3 files changed, 15 insertions, 0 deletions
diff --git a/src/routines/level2/xtrsv.cpp b/src/routines/level2/xtrsv.cpp
index c474db4c..5d2ddf54 100644
--- a/src/routines/level2/xtrsv.cpp
+++ b/src/routines/level2/xtrsv.cpp
@@ -87,6 +87,11 @@ void Xtrsv<T>::DoTrsv(const Layout layout, const Triangle triangle,
   // Makes sure all dimensions are larger than zero
   if (n == 0) { throw BLASError(StatusCode::kInvalidDimension); }
 
+  // Some parts of this kernel are not tunable and thus require some minimal OpenCL properties
+  if (device_.MaxWorkGroupSize() < 16) { // minimum of total local work size of 16
+    throw RuntimeErrorCode(StatusCode::kNotImplemented);
+  }
+
   // Tests the matrix and vector
   TestMatrixA(n, n, a_buffer, a_offset, a_ld);
   TestVectorX(n, b_buffer, b_offset, b_inc);
diff --git a/src/routines/level3/xtrsm.cpp b/src/routines/level3/xtrsm.cpp
index 255e1cca..905660ff 100644
--- a/src/routines/level3/xtrsm.cpp
+++ b/src/routines/level3/xtrsm.cpp
@@ -78,6 +78,11 @@ void Xtrsm<T>::TrsmColMajor(const Side side, const Triangle triangle,
   // Makes sure all dimensions are larger than zero
   if ((m == 0) || (n == 0)) { throw BLASError(StatusCode::kInvalidDimension); }
 
+  // Some parts of this kernel are not tunable and thus require some minimal OpenCL properties
+  if (device_.MaxWorkGroupSize() < 16) { // minimum of total local work size of 16
+    throw RuntimeErrorCode(StatusCode::kNotImplemented);
+  }
+
   // Computes the k dimension. This is based on whether or not matrix is A (on the left)
   // or B (on the right) in the Xgemm routine.
   const auto k = (side == Side::kLeft) ? m : n;
diff --git a/src/routines/levelx/xinvert.cpp b/src/routines/levelx/xinvert.cpp
index d851e0b9..65a28d73 100644
--- a/src/routines/levelx/xinvert.cpp
+++ b/src/routines/levelx/xinvert.cpp
@@ -49,6 +49,11 @@ void Xinvert<T>::InvertMatrixDiagonalBlocks(const Layout layout, const Triangle
     throw BLASError(StatusCode::kInvalidDimension);
   }
 
+  // Some parts of this kernel are not tunable and thus require some minimal OpenCL properties
+  if (device_.MaxWorkGroupSize() < 16) { // minimum of total local work size of 16
+    throw RuntimeErrorCode(StatusCode::kNotImplemented);
+  }
+
   // Helper variables
   const auto internal_block_size = static_cast<size_t>(db_["INTERNAL_BLOCK_SIZE"]);
   assert(internal_block_size == 16);