Add base case for wg size

Fix bug that was causing infinite loop in helper func

Author: hdelan

source/ur/ur.hpp

@@ -329,7 +329,8 @@ template <typename T> class Result {
 static inline void
 roundToHighestFactorOfGlobalSize(size_t &ThreadsPerBlockInDim,
                                  const size_t GlobalWorkSizeInDim) {
-  while (GlobalWorkSizeInDim % ThreadsPerBlockInDim) {
+  while (ThreadsPerBlockInDim > 1 &&
+         GlobalWorkSizeInDim % ThreadsPerBlockInDim) {
     --ThreadsPerBlockInDim;
   }
 }
@@ -359,8 +360,10 @@ static inline void roundToHighestFactorOfGlobalSizeIn3d(
 
   ThreadsPerBlock[0] = std::min(
       GlobalSize[0], MaxBlockSize / (ThreadsPerBlock[1] * ThreadsPerBlock[2]));
+
   // Make the X dim a factor of 2
   do {
     roundToHighestFactorOfGlobalSize(ThreadsPerBlock[0], GlobalSize[0]);
-  } while (!isPowerOf2(ThreadsPerBlock[0]));
+  } while (!isPowerOf2(ThreadsPerBlock[0]) && ThreadsPerBlock[0] > 32 &&
+           --ThreadsPerBlock[0]);
 }