Fixed thread synchronization bug

a5e3869f · Peter Eastman · b1a22214 · a5e3869f
Commit a5e3869f authored Jan 27, 2010 by Peter Eastman
Show whitespace changes
Inline Side-by-side

Showing with 4 additions and 2 deletions

platforms/opencl/src/kernels/sort.cl platforms/opencl/src/kernels/sort.cl +4 -2

No files found.
--- a/platforms/opencl/src/kernels/sort.cl
+++ b/platforms/opencl/src/kernels/sort.cl
@@ -23,6 +23,7 @@ __kernel void computeRange(__global TYPE* data, int length, __global float2* ran
    // Now reduce them.

    buffer[get_local_id(0)] = minimum;
+    barrier(CLK_LOCAL_MEM_FENCE);
    for (int step = 1; step < get_local_size(0); step *= 2) {
        if (get_local_id(0)+step < get_local_size(0) && get_local_id(0)%(2*step) == 0)
            buffer[get_local_id(0)] = min(buffer[get_local_id(0)], buffer[get_local_id(0)+step]);
@@ -30,6 +31,7 @@ __kernel void computeRange(__global TYPE* data, int length, __global float2* ran
    }
    minimum = buffer[0];
    buffer[get_local_id(0)] = maximum;
+    barrier(CLK_LOCAL_MEM_FENCE);
    for (int step = 1; step < get_local_size(0); step *= 2) {
        if (get_local_id(0)+step < get_local_size(0) && get_local_id(0)%(2*step) == 0)
            buffer[get_local_id(0)] = max(buffer[get_local_id(0)], buffer[get_local_id(0)+step]);