bugfixes for parallel reduction primitive

 - wrong cache size
 - illegal read from cache on overflow -> read from gloabal mem

2 files changed, 3 insertions, 3 deletions

diff --git a/src/com/jogamp/opencl/util/pp/Reduction.java b/src/com/jogamp/opencl/util/pp/Reduction.java
index c2d47f7c..77a37e78 100644
--- a/src/com/jogamp/opencl/util/pp/Reduction.java
+++ b/src/com/jogamp/opencl/util/pp/Reduction.java
@@ -122,7 +122,7 @@ public class Reduction<B extends Buffer> implements CLResource {
         int workItems = CLUtil.roundUp(realSize, groupSize*2) / 2;
 
         int groups = workItems / groupSize;
-        int sharedBufferSize = groupSize / 2 * ELEMENT.SIZE*VECTOR_SIZE;
+        int sharedBufferSize = groupSize * ELEMENT.SIZE*VECTOR_SIZE;
 
         int outputSize = groups * ELEMENT.SIZE*VECTOR_SIZE;
 
diff --git a/src/com/jogamp/opencl/util/pp/reduce.cl b/src/com/jogamp/opencl/util/pp/reduce.cl
index d820ffe2..5f124d62 100644
--- a/src/com/jogamp/opencl/util/pp/reduce.cl
+++ b/src/com/jogamp/opencl/util/pp/reduce.cl
@@ -50,9 +50,9 @@ kernel void reduce(const global TYPE* input, global TYPE* output, local TYPE* sh
     #elif OP_MUL
         shared[localID] = 1;
     #elif OP_MIN
-        shared[localID] = shared[0];
+        shared[localID] = input[0];
     #elif OP_MAX
-        shared[localID] = shared[0];
+        shared[localID] = input[0];
     #endif
     }