From 49eb490ee1c60ed254eb1c501e18ebd9948e0ab5 Mon Sep 17 00:00:00 2001
From: Cedric Nugteren <web@cedricnugteren.nl>
Date: Mon, 17 Feb 2020 22:07:51 +0100
Subject: Catches all exceptions of the tuners

---
 CHANGELOG                             |  1 +
 src/tuning/kernels/copy_fast.cpp      | 20 +++++++++++---------
 src/tuning/kernels/copy_pad.cpp       | 20 +++++++++++---------
 src/tuning/kernels/invert.cpp         | 20 +++++++++++---------
 src/tuning/kernels/transpose_fast.cpp | 20 +++++++++++---------
 src/tuning/kernels/transpose_pad.cpp  | 20 +++++++++++---------
 src/tuning/kernels/xaxpy.cpp          | 20 +++++++++++---------
 src/tuning/kernels/xconvgemm.cpp      |  6 ++++--
 src/tuning/kernels/xdot.cpp           |  8 +++++---
 src/tuning/kernels/xgemm.cpp          | 20 +++++++++++---------
 src/tuning/kernels/xgemm_direct.cpp   |  8 +++++---
 src/tuning/kernels/xgemv.cpp          | 10 ++++++----
 src/tuning/kernels/xger.cpp           | 20 +++++++++++---------
 src/tuning/routines/xgemm.cpp         | 20 +++++++++++---------
 src/tuning/routines/xtrsv.cpp         | 18 ++++++++++--------
 15 files changed, 130 insertions(+), 101 deletions(-)

diff --git a/CHANGELOG b/CHANGELOG
index 7a468e7c..ac51ac58 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -1,5 +1,6 @@
 Development (next version)
 - Implemented single-kernel version of convolution as GEMM
+- Now catches all exceptions thrown by the tuners
 - Fixed a bug in ISAMIN kernel
 - Fixed an out-of-bounds read/write in the XHAD routine (thanks to etomzak)
 - Various minor fixes and enhancements
diff --git a/src/tuning/kernels/copy_fast.cpp b/src/tuning/kernels/copy_fast.cpp
index 13f7ef3c..44860f6e 100644
--- a/src/tuning/kernels/copy_fast.cpp
+++ b/src/tuning/kernels/copy_fast.cpp
@@ -20,15 +20,17 @@ using double2 = clblast::double2;
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
-  switch(clblast::GetPrecision(command_line_args)) {
-    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<half>, clblast::CopyTestValidArguments<half>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<half>, clblast::CopySetArguments<half>); break;
-    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<float>, clblast::CopyTestValidArguments<float>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<float>, clblast::CopySetArguments<float>); break;
-    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<double>, clblast::CopyTestValidArguments<double>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<double>, clblast::CopySetArguments<double>); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<float2>, clblast::CopyTestValidArguments<float2>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<float2>, clblast::CopySetArguments<float2>); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<double2>, clblast::CopyTestValidArguments<double2>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<double2>, clblast::CopySetArguments<double2>); break;
-  }
-  return 0;
+  try {
+    const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+    switch(clblast::GetPrecision(command_line_args)) {
+      case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<half>, clblast::CopyTestValidArguments<half>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<half>, clblast::CopySetArguments<half>); break;
+      case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<float>, clblast::CopyTestValidArguments<float>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<float>, clblast::CopySetArguments<float>); break;
+      case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<double>, clblast::CopyTestValidArguments<double>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<double>, clblast::CopySetArguments<double>); break;
+      case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<float2>, clblast::CopyTestValidArguments<float2>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<float2>, clblast::CopySetArguments<float2>); break;
+      case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<double2>, clblast::CopyTestValidArguments<double2>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<double2>, clblast::CopySetArguments<double2>); break;
+    }
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/copy_pad.cpp b/src/tuning/kernels/copy_pad.cpp
index ffaed6ed..589900a9 100644
--- a/src/tuning/kernels/copy_pad.cpp
+++ b/src/tuning/kernels/copy_pad.cpp
@@ -20,15 +20,17 @@ using double2 = clblast::double2;
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
-  switch(clblast::GetPrecision(command_line_args)) {
-    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::PadGetTunerDefaults, clblast::PadGetTunerSettings<half>, clblast::PadTestValidArguments<half>, clblast::PadSetConstraints, clblast::PadComputeLocalMemSize<half>, clblast::PadSetArguments<half>); break;
-    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::PadGetTunerDefaults, clblast::PadGetTunerSettings<float>, clblast::PadTestValidArguments<float>, clblast::PadSetConstraints, clblast::PadComputeLocalMemSize<float>, clblast::PadSetArguments<float>); break;
-    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::PadGetTunerDefaults, clblast::PadGetTunerSettings<double>, clblast::PadTestValidArguments<double>, clblast::PadSetConstraints, clblast::PadComputeLocalMemSize<double>, clblast::PadSetArguments<double>); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::PadGetTunerDefaults, clblast::PadGetTunerSettings<float2>, clblast::PadTestValidArguments<float2>, clblast::PadSetConstraints, clblast::PadComputeLocalMemSize<float2>, clblast::PadSetArguments<float2>); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::PadGetTunerDefaults, clblast::PadGetTunerSettings<double2>, clblast::PadTestValidArguments<double2>, clblast::PadSetConstraints, clblast::PadComputeLocalMemSize<double2>, clblast::PadSetArguments<double2>); break;
-  }
-  return 0;
+  try {
+    const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+    switch(clblast::GetPrecision(command_line_args)) {
+      case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::PadGetTunerDefaults, clblast::PadGetTunerSettings<half>, clblast::PadTestValidArguments<half>, clblast::PadSetConstraints, clblast::PadComputeLocalMemSize<half>, clblast::PadSetArguments<half>); break;
+      case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::PadGetTunerDefaults, clblast::PadGetTunerSettings<float>, clblast::PadTestValidArguments<float>, clblast::PadSetConstraints, clblast::PadComputeLocalMemSize<float>, clblast::PadSetArguments<float>); break;
+      case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::PadGetTunerDefaults, clblast::PadGetTunerSettings<double>, clblast::PadTestValidArguments<double>, clblast::PadSetConstraints, clblast::PadComputeLocalMemSize<double>, clblast::PadSetArguments<double>); break;
+      case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::PadGetTunerDefaults, clblast::PadGetTunerSettings<float2>, clblast::PadTestValidArguments<float2>, clblast::PadSetConstraints, clblast::PadComputeLocalMemSize<float2>, clblast::PadSetArguments<float2>); break;
+      case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::PadGetTunerDefaults, clblast::PadGetTunerSettings<double2>, clblast::PadTestValidArguments<double2>, clblast::PadSetConstraints, clblast::PadComputeLocalMemSize<double2>, clblast::PadSetArguments<double2>); break;
+    }
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/invert.cpp b/src/tuning/kernels/invert.cpp
index 3795da88..9b2c0467 100644
--- a/src/tuning/kernels/invert.cpp
+++ b/src/tuning/kernels/invert.cpp
@@ -20,15 +20,17 @@ using double2 = clblast::double2;
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
-  switch(clblast::GetPrecision(command_line_args)) {
-    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::InvertGetTunerDefaults, clblast::InvertGetTunerSettings<half>, clblast::InvertTestValidArguments<half>, clblast::InvertSetConstraints, clblast::InvertComputeLocalMemSize<half>, clblast::InvertSetArguments<half>); break;
-    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::InvertGetTunerDefaults, clblast::InvertGetTunerSettings<float>, clblast::InvertTestValidArguments<float>, clblast::InvertSetConstraints, clblast::InvertComputeLocalMemSize<float>, clblast::InvertSetArguments<float>); break;
-    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::InvertGetTunerDefaults, clblast::InvertGetTunerSettings<double>, clblast::InvertTestValidArguments<double>, clblast::InvertSetConstraints, clblast::InvertComputeLocalMemSize<double>, clblast::InvertSetArguments<double>); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::InvertGetTunerDefaults, clblast::InvertGetTunerSettings<float2>, clblast::InvertTestValidArguments<float2>, clblast::InvertSetConstraints, clblast::InvertComputeLocalMemSize<float2>, clblast::InvertSetArguments<float2>); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::InvertGetTunerDefaults, clblast::InvertGetTunerSettings<double2>, clblast::InvertTestValidArguments<double2>, clblast::InvertSetConstraints, clblast::InvertComputeLocalMemSize<double2>, clblast::InvertSetArguments<double2>); break;
-  }
-  return 0;
+  try {
+    const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+    switch(clblast::GetPrecision(command_line_args)) {
+      case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::InvertGetTunerDefaults, clblast::InvertGetTunerSettings<half>, clblast::InvertTestValidArguments<half>, clblast::InvertSetConstraints, clblast::InvertComputeLocalMemSize<half>, clblast::InvertSetArguments<half>); break;
+      case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::InvertGetTunerDefaults, clblast::InvertGetTunerSettings<float>, clblast::InvertTestValidArguments<float>, clblast::InvertSetConstraints, clblast::InvertComputeLocalMemSize<float>, clblast::InvertSetArguments<float>); break;
+      case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::InvertGetTunerDefaults, clblast::InvertGetTunerSettings<double>, clblast::InvertTestValidArguments<double>, clblast::InvertSetConstraints, clblast::InvertComputeLocalMemSize<double>, clblast::InvertSetArguments<double>); break;
+      case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::InvertGetTunerDefaults, clblast::InvertGetTunerSettings<float2>, clblast::InvertTestValidArguments<float2>, clblast::InvertSetConstraints, clblast::InvertComputeLocalMemSize<float2>, clblast::InvertSetArguments<float2>); break;
+      case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::InvertGetTunerDefaults, clblast::InvertGetTunerSettings<double2>, clblast::InvertTestValidArguments<double2>, clblast::InvertSetConstraints, clblast::InvertComputeLocalMemSize<double2>, clblast::InvertSetArguments<double2>); break;
+    }
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/transpose_fast.cpp b/src/tuning/kernels/transpose_fast.cpp
index 024f7385..b48293ee 100644
--- a/src/tuning/kernels/transpose_fast.cpp
+++ b/src/tuning/kernels/transpose_fast.cpp
@@ -20,15 +20,17 @@ using double2 = clblast::double2;
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
-  switch(clblast::GetPrecision(command_line_args)) {
-    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::TransposeGetTunerDefaults, clblast::TransposeGetTunerSettings<half>, clblast::TransposeTestValidArguments<half>, clblast::TransposeSetConstraints, clblast::TransposeComputeLocalMemSize<half>, clblast::TransposeSetArguments<half>); break;
-    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::TransposeGetTunerDefaults, clblast::TransposeGetTunerSettings<float>, clblast::TransposeTestValidArguments<float>, clblast::TransposeSetConstraints, clblast::TransposeComputeLocalMemSize<float>, clblast::TransposeSetArguments<float>); break;
-    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::TransposeGetTunerDefaults, clblast::TransposeGetTunerSettings<double>, clblast::TransposeTestValidArguments<double>, clblast::TransposeSetConstraints, clblast::TransposeComputeLocalMemSize<double>, clblast::TransposeSetArguments<double>); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::TransposeGetTunerDefaults, clblast::TransposeGetTunerSettings<float2>, clblast::TransposeTestValidArguments<float2>, clblast::TransposeSetConstraints, clblast::TransposeComputeLocalMemSize<float2>, clblast::TransposeSetArguments<float2>); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::TransposeGetTunerDefaults, clblast::TransposeGetTunerSettings<double2>, clblast::TransposeTestValidArguments<double2>, clblast::TransposeSetConstraints, clblast::TransposeComputeLocalMemSize<double2>, clblast::TransposeSetArguments<double2>); break;
-  }
-  return 0;
+  try {
+    const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+    switch(clblast::GetPrecision(command_line_args)) {
+      case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::TransposeGetTunerDefaults, clblast::TransposeGetTunerSettings<half>, clblast::TransposeTestValidArguments<half>, clblast::TransposeSetConstraints, clblast::TransposeComputeLocalMemSize<half>, clblast::TransposeSetArguments<half>); break;
+      case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::TransposeGetTunerDefaults, clblast::TransposeGetTunerSettings<float>, clblast::TransposeTestValidArguments<float>, clblast::TransposeSetConstraints, clblast::TransposeComputeLocalMemSize<float>, clblast::TransposeSetArguments<float>); break;
+      case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::TransposeGetTunerDefaults, clblast::TransposeGetTunerSettings<double>, clblast::TransposeTestValidArguments<double>, clblast::TransposeSetConstraints, clblast::TransposeComputeLocalMemSize<double>, clblast::TransposeSetArguments<double>); break;
+      case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::TransposeGetTunerDefaults, clblast::TransposeGetTunerSettings<float2>, clblast::TransposeTestValidArguments<float2>, clblast::TransposeSetConstraints, clblast::TransposeComputeLocalMemSize<float2>, clblast::TransposeSetArguments<float2>); break;
+      case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::TransposeGetTunerDefaults, clblast::TransposeGetTunerSettings<double2>, clblast::TransposeTestValidArguments<double2>, clblast::TransposeSetConstraints, clblast::TransposeComputeLocalMemSize<double2>, clblast::TransposeSetArguments<double2>); break;
+    }
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/transpose_pad.cpp b/src/tuning/kernels/transpose_pad.cpp
index ffaa252b..73c8ef27 100644
--- a/src/tuning/kernels/transpose_pad.cpp
+++ b/src/tuning/kernels/transpose_pad.cpp
@@ -20,15 +20,17 @@ using double2 = clblast::double2;
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
-  switch(clblast::GetPrecision(command_line_args)) {
-    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::PadtransposeGetTunerDefaults, clblast::PadtransposeGetTunerSettings<half>, clblast::PadtransposeTestValidArguments<half>, clblast::PadtransposeSetConstraints, clblast::PadtransposeComputeLocalMemSize<half>, clblast::PadtransposeSetArguments<half>); break;
-    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::PadtransposeGetTunerDefaults, clblast::PadtransposeGetTunerSettings<float>, clblast::PadtransposeTestValidArguments<float>, clblast::PadtransposeSetConstraints, clblast::PadtransposeComputeLocalMemSize<float>, clblast::PadtransposeSetArguments<float>); break;
-    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::PadtransposeGetTunerDefaults, clblast::PadtransposeGetTunerSettings<double>, clblast::PadtransposeTestValidArguments<double>, clblast::PadtransposeSetConstraints, clblast::PadtransposeComputeLocalMemSize<double>, clblast::PadtransposeSetArguments<double>); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::PadtransposeGetTunerDefaults, clblast::PadtransposeGetTunerSettings<float2>, clblast::PadtransposeTestValidArguments<float2>, clblast::PadtransposeSetConstraints, clblast::PadtransposeComputeLocalMemSize<float2>, clblast::PadtransposeSetArguments<float2>); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::PadtransposeGetTunerDefaults, clblast::PadtransposeGetTunerSettings<double2>, clblast::PadtransposeTestValidArguments<double2>, clblast::PadtransposeSetConstraints, clblast::PadtransposeComputeLocalMemSize<double2>, clblast::PadtransposeSetArguments<double2>); break;
-  }
-  return 0;
+  try {
+    const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+    switch(clblast::GetPrecision(command_line_args)) {
+      case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::PadtransposeGetTunerDefaults, clblast::PadtransposeGetTunerSettings<half>, clblast::PadtransposeTestValidArguments<half>, clblast::PadtransposeSetConstraints, clblast::PadtransposeComputeLocalMemSize<half>, clblast::PadtransposeSetArguments<half>); break;
+      case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::PadtransposeGetTunerDefaults, clblast::PadtransposeGetTunerSettings<float>, clblast::PadtransposeTestValidArguments<float>, clblast::PadtransposeSetConstraints, clblast::PadtransposeComputeLocalMemSize<float>, clblast::PadtransposeSetArguments<float>); break;
+      case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::PadtransposeGetTunerDefaults, clblast::PadtransposeGetTunerSettings<double>, clblast::PadtransposeTestValidArguments<double>, clblast::PadtransposeSetConstraints, clblast::PadtransposeComputeLocalMemSize<double>, clblast::PadtransposeSetArguments<double>); break;
+      case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::PadtransposeGetTunerDefaults, clblast::PadtransposeGetTunerSettings<float2>, clblast::PadtransposeTestValidArguments<float2>, clblast::PadtransposeSetConstraints, clblast::PadtransposeComputeLocalMemSize<float2>, clblast::PadtransposeSetArguments<float2>); break;
+      case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::PadtransposeGetTunerDefaults, clblast::PadtransposeGetTunerSettings<double2>, clblast::PadtransposeTestValidArguments<double2>, clblast::PadtransposeSetConstraints, clblast::PadtransposeComputeLocalMemSize<double2>, clblast::PadtransposeSetArguments<double2>); break;
+    }
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/xaxpy.cpp b/src/tuning/kernels/xaxpy.cpp
index 681876ea..e7f3b26a 100644
--- a/src/tuning/kernels/xaxpy.cpp
+++ b/src/tuning/kernels/xaxpy.cpp
@@ -20,15 +20,17 @@ using double2 = clblast::double2;
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
-  switch(clblast::GetPrecision(command_line_args)) {
-    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::XaxpyGetTunerDefaults, clblast::XaxpyGetTunerSettings<half>, clblast::XaxpyTestValidArguments<half>, clblast::XaxpySetConstraints, clblast::XaxpyComputeLocalMemSize<half>, clblast::XaxpySetArguments<half>); break;
-    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::XaxpyGetTunerDefaults, clblast::XaxpyGetTunerSettings<float>, clblast::XaxpyTestValidArguments<float>, clblast::XaxpySetConstraints, clblast::XaxpyComputeLocalMemSize<float>, clblast::XaxpySetArguments<float>); break;
-    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::XaxpyGetTunerDefaults, clblast::XaxpyGetTunerSettings<double>, clblast::XaxpyTestValidArguments<double>, clblast::XaxpySetConstraints, clblast::XaxpyComputeLocalMemSize<double>, clblast::XaxpySetArguments<double>); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::XaxpyGetTunerDefaults, clblast::XaxpyGetTunerSettings<float2>, clblast::XaxpyTestValidArguments<float2>, clblast::XaxpySetConstraints, clblast::XaxpyComputeLocalMemSize<float2>, clblast::XaxpySetArguments<float2>); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::XaxpyGetTunerDefaults, clblast::XaxpyGetTunerSettings<double2>, clblast::XaxpyTestValidArguments<double2>, clblast::XaxpySetConstraints, clblast::XaxpyComputeLocalMemSize<double2>, clblast::XaxpySetArguments<double2>); break;
-  }
-  return 0;
+  try {
+    const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+    switch(clblast::GetPrecision(command_line_args)) {
+      case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::XaxpyGetTunerDefaults, clblast::XaxpyGetTunerSettings<half>, clblast::XaxpyTestValidArguments<half>, clblast::XaxpySetConstraints, clblast::XaxpyComputeLocalMemSize<half>, clblast::XaxpySetArguments<half>); break;
+      case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::XaxpyGetTunerDefaults, clblast::XaxpyGetTunerSettings<float>, clblast::XaxpyTestValidArguments<float>, clblast::XaxpySetConstraints, clblast::XaxpyComputeLocalMemSize<float>, clblast::XaxpySetArguments<float>); break;
+      case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::XaxpyGetTunerDefaults, clblast::XaxpyGetTunerSettings<double>, clblast::XaxpyTestValidArguments<double>, clblast::XaxpySetConstraints, clblast::XaxpyComputeLocalMemSize<double>, clblast::XaxpySetArguments<double>); break;
+      case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::XaxpyGetTunerDefaults, clblast::XaxpyGetTunerSettings<float2>, clblast::XaxpyTestValidArguments<float2>, clblast::XaxpySetConstraints, clblast::XaxpyComputeLocalMemSize<float2>, clblast::XaxpySetArguments<float2>); break;
+      case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::XaxpyGetTunerDefaults, clblast::XaxpyGetTunerSettings<double2>, clblast::XaxpyTestValidArguments<double2>, clblast::XaxpySetConstraints, clblast::XaxpyComputeLocalMemSize<double2>, clblast::XaxpySetArguments<double2>); break;
+    }
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/xconvgemm.cpp b/src/tuning/kernels/xconvgemm.cpp
index 15dfe829..e9044279 100644
--- a/src/tuning/kernels/xconvgemm.cpp
+++ b/src/tuning/kernels/xconvgemm.cpp
@@ -31,8 +31,10 @@ void StartVariation(int argc, char *argv[]) {
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  StartVariation<1>(argc, argv);
-  return 0;
+  try {
+    StartVariation<1>(argc, argv);
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/xdot.cpp b/src/tuning/kernels/xdot.cpp
index a481f23b..bdd93067 100644
--- a/src/tuning/kernels/xdot.cpp
+++ b/src/tuning/kernels/xdot.cpp
@@ -34,9 +34,11 @@ void StartVariation(int argc, char *argv[]) {
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  StartVariation<1>(argc, argv);
-  StartVariation<2>(argc, argv);
-  return 0;
+  try {
+    StartVariation<1>(argc, argv);
+    StartVariation<2>(argc, argv);
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/xgemm.cpp b/src/tuning/kernels/xgemm.cpp
index dd907ba4..5a48a50d 100644
--- a/src/tuning/kernels/xgemm.cpp
+++ b/src/tuning/kernels/xgemm.cpp
@@ -33,15 +33,17 @@ void StartVariation(int argc, char *argv[]) {
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  printf("* (1/4) Tuning main GEMM kernel (GEMMK == 0) for fixed set of parameters\n\n");
-  StartVariation<1>(argc, argv);
-  printf("* (2/4) Tuning main GEMM kernel (GEMMK == 0) for random parameters out of larger set\n\n");
-  StartVariation<2>(argc, argv);
-  printf("* (3/4) Tuning secondary GEMM kernel (GEMMK == 1) for fixed set of parameters\n\n");
-  StartVariation<11>(argc, argv);
-  printf("* (4/4) Tuning secondary GEMM kernel (GEMMK == 1) for random parameters out of larger set\n\n");
-  StartVariation<12>(argc, argv);
-  return 0;
+  try {
+    printf("* (1/4) Tuning main GEMM kernel (GEMMK == 0) for fixed set of parameters\n\n");
+    StartVariation<1>(argc, argv);
+    printf("* (2/4) Tuning main GEMM kernel (GEMMK == 0) for random parameters out of larger set\n\n");
+    StartVariation<2>(argc, argv);
+    printf("* (3/4) Tuning secondary GEMM kernel (GEMMK == 1) for fixed set of parameters\n\n");
+    StartVariation<11>(argc, argv);
+    printf("* (4/4) Tuning secondary GEMM kernel (GEMMK == 1) for random parameters out of larger set\n\n");
+    StartVariation<12>(argc, argv);
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/xgemm_direct.cpp b/src/tuning/kernels/xgemm_direct.cpp
index 73c2217c..330e6d51 100644
--- a/src/tuning/kernels/xgemm_direct.cpp
+++ b/src/tuning/kernels/xgemm_direct.cpp
@@ -33,9 +33,11 @@ void StartVariation(int argc, char *argv[]) {
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  StartVariation<1>(argc, argv);
-  StartVariation<2>(argc, argv);
-  return 0;
+  try {
+    StartVariation<1>(argc, argv);
+    StartVariation<2>(argc, argv);
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/xgemv.cpp b/src/tuning/kernels/xgemv.cpp
index 6505a081..065f6b6f 100644
--- a/src/tuning/kernels/xgemv.cpp
+++ b/src/tuning/kernels/xgemv.cpp
@@ -33,10 +33,12 @@ void StartVariation(int argc, char *argv[]) {
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  StartVariation<1>(argc, argv);
-  StartVariation<2>(argc, argv);
-  StartVariation<3>(argc, argv);
-  return 0;
+  try {
+    StartVariation<1>(argc, argv);
+    StartVariation<2>(argc, argv);
+    StartVariation<3>(argc, argv);
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/kernels/xger.cpp b/src/tuning/kernels/xger.cpp
index e4c9fc03..372d9d05 100644
--- a/src/tuning/kernels/xger.cpp
+++ b/src/tuning/kernels/xger.cpp
@@ -20,15 +20,17 @@ using double2 = clblast::double2;
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
-  switch(clblast::GetPrecision(command_line_args)) {
-    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::XgerGetTunerDefaults, clblast::XgerGetTunerSettings<half>, clblast::XgerTestValidArguments<half>, clblast::XgerSetConstraints, clblast::XgerComputeLocalMemSize<half>, clblast::XgerSetArguments<half>); break;
-    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::XgerGetTunerDefaults, clblast::XgerGetTunerSettings<float>, clblast::XgerTestValidArguments<float>, clblast::XgerSetConstraints, clblast::XgerComputeLocalMemSize<float>, clblast::XgerSetArguments<float>); break;
-    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::XgerGetTunerDefaults, clblast::XgerGetTunerSettings<double>, clblast::XgerTestValidArguments<double>, clblast::XgerSetConstraints, clblast::XgerComputeLocalMemSize<double>, clblast::XgerSetArguments<double>); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::XgerGetTunerDefaults, clblast::XgerGetTunerSettings<float2>, clblast::XgerTestValidArguments<float2>, clblast::XgerSetConstraints, clblast::XgerComputeLocalMemSize<float2>, clblast::XgerSetArguments<float2>); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::XgerGetTunerDefaults, clblast::XgerGetTunerSettings<double2>, clblast::XgerTestValidArguments<double2>, clblast::XgerSetConstraints, clblast::XgerComputeLocalMemSize<double2>, clblast::XgerSetArguments<double2>); break;
-  }
-  return 0;
+  try {
+    const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+    switch(clblast::GetPrecision(command_line_args)) {
+      case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::XgerGetTunerDefaults, clblast::XgerGetTunerSettings<half>, clblast::XgerTestValidArguments<half>, clblast::XgerSetConstraints, clblast::XgerComputeLocalMemSize<half>, clblast::XgerSetArguments<half>); break;
+      case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::XgerGetTunerDefaults, clblast::XgerGetTunerSettings<float>, clblast::XgerTestValidArguments<float>, clblast::XgerSetConstraints, clblast::XgerComputeLocalMemSize<float>, clblast::XgerSetArguments<float>); break;
+      case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::XgerGetTunerDefaults, clblast::XgerGetTunerSettings<double>, clblast::XgerTestValidArguments<double>, clblast::XgerSetConstraints, clblast::XgerComputeLocalMemSize<double>, clblast::XgerSetArguments<double>); break;
+      case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::XgerGetTunerDefaults, clblast::XgerGetTunerSettings<float2>, clblast::XgerTestValidArguments<float2>, clblast::XgerSetConstraints, clblast::XgerComputeLocalMemSize<float2>, clblast::XgerSetArguments<float2>); break;
+      case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::XgerGetTunerDefaults, clblast::XgerGetTunerSettings<double2>, clblast::XgerTestValidArguments<double2>, clblast::XgerSetConstraints, clblast::XgerComputeLocalMemSize<double2>, clblast::XgerSetArguments<double2>); break;
+    }
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/routines/xgemm.cpp b/src/tuning/routines/xgemm.cpp
index 7d886ebf..f5c4569e 100644
--- a/src/tuning/routines/xgemm.cpp
+++ b/src/tuning/routines/xgemm.cpp
@@ -214,15 +214,17 @@ using double2 = clblast::double2;
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
-  switch(clblast::GetPrecision(command_line_args)) {
-    case clblast::Precision::kHalf: clblast::TuneXgemm<half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::TuneXgemm<float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::TuneXgemm<double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::TuneXgemm<float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::TuneXgemm<double2>(argc, argv); break;
-  }
-  return 0;
+  try {
+    const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+    switch(clblast::GetPrecision(command_line_args)) {
+      case clblast::Precision::kHalf: clblast::TuneXgemm<half>(argc, argv); break;
+      case clblast::Precision::kSingle: clblast::TuneXgemm<float>(argc, argv); break;
+      case clblast::Precision::kDouble: clblast::TuneXgemm<double>(argc, argv); break;
+      case clblast::Precision::kComplexSingle: clblast::TuneXgemm<float2>(argc, argv); break;
+      case clblast::Precision::kComplexDouble: clblast::TuneXgemm<double2>(argc, argv); break;
+    }
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
diff --git a/src/tuning/routines/xtrsv.cpp b/src/tuning/routines/xtrsv.cpp
index 7551d1f5..77252273 100644
--- a/src/tuning/routines/xtrsv.cpp
+++ b/src/tuning/routines/xtrsv.cpp
@@ -132,14 +132,16 @@ using double2 = clblast::double2;
 
 // Main function (not within the clblast namespace)
 int main(int argc, char *argv[]) {
-  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
-  switch(clblast::GetPrecision(command_line_args)) {
-    case clblast::Precision::kSingle: clblast::TuneXtrsv<float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::TuneXtrsv<double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::TuneXtrsv<float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::TuneXtrsv<double2>(argc, argv); break;
-  }
-  return 0;
+  try {
+    const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+    switch(clblast::GetPrecision(command_line_args)) {
+      case clblast::Precision::kSingle: clblast::TuneXtrsv<float>(argc, argv); break;
+      case clblast::Precision::kDouble: clblast::TuneXtrsv<double>(argc, argv); break;
+      case clblast::Precision::kComplexSingle: clblast::TuneXtrsv<float2>(argc, argv); break;
+      case clblast::Precision::kComplexDouble: clblast::TuneXtrsv<double2>(argc, argv); break;
+    }
+    return 0;
+  } catch (...) { return static_cast<int>(clblast::DispatchException()); }
 }
 
 // =================================================================================================
-- 
cgit v1.2.3