threaded iterator construct
(see D5372), but that proved to be relatively slower (presumably due to
the very high dissymmetry between tasks, usually during a paint stroke
only a few chunks will require most of the computing effort, overhead of
threaded foreach management is then noticeable).

This concludes (for now) the work on
T51133 Bad performance with texture painting depending on multi-thread settings.

2 files changed, 19 insertions, 29 deletions

diff --git a/source/blender/editors/sculpt_paint/CMakeLists.txt b/source/blender/editors/sculpt_paint/CMakeLists.txt
index 23617e687ea..2a8ff9d4f78 100644
--- a/source/blender/editors/sculpt_paint/CMakeLists.txt
+++ b/source/blender/editors/sculpt_paint/CMakeLists.txt
@@ -29,6 +29,7 @@ set(INC
   ../../makesrna
   ../../render/extern/include
   ../../windowmanager
+  ../../../../intern/atomic
   ../../../../intern/glew-mx
   ../../../../intern/guardedalloc
 )
diff --git a/source/blender/editors/sculpt_paint/paint_image_proj.c b/source/blender/editors/sculpt_paint/paint_image_proj.c
index 1220b071a17..f768a66a41c 100644
--- a/source/blender/editors/sculpt_paint/paint_image_proj.c
+++ b/source/blender/editors/sculpt_paint/paint_image_proj.c
@@ -44,6 +44,8 @@
 #include "BLI_threads.h"
 #include "BLI_utildefines.h"
 
+#include "atomic_ops.h"
+
 #include "BLT_translation.h"
 
 #include "IMB_imbuf.h"
@@ -366,7 +368,7 @@ typedef struct ProjPaintState {
   int bucketMin[2];
   int bucketMax[2];
   /** must lock threads while accessing these. */
-  int context_bucket_x, context_bucket_y;
+  int context_bucket_index;
 
   struct CurveMapping *cavity_curve;
   BlurKernel *blurkernel;
@@ -4712,11 +4714,8 @@ static bool project_bucket_iter_init(ProjPaintState *ps, const float mval_f[2])
 
     /* mouse outside the model areas? */
     if (ps->bucketMin[0] == ps->bucketMax[0] || ps->bucketMin[1] == ps->bucketMax[1]) {
-      return 0;
+      return false;
     }
-
-    ps->context_bucket_x = ps->bucketMin[0];
-    ps->context_bucket_y = ps->bucketMin[1];
   }
   else { /* reproject: PROJ_SRC_* */
     ps->bucketMin[0] = 0;
@@ -4724,11 +4723,10 @@ static bool project_bucket_iter_init(ProjPaintState *ps, const float mval_f[2])
 
     ps->bucketMax[0] = ps->buckets_x;
     ps->bucketMax[1] = ps->buckets_y;
-
-    ps->context_bucket_x = 0;
-    ps->context_bucket_y = 0;
   }
-  return 1;
+
+  ps->context_bucket_index = ps->bucketMin[0] + ps->bucketMin[1] * ps->buckets_x;
+  return true;
 }
 
 static bool project_bucket_iter_next(ProjPaintState *ps,
@@ -4738,37 +4736,28 @@ static bool project_bucket_iter_next(ProjPaintState *ps,
 {
   const int diameter = 2 * ps->brush_size;
 
-  if (ps->thread_tot > 1) {
-    BLI_thread_lock(LOCK_CUSTOM1);
-  }
-
-  // printf("%d %d\n", ps->context_bucket_x, ps->context_bucket_y);
+  const int max_bucket_idx = ps->bucketMax[0] + (ps->bucketMax[1] - 1) * ps->buckets_x;
 
-  for (; ps->context_bucket_y < ps->bucketMax[1]; ps->context_bucket_y++) {
-    for (; ps->context_bucket_x < ps->bucketMax[0]; ps->context_bucket_x++) {
+  for (int bidx = atomic_fetch_and_add_int32(&ps->context_bucket_index, 1); bidx < max_bucket_idx;
+       bidx = atomic_fetch_and_add_int32(&ps->context_bucket_index, 1)) {
+    const int bucket_y = bidx / ps->buckets_x;
+    const int bucket_x = bidx - (bucket_y * ps->buckets_x);
 
+    BLI_assert(bucket_y >= ps->bucketMin[1] && bucket_y < ps->bucketMax[1]);
+    if (bucket_x >= ps->bucketMin[0] && bucket_x < ps->bucketMax[0]) {
       /* use bucket_bounds for project_bucket_isect_circle and project_bucket_init*/
-      project_bucket_bounds(ps, ps->context_bucket_x, ps->context_bucket_y, bucket_bounds);
+      project_bucket_bounds(ps, bucket_x, bucket_y, bucket_bounds);
 
       if ((ps->source != PROJ_SRC_VIEW) ||
           project_bucket_isect_circle(mval, (float)(diameter * diameter), bucket_bounds)) {
-        *bucket_index = ps->context_bucket_x + (ps->context_bucket_y * ps->buckets_x);
-        ps->context_bucket_x++;
-
-        if (ps->thread_tot > 1) {
-          BLI_thread_unlock(LOCK_CUSTOM1);
-        }
+        *bucket_index = bidx;
 
-        return 1;
+        return true;
       }
     }
-    ps->context_bucket_x = ps->bucketMin[0];
   }
 
-  if (ps->thread_tot > 1) {
-    BLI_thread_unlock(LOCK_CUSTOM1);
-  }
-  return 0;
+  return false;
 }
 
 /* Each thread gets one of these, also used as an argument to pass to project_paint_op */