From 249bdaa8e9a111573f5c3a821230bba6437817c7 Mon Sep 17 00:00:00 2001
From: Cedric Nugteren <web@cedricnugteren.nl>
Date: Mon, 18 Dec 2017 21:34:07 +0100
Subject: Reformatted tuning code to make compilation faster

---
 src/tuning/kernels/copy_fast.cpp      | 140 ++++++++--------
 src/tuning/kernels/copy_pad.cpp       | 156 +++++++++---------
 src/tuning/kernels/transpose_fast.cpp | 140 ++++++++--------
 src/tuning/kernels/transpose_pad.cpp  | 154 +++++++++---------
 src/tuning/kernels/xaxpy.cpp          | 142 ++++++++---------
 src/tuning/kernels/xdot.cpp           | 142 ++++++++---------
 src/tuning/kernels/xgemm.cpp          | 270 ++++++++++++++++---------------
 src/tuning/kernels/xgemm_direct.cpp   | 266 +++++++++++++++----------------
 src/tuning/kernels/xgemv.cpp          | 216 +++++++++++++------------
 src/tuning/kernels/xger.cpp           | 158 +++++++++---------
 src/tuning/tuning.cpp                 | 288 +++++++++++++++++++++++++++++++++
 src/tuning/tuning.hpp                 | 290 ++--------------------------------
 12 files changed, 1177 insertions(+), 1185 deletions(-)

diff --git a/src/tuning/kernels/copy_fast.cpp b/src/tuning/kernels/copy_fast.cpp
index 462107d3..d046c9e9 100644
--- a/src/tuning/kernels/copy_fast.cpp
+++ b/src/tuning/kernels/copy_fast.cpp
@@ -20,78 +20,74 @@
 namespace clblast {
 // =================================================================================================
 
-// See comment at top of file for a description of the class
-template <typename T>
-class TuneCopy {
- public:
-
-  // Settings for this kernel (default command-line arguments)
-  static TunerDefaults GetTunerDefaults() {
-    auto settings = TunerDefaults();
-    settings.options = {kArgM, kArgN, kArgAlpha};
-    settings.default_m = 1024;
-    settings.default_n = 1024;
-    return settings;
-  }
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgM, kArgN, kArgAlpha};
+  settings.default_m = 1024;
+  settings.default_n = 1024;
+  return settings;
+}
 
-  // Settings for this kernel (general)
-  static TunerSettings GetTunerSettings(const Arguments<T> &args) {
-    auto settings = TunerSettings();
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings GetTunerSettings(const int, const Arguments<T> &args) {
+  auto settings = TunerSettings();
 
-    // Identification of the kernel
-    settings.kernel_family = "copy";
-    settings.kernel_name = "CopyMatrixFast";
-    settings.sources =
+  // Identification of the kernel
+  settings.kernel_family = "copy";
+  settings.kernel_name = "CopyMatrixFast";
+  settings.sources =
 #include "../src/kernels/level3/level3.opencl"
 #include "../src/kernels/level3/copy_fast.opencl"
-    ;
-
-    // Buffer sizes
-    settings.size_a = args.m * args.n;
-    settings.size_b = args.m * args.n;
-
-    // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
-    settings.inputs = {2, 3};
-    settings.outputs = {3};
-
-    // Sets the base thread configuration
-    settings.global_size = {args.m, args.n};
-    settings.global_size_ref = settings.global_size;
-    settings.local_size = {1, 1};
-    settings.local_size_ref = {8, 8};
-
-    // Transforms the thread configuration based on the parameters
-    settings.mul_local = {{"COPY_DIMX", "COPY_DIMY"}};
-    settings.div_global = {{"COPY_VW", "COPY_WPT"}};
-
-    // Sets the tuning parameters and their possible values
-    settings.parameters = {
-      {"COPY_DIMX", {8, 16, 32}},
-      {"COPY_DIMY", {8, 16, 32}},
-      {"COPY_WPT", {1, 2, 4, 8}},
-      {"COPY_VW", {1, 2, 4, 8}},
-    };
-
-    // Describes how to compute the performance metrics
-    settings.metric_amount = 2 * args.m * args.n * GetBytes(args.precision);
-    settings.performance_unit = "GB/s";
-
-    return settings;
-  }
+  ;
+
+  // Buffer sizes
+  settings.size_a = args.m * args.n;
+  settings.size_b = args.m * args.n;
+
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {2, 3};
+  settings.outputs = {3};
+
+  // Sets the base thread configuration
+  settings.global_size = {args.m, args.n};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1, 1};
+  settings.local_size_ref = {8, 8};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"COPY_DIMX", "COPY_DIMY"}};
+  settings.div_global = {{"COPY_VW", "COPY_WPT"}};
+
+  // Sets the tuning parameters and their possible values
+  settings.parameters = {
+    {"COPY_DIMX", {8, 16, 32}},
+    {"COPY_DIMY", {8, 16, 32}},
+    {"COPY_WPT", {1, 2, 4, 8}},
+    {"COPY_VW", {1, 2, 4, 8}},
+  };
+
+  // Describes how to compute the performance metrics
+  settings.metric_amount = 2 * args.m * args.n * GetBytes(args.precision);
+  settings.performance_unit = "GB/s";
+
+  return settings;
+}
 
-  // Tests for valid arguments
-  static void TestValidArguments(const Arguments<T> &) { }
-  static std::vector<Constraint> SetConstraints() { return {}; }
-
-  // Sets the kernel's arguments
-  static void SetArguments(Kernel &kernel, const Arguments<T> &args,
-                           std::vector<Buffer<T>>& buffers) {
-    kernel.SetArgument(0, static_cast<int>(args.m));
-    kernel.SetArgument(1, buffers[2]()); // 2 == A matrix
-    kernel.SetArgument(2, buffers[3]()); // 3 == B matrix
-    kernel.SetArgument(3, GetRealArg(args.alpha));
-  }
-};
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int, const Arguments<T> &) { }
+std::vector<Constraint> SetConstraints(const int) { return {}; }
+
+// Sets the kernel's arguments
+template <typename T>
+void SetArguments(const int, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  kernel.SetArgument(0, static_cast<int>(args.m));
+  kernel.SetArgument(1, buffers[2]()); // 2 == A matrix
+  kernel.SetArgument(2, buffers[3]()); // 3 == B matrix
+  kernel.SetArgument(3, GetRealArg(args.alpha));
+}
 
 // =================================================================================================
 } // namespace clblast
@@ -105,11 +101,11 @@ using double2 = clblast::double2;
 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<clblast::TuneCopy<half>, half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneCopy<float>, float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneCopy<double>, double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneCopy<float2>, float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneCopy<double2>, double2>(argc, argv); break;
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break;
+    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break;
+    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break;
   }
   return 0;
 }
diff --git a/src/tuning/kernels/copy_pad.cpp b/src/tuning/kernels/copy_pad.cpp
index 24557517..1b483e86 100644
--- a/src/tuning/kernels/copy_pad.cpp
+++ b/src/tuning/kernels/copy_pad.cpp
@@ -20,86 +20,82 @@
 namespace clblast {
 // =================================================================================================
 
-// See comment at top of file for a description of the class
-template <typename T>
-class TunePad {
- public:
-
-  // Settings for this kernel (default command-line arguments)
-  static TunerDefaults GetTunerDefaults() {
-    auto settings = TunerDefaults();
-    settings.options = {kArgM, kArgN, kArgAlpha};
-    settings.default_m = 1024;
-    settings.default_n = 1024;
-    return settings;
-  }
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgM, kArgN, kArgAlpha};
+  settings.default_m = 1024;
+  settings.default_n = 1024;
+  return settings;
+}
 
-  // Settings for this kernel (general)
-  static TunerSettings GetTunerSettings(const Arguments<T> &args) {
-    auto settings = TunerSettings();
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings GetTunerSettings(const int, const Arguments<T> &args) {
+  auto settings = TunerSettings();
 
-    // Identification of the kernel
-    settings.kernel_family = "pad";
-    settings.kernel_name = "CopyPadMatrix";
-    settings.sources =
+  // Identification of the kernel
+  settings.kernel_family = "pad";
+  settings.kernel_name = "CopyPadMatrix";
+  settings.sources =
 #include "../src/kernels/level3/level3.opencl"
 #include "../src/kernels/level3/copy_pad.opencl"
-    ;
-
-    // Buffer sizes
-    settings.size_a = args.m * args.n;
-    settings.size_b = args.m * args.n;
-
-    // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
-    settings.inputs = {2, 3};
-    settings.outputs = {3};
-
-    // Sets the base thread configuration
-    settings.global_size = {args.m, args.n};
-    settings.global_size_ref = settings.global_size;
-    settings.local_size = {1, 1};
-    settings.local_size_ref = {8, 8};
-
-    // Transforms the thread configuration based on the parameters
-    settings.mul_local = {{"PAD_DIMX", "PAD_DIMY"}};
-    settings.div_global = {{"PAD_WPTX", "PAD_WPTY"}};
-
-    // Sets the tuning parameters and their possible values
-    settings.parameters = {
-      {"PAD_DIMX", {8, 16, 32}},
-      {"PAD_DIMY", {8, 16, 32}},
-      {"PAD_WPTX", {1, 2, 4}},
-      {"PAD_WPTY", {1, 2, 4}},
-    };
-
-    // Describes how to compute the performance metrics
-    settings.metric_amount = 2 * args.m * args.n * GetBytes(args.precision);
-    settings.performance_unit = "GB/s";
-
-    return settings;
-  }
+  ;
+
+  // Buffer sizes
+  settings.size_a = args.m * args.n;
+  settings.size_b = args.m * args.n;
+
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {2, 3};
+  settings.outputs = {3};
+
+  // Sets the base thread configuration
+  settings.global_size = {args.m, args.n};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1, 1};
+  settings.local_size_ref = {8, 8};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"PAD_DIMX", "PAD_DIMY"}};
+  settings.div_global = {{"PAD_WPTX", "PAD_WPTY"}};
+
+  // Sets the tuning parameters and their possible values
+  settings.parameters = {
+    {"PAD_DIMX", {8, 16, 32}},
+    {"PAD_DIMY", {8, 16, 32}},
+    {"PAD_WPTX", {1, 2, 4}},
+    {"PAD_WPTY", {1, 2, 4}},
+  };
+
+  // Describes how to compute the performance metrics
+  settings.metric_amount = 2 * args.m * args.n * GetBytes(args.precision);
+  settings.performance_unit = "GB/s";
+
+  return settings;
+}
 
-  // Tests for valid arguments
-  static void TestValidArguments(const Arguments<T> &) { }
-  static std::vector<Constraint> SetConstraints() { return {}; }
-
-  // Sets the kernel's arguments
-  static void SetArguments(Kernel &kernel, const Arguments<T> &args,
-                           std::vector<Buffer<T>>& buffers) {
-    kernel.SetArgument(0, static_cast<int>(args.m));
-    kernel.SetArgument(1, static_cast<int>(args.n));
-    kernel.SetArgument(2, static_cast<int>(args.m));
-    kernel.SetArgument(3, 0);
-    kernel.SetArgument(4, buffers[2]()); // 2 == A matrix
-    kernel.SetArgument(5, static_cast<int>(args.m));
-    kernel.SetArgument(6, static_cast<int>(args.n));
-    kernel.SetArgument(7, static_cast<int>(args.m));
-    kernel.SetArgument(8, 0);
-    kernel.SetArgument(9, buffers[3]()); // 3 == B matrix
-    kernel.SetArgument(10, GetRealArg(args.alpha));
-    kernel.SetArgument(11, 0);
-  }
-};
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int, const Arguments<T> &) { }
+std::vector<Constraint> SetConstraints(const int) { return {}; }
+
+// Sets the kernel's arguments
+template <typename T>
+void SetArguments(const int, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  kernel.SetArgument(0, static_cast<int>(args.m));
+  kernel.SetArgument(1, static_cast<int>(args.n));
+  kernel.SetArgument(2, static_cast<int>(args.m));
+  kernel.SetArgument(3, 0);
+  kernel.SetArgument(4, buffers[2]()); // 2 == A matrix
+  kernel.SetArgument(5, static_cast<int>(args.m));
+  kernel.SetArgument(6, static_cast<int>(args.n));
+  kernel.SetArgument(7, static_cast<int>(args.m));
+  kernel.SetArgument(8, 0);
+  kernel.SetArgument(9, buffers[3]()); // 3 == B matrix
+  kernel.SetArgument(10, GetRealArg(args.alpha));
+  kernel.SetArgument(11, 0);
+}
 
 // =================================================================================================
 } // namespace clblast
@@ -113,11 +109,11 @@ using double2 = clblast::double2;
 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<clblast::TunePad<half>, half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::Tuner<clblast::TunePad<float>, float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::Tuner<clblast::TunePad<double>, double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TunePad<float2>, float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TunePad<double2>, double2>(argc, argv); break;
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break;
+    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break;
+    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break;
   }
   return 0;
 }
diff --git a/src/tuning/kernels/transpose_fast.cpp b/src/tuning/kernels/transpose_fast.cpp
index 1e0d3c7b..5b701a5b 100644
--- a/src/tuning/kernels/transpose_fast.cpp
+++ b/src/tuning/kernels/transpose_fast.cpp
@@ -20,78 +20,74 @@
 namespace clblast {
 // =================================================================================================
 
-// See comment at top of file for a description of the class
-template <typename T>
-class TuneTranspose {
- public:
-
-  // Settings for this kernel (default command-line arguments)
-  static TunerDefaults GetTunerDefaults() {
-    auto settings = TunerDefaults();
-    settings.options = {kArgM, kArgN, kArgAlpha};
-    settings.default_m = 1024;
-    settings.default_n = 1024;
-    return settings;
-  }
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgM, kArgN, kArgAlpha};
+  settings.default_m = 1024;
+  settings.default_n = 1024;
+  return settings;
+}
 
-  // Settings for this kernel (general)
-  static TunerSettings GetTunerSettings(const Arguments<T> &args) {
-    auto settings = TunerSettings();
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings GetTunerSettings(const int, const Arguments<T> &args) {
+  auto settings = TunerSettings();
 
-    // Identification of the kernel
-    settings.kernel_family = "transpose";
-    settings.kernel_name = "TransposeMatrixFast";
-    settings.sources =
+  // Identification of the kernel
+  settings.kernel_family = "transpose";
+  settings.kernel_name = "TransposeMatrixFast";
+  settings.sources =
 #include "../src/kernels/level3/level3.opencl"
 #include "../src/kernels/level3/transpose_fast.opencl"
-    ;
-
-    // Buffer sizes
-    settings.size_a = args.m * args.n;
-    settings.size_b = args.m * args.n;
-
-    // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
-    settings.inputs = {2, 3};
-    settings.outputs = {3};
-
-    // Sets the base thread configuration
-    settings.global_size = {args.m, args.n};
-    settings.global_size_ref = settings.global_size;
-    settings.local_size = {1, 1};
-    settings.local_size_ref = {8, 8};
-
-    // Transforms the thread configuration based on the parameters
-    settings.mul_local = {{"TRA_DIM", "TRA_DIM"}};
-    settings.div_global = {{"TRA_WPT", "TRA_WPT"}};
-
-    // Sets the tuning parameters and their possible values
-    settings.parameters = {
-      {"TRA_DIM", {4, 8, 16, 32, 64}},
-      {"TRA_WPT", {1, 2, 4, 8, 16}},
-      {"TRA_PAD", {0, 1}},
-      {"TRA_SHUFFLE", {0, 1}},
-    };
-
-    // Describes how to compute the performance metrics
-    settings.metric_amount = 2 * args.m * args.n * GetBytes(args.precision);
-    settings.performance_unit = "GB/s";
-
-    return settings;
-  }
+  ;
+
+  // Buffer sizes
+  settings.size_a = args.m * args.n;
+  settings.size_b = args.m * args.n;
+
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {2, 3};
+  settings.outputs = {3};
+
+  // Sets the base thread configuration
+  settings.global_size = {args.m, args.n};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1, 1};
+  settings.local_size_ref = {8, 8};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"TRA_DIM", "TRA_DIM"}};
+  settings.div_global = {{"TRA_WPT", "TRA_WPT"}};
+
+  // Sets the tuning parameters and their possible values
+  settings.parameters = {
+    {"TRA_DIM", {4, 8, 16, 32, 64}},
+    {"TRA_WPT", {1, 2, 4, 8, 16}},
+    {"TRA_PAD", {0, 1}},
+    {"TRA_SHUFFLE", {0, 1}},
+  };
+
+  // Describes how to compute the performance metrics
+  settings.metric_amount = 2 * args.m * args.n * GetBytes(args.precision);
+  settings.performance_unit = "GB/s";
+
+  return settings;
+}
 
-  // Tests for valid arguments
-  static void TestValidArguments(const Arguments<T> &) { }
-  static std::vector<Constraint> SetConstraints() { return {}; }
-
-  // Sets the kernel's arguments
-  static void SetArguments(Kernel &kernel, const Arguments<T> &args,
-                           std::vector<Buffer<T>>& buffers) {
-    kernel.SetArgument(0, static_cast<int>(args.m));
-    kernel.SetArgument(1, buffers[2]()); // 2 == A matrix
-    kernel.SetArgument(2, buffers[3]()); // 3 == B matrix
-    kernel.SetArgument(3, GetRealArg(args.alpha));
-  }
-};
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int, const Arguments<T> &) { }
+std::vector<Constraint> SetConstraints(const int) { return {}; }
+
+// Sets the kernel's arguments
+template <typename T>
+void SetArguments(const int, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  kernel.SetArgument(0, static_cast<int>(args.m));
+  kernel.SetArgument(1, buffers[2]()); // 2 == A matrix
+  kernel.SetArgument(2, buffers[3]()); // 3 == B matrix
+  kernel.SetArgument(3, GetRealArg(args.alpha));
+}
 
 // =================================================================================================
 } // namespace clblast
@@ -105,11 +101,11 @@ using double2 = clblast::double2;
 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<clblast::TuneTranspose<half>, half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneTranspose<float>, float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneTranspose<double>, double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneTranspose<float2>, float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneTranspose<double2>, double2>(argc, argv); break;
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break;
+    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break;
+    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break;
   }
   return 0;
 }
diff --git a/src/tuning/kernels/transpose_pad.cpp b/src/tuning/kernels/transpose_pad.cpp
index 087f8e67..ed24fb04 100644
--- a/src/tuning/kernels/transpose_pad.cpp
+++ b/src/tuning/kernels/transpose_pad.cpp
@@ -20,85 +20,81 @@
 namespace clblast {
 // =================================================================================================
 
-// See comment at top of file for a description of the class
-template <typename T>
-class TunePadTranspose {
- public:
-
-  // Settings for this kernel (default command-line arguments)
-  static TunerDefaults GetTunerDefaults() {
-    auto settings = TunerDefaults();
-    settings.options = {kArgM, kArgN, kArgAlpha};
-    settings.default_m = 1024;
-    settings.default_n = 1024;
-    return settings;
-  }
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgM, kArgN, kArgAlpha};
+  settings.default_m = 1024;
+  settings.default_n = 1024;
+  return settings;
+}
 
-  // Settings for this kernel (general)
-  static TunerSettings GetTunerSettings(const Arguments<T> &args) {
-    auto settings = TunerSettings();
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings GetTunerSettings(const int, const Arguments<T> &args) {
+  auto settings = TunerSettings();
 
-    // Identification of the kernel
-    settings.kernel_family = "padtranspose";
-    settings.kernel_name = "TransposePadMatrix";
-    settings.sources =
+  // Identification of the kernel
+  settings.kernel_family = "padtranspose";
+  settings.kernel_name = "TransposePadMatrix";
+  settings.sources =
 #include "../src/kernels/level3/level3.opencl"
 #include "../src/kernels/level3/transpose_pad.opencl"
-    ;
-
-    // Buffer sizes
-    settings.size_a = args.m * args.n;
-    settings.size_b = args.m * args.n;
-
-    // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
-    settings.inputs = {2, 3};
-    settings.outputs = {3};
-
-    // Sets the base thread configuration
-    settings.global_size = {args.m, args.n};
-    settings.global_size_ref = settings.global_size;
-    settings.local_size = {1, 1};
-    settings.local_size_ref = {8, 8};
-
-    // Transforms the thread configuration based on the parameters
-    settings.mul_local = {{"PADTRA_TILE", "PADTRA_TILE"}};
-    settings.div_global = {{"PADTRA_WPT", "PADTRA_WPT"}};
-
-    // Sets the tuning parameters and their possible values
-    settings.parameters = {
-      {"PADTRA_TILE", {8, 16, 32, 64}},
-      {"PADTRA_WPT", {1, 2, 4, 8, 16}},
-      {"PADTRA_PAD", {0, 1}},
-    };
-
-    // Describes how to compute the performance metrics
-    settings.metric_amount = 2 * args.m * args.n * GetBytes(args.precision);
-    settings.performance_unit = "GB/s";
-
-    return settings;
-  }
+  ;
+
+  // Buffer sizes
+  settings.size_a = args.m * args.n;
+  settings.size_b = args.m * args.n;
+
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {2, 3};
+  settings.outputs = {3};
+
+  // Sets the base thread configuration
+  settings.global_size = {args.m, args.n};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1, 1};
+  settings.local_size_ref = {8, 8};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"PADTRA_TILE", "PADTRA_TILE"}};
+  settings.div_global = {{"PADTRA_WPT", "PADTRA_WPT"}};
+
+  // Sets the tuning parameters and their possible values
+  settings.parameters = {
+    {"PADTRA_TILE", {8, 16, 32, 64}},
+    {"PADTRA_WPT", {1, 2, 4, 8, 16}},
+    {"PADTRA_PAD", {0, 1}},
+  };
+
+  // Describes how to compute the performance metrics
+  settings.metric_amount = 2 * args.m * args.n * GetBytes(args.precision);
+  settings.performance_unit = "GB/s";
+
+  return settings;
+}
 
-  // Tests for valid arguments
-  static void TestValidArguments(const Arguments<T> &) { }
-  static std::vector<Constraint> SetConstraints() { return {}; }
-
-  // Sets the kernel's arguments
-  static void SetArguments(Kernel &kernel, const Arguments<T> &args,
-                           std::vector<Buffer<T>>& buffers) {
-    kernel.SetArgument(0, static_cast<int>(args.m));
-    kernel.SetArgument(1, static_cast<int>(args.n));
-    kernel.SetArgument(2, static_cast<int>(args.m));
-    kernel.SetArgument(3, 0);
-    kernel.SetArgument(4, buffers[2]()); // 2 == A matrix
-    kernel.SetArgument(5, static_cast<int>(args.n));
-    kernel.SetArgument(6, static_cast<int>(args.m));
-    kernel.SetArgument(7, static_cast<int>(args.n));
-    kernel.SetArgument(8, 0);
-    kernel.SetArgument(9, buffers[3]()); // 3 == B matrix
-    kernel.SetArgument(10, GetRealArg(args.alpha));
-    kernel.SetArgument(11, 0);
-  }
-};
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int, const Arguments<T> &) { }
+std::vector<Constraint> SetConstraints(const int) { return {}; }
+
+// Sets the kernel's arguments
+template <typename T>
+void SetArguments(const int, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  kernel.SetArgument(0, static_cast<int>(args.m));
+  kernel.SetArgument(1, static_cast<int>(args.n));
+  kernel.SetArgument(2, static_cast<int>(args.m));
+  kernel.SetArgument(3, 0);
+  kernel.SetArgument(4, buffers[2]()); // 2 == A matrix
+  kernel.SetArgument(5, static_cast<int>(args.n));
+  kernel.SetArgument(6, static_cast<int>(args.m));
+  kernel.SetArgument(7, static_cast<int>(args.n));
+  kernel.SetArgument(8, 0);
+  kernel.SetArgument(9, buffers[3]()); // 3 == B matrix
+  kernel.SetArgument(10, GetRealArg(args.alpha));
+  kernel.SetArgument(11, 0);
+}
 
 // =================================================================================================
 } // namespace clblast
@@ -112,11 +108,11 @@ using double2 = clblast::double2;
 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<clblast::TunePadTranspose<half>, half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::Tuner<clblast::TunePadTranspose<float>, float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::Tuner<clblast::TunePadTranspose<double>, double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TunePadTranspose<float2>, float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TunePadTranspose<double2>, double2>(argc, argv); break;
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break;
+    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break;
+    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break;
   }
   return 0;
 }
diff --git a/src/tuning/kernels/xaxpy.cpp b/src/tuning/kernels/xaxpy.cpp
index d843ea78..dd44018c 100644
--- a/src/tuning/kernels/xaxpy.cpp
+++ b/src/tuning/kernels/xaxpy.cpp
@@ -20,80 +20,76 @@
 namespace clblast {
 // =================================================================================================
 
-// See comment at top of file for a description of the class
-template <typename T>
-class TuneXaxpy {
- public:
-
-  // Settings for this kernel (default command-line arguments)
-  static TunerDefaults GetTunerDefaults() {
-    auto settings = TunerDefaults();
-    settings.options = {kArgN, kArgAlpha};
-    settings.default_n = 4096*1024;
-    return settings;
-  }
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgN, kArgAlpha};
+  settings.default_n = 4096*1024;
+  return settings;
+}
 
-  // Settings for this kernel (general)
-  static TunerSettings GetTunerSettings(const Arguments<T> &args) {
-    auto settings = TunerSettings();
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings GetTunerSettings(const int, const Arguments<T> &args) {
+  auto settings = TunerSettings();
 
-    // Identification of the kernel
-    settings.kernel_family = "xaxpy";
-    settings.kernel_name = "XaxpyFastest";
-    settings.sources =
+  // Identification of the kernel
+  settings.kernel_family = "xaxpy";
+  settings.kernel_name = "XaxpyFastest";
+  settings.sources =
 #include "../src/kernels/level1/level1.opencl"
 #include "../src/kernels/level1/xaxpy.opencl"
-    ;
-
-    // Buffer sizes
-    settings.size_x = args.n;
-    settings.size_y = args.n;
-
-    // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
-    settings.inputs = {0, 1};
-    settings.outputs = {1};
-
-    // Sets the base thread configuration
-    settings.global_size = {args.n};
-    settings.global_size_ref = settings.global_size;
-    settings.local_size = {1};
-    settings.local_size_ref = {64};
-
-    // Transforms the thread configuration based on the parameters
-    settings.mul_local = {{"WGS"}};
-    settings.div_global = {{"WPT"},{"VW"}};
-
-    // Sets the tuning parameters and their possible values
-    settings.parameters = {
-      {"WGS", {64, 128, 256, 512, 1024, 2048}},
-      {"WPT", {1, 2, 4, 8}},
-      {"VW", {1, 2, 4, 8}},
-    };
-
-    // Describes how to compute the performance metrics
-    settings.metric_amount = 3 * args.n * GetBytes(args.precision);
-    settings.performance_unit = "GB/s";
-
-    return settings;
-  }
+  ;
+
+  // Buffer sizes
+  settings.size_x = args.n;
+  settings.size_y = args.n;
+
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {0, 1};
+  settings.outputs = {1};
+
+  // Sets the base thread configuration
+  settings.global_size = {args.n};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1};
+  settings.local_size_ref = {64};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"WGS"}};
+  settings.div_global = {{"WPT"},{"VW"}};
+
+  // Sets the tuning parameters and their possible values
+  settings.parameters = {
+    {"WGS", {64, 128, 256, 512, 1024, 2048}},
+    {"WPT", {1, 2, 4, 8}},
+    {"VW", {1, 2, 4, 8}},
+  };
+
+  // Describes how to compute the performance metrics
+  settings.metric_amount = 3 * args.n * GetBytes(args.precision);
+  settings.performance_unit = "GB/s";
+
+  return settings;
+}
 
-  // Tests for valid arguments
-  static void TestValidArguments(const Arguments<T> &args) {
-    if (!IsMultiple(args.n, 64)) {
-      throw std::runtime_error("'XaxpyFastest' requires 'n' to be a multiple of WGS*WPT*VW");
-    }
-  }
-  static std::vector<Constraint> SetConstraints() { return {}; }
-
-  // Sets the kernel's arguments
-  static void SetArguments(Kernel &kernel, const Arguments<T> &args,
-                           std::vector<Buffer<T>>& buffers) {
-    kernel.SetArgument(0, static_cast<int>(args.n));
-    kernel.SetArgument(1, GetRealArg(args.alpha));
-    kernel.SetArgument(2, buffers[0]()); // 0 == X vector
-    kernel.SetArgument(3, buffers[1]()); // 1 == Y vector
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int, const Arguments<T> &args) {
+  if (!IsMultiple(args.n, 64)) {
+    throw std::runtime_error("'XaxpyFastest' requires 'n' to be a multiple of WGS*WPT*VW");
   }
-};
+}
+std::vector<Constraint> SetConstraints(const int) { return {}; }
+
+// Sets the kernel's arguments
+template <typename T>
+void SetArguments(const int, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  kernel.SetArgument(0, static_cast<int>(args.n));
+  kernel.SetArgument(1, GetRealArg(args.alpha));
+  kernel.SetArgument(2, buffers[0]()); // 0 == X vector
+  kernel.SetArgument(3, buffers[1]()); // 1 == Y vector
+}
 
 // =================================================================================================
 } // namespace clblast
@@ -107,11 +103,11 @@ using double2 = clblast::double2;
 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<clblast::TuneXaxpy<half>, half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneXaxpy<float>, float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneXaxpy<double>, double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneXaxpy<float2>, float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneXaxpy<double2>, double2>(argc, argv); break;
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break;
+    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break;
+    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break;
   }
   return 0;
 }
diff --git a/src/tuning/kernels/xdot.cpp b/src/tuning/kernels/xdot.cpp
index 12350657..635d012a 100644
--- a/src/tuning/kernels/xdot.cpp
+++ b/src/tuning/kernels/xdot.cpp
@@ -21,86 +21,82 @@
 namespace clblast {
 // =================================================================================================
 
-// See comment at top of file for a description of the class
-template <typename T, int V>
-class TuneXdot {
- public:
-
-  // Settings for this kernel (default command-line arguments)
-  static TunerDefaults GetTunerDefaults() {
-    auto settings = TunerDefaults();
-    settings.options = {kArgN};
-    settings.default_n = 2*1024*1024;
-    return settings;
-  }
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgN};
+  settings.default_n = 2*1024*1024;
+  return settings;
+}
 
-  // Settings for this kernel (general)
-  static TunerSettings GetTunerSettings(const Arguments<T> &args) {
-    auto settings = TunerSettings();
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings GetTunerSettings(const int V, const Arguments<T> &args) {
+  auto settings = TunerSettings();
 
-    // Identification of the kernel
-    settings.kernel_family = "xdot_"+std::to_string(V);
-    settings.kernel_name = (V==1) ? "Xdot" : "XdotEpilogue";
-    settings.sources =
+  // Identification of the kernel
+  settings.kernel_family = "xdot_"+std::to_string(V);
+  settings.kernel_name = (V==1) ? "Xdot" : "XdotEpilogue";
+  settings.sources =
 #include "../src/kernels/level1/xdot.opencl"
-    ;
+  ;
 
-    // Buffer sizes
-    settings.size_x = args.n;
-    settings.size_y = args.n;
-    settings.size_temp = args.n; // Worst case
+  // Buffer sizes
+  settings.size_x = args.n;
+  settings.size_y = args.n;
+  settings.size_temp = args.n; // Worst case
 
-    // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
-    settings.inputs = {0, 1, 5};
-    settings.outputs = {}; // no output checking
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {0, 1, 5};
+  settings.outputs = {}; // no output checking
 
-    // Sets the base thread configuration
-    settings.global_size = (V==1) ? std::vector<size_t>{2*64} : std::vector<size_t>{1};
-    settings.global_size_ref = (V==1) ? std::vector<size_t>{2*64*64} : std::vector<size_t>{64};
-    settings.local_size = {1};
-    settings.local_size_ref = {64};
+  // Sets the base thread configuration
+  settings.global_size = (V==1) ? std::vector<size_t>{2*64} : std::vector<size_t>{1};
+  settings.global_size_ref = (V==1) ? std::vector<size_t>{2*64*64} : std::vector<size_t>{64};
+  settings.local_size = {1};
+  settings.local_size_ref = {64};
 
-    // Transforms the thread configuration based on the parameters
-    settings.mul_local = (V==1) ? TransformVector{{"WGS1"}} : TransformVector{{"WGS2"}};
-    settings.mul_global = (V==1) ? TransformVector{{"WGS1"}} : TransformVector{{"WGS2"}};
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = (V==1) ? TransformVector{{"WGS1"}} : TransformVector{{"WGS2"}};
+  settings.mul_global = (V==1) ? TransformVector{{"WGS1"}} : TransformVector{{"WGS2"}};
 
-    // Sets the tuning parameters and their possible values
-    settings.parameters = {
-      {"WGS"+std::to_string(V), {32, 64, 128, 256, 512, 1024}},
-    };
+  // Sets the tuning parameters and their possible values
+  settings.parameters = {
+    {"WGS"+std::to_string(V), {32, 64, 128, 256, 512, 1024}},
+  };
 
-    // Describes how to compute the performance metrics
-    settings.metric_amount = (V==1) ? (2*args.n + 1) * GetBytes(args.precision) : 1 * GetBytes(args.precision);
-    settings.performance_unit = (V==1) ? "GB/s" : "N/A";
+  // Describes how to compute the performance metrics
+  settings.metric_amount = (V==1) ? (2*args.n + 1) * GetBytes(args.precision) : 1 * GetBytes(args.precision);
+  settings.performance_unit = (V==1) ? "GB/s" : "N/A";
 
-    return settings;
-  }
+  return settings;
+}
 
-  // Tests for valid arguments
-  static void TestValidArguments(const Arguments<T> &) { }
-  static std::vector<Constraint> SetConstraints() { return {}; }
-
-  // Sets the kernel's arguments
-  static void SetArguments(Kernel &kernel, const Arguments<T> &args,
-                           std::vector<Buffer<T>>& buffers) {
-    if (V == 1) {
-      kernel.SetArgument(0, static_cast<int>(args.n));
-      kernel.SetArgument(1, buffers[0]()); // 0 == X vector
-      kernel.SetArgument(2, 0);
-      kernel.SetArgument(3, 1);
-      kernel.SetArgument(4, buffers[1]()); // 1 == Y vector
-      kernel.SetArgument(5, 0);
-      kernel.SetArgument(6, 1);
-      kernel.SetArgument(7, buffers[5]()); // 5 == temp; no output checking - size varies
-      kernel.SetArgument(8, static_cast<int>(false));
-    }
-    else {
-      kernel.SetArgument(0, buffers[5]()); // 5 == temp
-      kernel.SetArgument(1, buffers[0]()); // 0 == X vector; no output checking - size varies
-      kernel.SetArgument(2, 0);
-    }
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int, const Arguments<T> &) { }
+std::vector<Constraint> SetConstraints(const int) { return {}; }
+
+// Sets the kernel's arguments
+template <typename T>
+void SetArguments(const int V, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  if (V == 1) {
+    kernel.SetArgument(0, static_cast<int>(args.n));
+    kernel.SetArgument(1, buffers[0]()); // 0 == X vector
+    kernel.SetArgument(2, 0);
+    kernel.SetArgument(3, 1);
+    kernel.SetArgument(4, buffers[1]()); // 1 == Y vector
+    kernel.SetArgument(5, 0);
+    kernel.SetArgument(6, 1);
+    kernel.SetArgument(7, buffers[5]()); // 5 == temp; no output checking - size varies
+    kernel.SetArgument(8, static_cast<int>(false));
+  }
+  else {
+    kernel.SetArgument(0, buffers[5]()); // 5 == temp
+    kernel.SetArgument(1, buffers[0]()); // 0 == X vector; no output checking - size varies
+    kernel.SetArgument(2, 0);
   }
-};
+}
 
 // =================================================================================================
 } // namespace clblast
@@ -115,11 +111,11 @@ template <int V>
 void StartVariation(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<clblast::TuneXdot<half, V>, half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneXdot<float, V>, float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneXdot<double, V>, double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneXdot<float2, V>, float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneXdot<double2, V>, double2>(argc, argv); break;
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break;
+    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break;
+    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break;
   }
 }
 
diff --git a/src/tuning/kernels/xgemm.cpp b/src/tuning/kernels/xgemm.cpp
index d38ce077..b25ba302 100644
--- a/src/tuning/kernels/xgemm.cpp
+++ b/src/tuning/kernels/xgemm.cpp
@@ -22,148 +22,144 @@
 namespace clblast {
 // =================================================================================================
 
-// See comment at top of file for a description of the class
-template <typename T, int V>
-class TuneXgemm {
- public:
-
-  // Settings for this kernel (default command-line arguments)
-  static TunerDefaults GetTunerDefaults() {
-    auto settings = TunerDefaults();
-    settings.options = {kArgM, kArgN, kArgK, kArgAlpha, kArgBeta, kArgFraction,
-                        kArgHeuristicSelection, kArgPsoSwarmSize,
-                        kArgPsoInfGlobal, kArgPsoInfLocal, kArgPsoInfRandom};
-    settings.default_m = 1024;
-    settings.default_n = 1024;
-    settings.default_k = 1024;
-    settings.default_fraction = (V==1) ? 1.0 : 512.0; // test all or sample randomly
-    settings.default_num_runs = 2;
-    return settings;
-  }
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int V) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgM, kArgN, kArgK, kArgAlpha, kArgBeta, kArgFraction,
+                      kArgHeuristicSelection, kArgPsoSwarmSize,
+                      kArgPsoInfGlobal, kArgPsoInfLocal, kArgPsoInfRandom};
+  settings.default_m = 1024;
+  settings.default_n = 1024;
+  settings.default_k = 1024;
+  settings.default_fraction = (V==1) ? 1.0 : 512.0; // test all or sample randomly
+  settings.default_num_runs = 2;
+  return settings;
+}
 
-  // Settings for this kernel (general)
-  static TunerSettings GetTunerSettings(const Arguments<T> &args) {
-    auto settings = TunerSettings();
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings GetTunerSettings(const int V, const Arguments<T> &args) {
+  auto settings = TunerSettings();
 
-    // Identification of the kernel
-    settings.kernel_family = (V==1) ? "xgemm_1" : "xgemm_2";
-    settings.kernel_name = "Xgemm";
-    settings.sources =
+  // Identification of the kernel
+  settings.kernel_family = (V==1) ? "xgemm_1" : "xgemm_2";
+  settings.kernel_name = "Xgemm";
+  settings.sources =
 #include "../src/kernels/level3/xgemm_part1.opencl"
 #include "../src/kernels/level3/xgemm_part2.opencl"
 #include "../src/kernels/level3/xgemm_part3.opencl"
 #include "../src/kernels/level3/xgemm_part4.opencl"
-    ;
-
-    // Buffer sizes
-    settings.size_a = args.m * args.k;
-    settings.size_b = args.n * args.k;
-    settings.size_c = args.m * args.n;
-
-    // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
-    settings.inputs = {2, 3, 4};
-    settings.outputs = {4};
-
-    // Sets the base thread configuration
-    settings.global_size = {args.m, args.n};
-    settings.global_size_ref = settings.global_size;
-    settings.local_size = {1, 1};
-    settings.local_size_ref = {8, 8};
-
-    // Transforms the thread configuration based on the parameters
-    settings.mul_local = {{"MDIMC", "NDIMC"}};
-    settings.mul_global = {{"MDIMC", "NDIMC"}};
-    settings.div_global = {{"MWG", "NWG"}};
-
-    // Sets the tuning parameters and their possible values
-    if (V==1) { // limited subset of tuning parameters - but explorable exhaustively
-      settings.parameters = {
-        {"MWG", {16, 32, 64}},
-        {"NWG", {16, 32, 64}},
-        {"KWG", {32}},
-        {"MDIMC", {8, 16, 32}},
-        {"NDIMC", {8, 16, 32}},
-        {"MDIMA", {8, 16, 32}},
-        {"NDIMB", {8, 16, 32}},
-        {"KWI", {2}},
-        {"VWM", {1, 2, 4}},
-        {"VWN", {1, 2, 4}},
-        {"STRM", {0}},
-        {"STRN", {0}},
-        {"SA", {0, 1}},
-        {"SB", {0, 1}},
-      };
-    }
-    else { // a lot more tuning parameters - has to be sampled randomly, too much to test all
-      settings.parameters = {
-        {"MWG", {16, 32, 64, 128}},
-        {"NWG", {16, 32, 64, 128}},
-        {"KWG", {16, 32}},
-        {"MDIMC", {8, 16, 32}},
-        {"NDIMC", {8, 16, 32}},
-        {"MDIMA", {8, 16, 32}},
-        {"NDIMB", {8, 16, 32}},
-        {"KWI", {2}},
-        {"VWM", {1, 2, 4, 8}},
-        {"VWN", {1, 2, 4, 8}},
-        {"STRM", {0, 1}},
-        {"STRN", {0, 1}},
-        {"SA", {0, 1}},
-        {"SB", {0, 1}},
-      };
-    }
-
-    // Describes how to compute the performance metrics
-    settings.metric_amount = 2 * args.m * args.n * args.k;
-    settings.performance_unit = "GFLOPS";
-
-    return settings;
+  ;
+
+  // Buffer sizes
+  settings.size_a = args.m * args.k;
+  settings.size_b = args.n * args.k;
+  settings.size_c = args.m * args.n;
+
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {2, 3, 4};
+  settings.outputs = {4};
+
+  // Sets the base thread configuration
+  settings.global_size = {args.m, args.n};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1, 1};
+  settings.local_size_ref = {8, 8};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"MDIMC", "NDIMC"}};
+  settings.mul_global = {{"MDIMC", "NDIMC"}};
+  settings.div_global = {{"MWG", "NWG"}};
+
+  // Sets the tuning parameters and their possible values
+  if (V==1) { // limited subset of tuning parameters - but explorable exhaustively
+    settings.parameters = {
+      {"MWG", {16, 32, 64}},
+      {"NWG", {16, 32, 64}},
+      {"KWG", {32}},
+      {"MDIMC", {8, 16, 32}},
+      {"NDIMC", {8, 16, 32}},
+      {"MDIMA", {8, 16, 32}},
+      {"NDIMB", {8, 16, 32}},
+      {"KWI", {2}},
+      {"VWM", {1, 2, 4}},
+      {"VWN", {1, 2, 4}},
+      {"STRM", {0}},
+      {"STRN", {0}},
+      {"SA", {0, 1}},
+      {"SB", {0, 1}},
+    };
   }
-
-  // Tests for valid arguments
-  static void TestValidArguments(const Arguments<T> &) { }
-  static std::vector<Constraint> SetConstraints() {
-    auto constraints = std::vector<Constraint>();
-    auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
-    auto MultipleOfXMulY = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]*v[2]); };
-    auto MultipleOfXMulYDivZ = [] (std::vector<size_t> v) { return IsMultiple(v[0], (v[1]*v[2])/v[3]); };
-    // Requirement for unrolling the KWG loop
-    constraints.push_back({MultipleOfX, {"KWG", "KWI"}});
-    // Required for integer MWI and NWI
-    constraints.push_back({MultipleOfXMulY, {"MWG", "MDIMC", "VWM"}});
-    constraints.push_back({MultipleOfXMulY, {"NWG", "NDIMC", "VWN"}});
-    // Required for integer MWIA and NWIB
-    constraints.push_back({MultipleOfXMulY, {"MWG", "MDIMA", "VWM"}});
-    constraints.push_back({MultipleOfXMulY, {"NWG", "NDIMB", "VWN"}});
-    // KWG has to be a multiple of KDIMA = ((MDIMC*NDIMC)/(MDIMA)) and KDIMB = (...)
-    constraints.push_back({MultipleOfXMulYDivZ, {"KWG", "MDIMC", "NDIMC", "MDIMA"}});
-    constraints.push_back({MultipleOfXMulYDivZ, {"KWG", "MDIMC", "NDIMC", "NDIMB"}});
-
-    // Extra constraints for variation 1 to limit the set of options significantly
-    if (V==1) {
-      auto IsEqual = [] (std::vector<size_t> v) { return v[0] == v[1]; };
-      constraints.push_back({IsEqual, {"MDIMC", "MDIMA"}});
-      constraints.push_back({IsEqual, {"NDIMC", "NDIMB"}});
-      constraints.push_back({IsEqual, {"SA", "SB"}});
-    }
-    return constraints;
+  else { // a lot more tuning parameters - has to be sampled randomly, too much to test all
+    settings.parameters = {
+      {"MWG", {16, 32, 64, 128}},
+      {"NWG", {16, 32, 64, 128}},
+      {"KWG", {16, 32}},
+      {"MDIMC", {8, 16, 32}},
+      {"NDIMC", {8, 16, 32}},
+      {"MDIMA", {8, 16, 32}},
+      {"NDIMB", {8, 16, 32}},
+      {"KWI", {2}},
+      {"VWM", {1, 2, 4, 8}},
+      {"VWN", {1, 2, 4, 8}},
+      {"STRM", {0, 1}},
+      {"STRN", {0, 1}},
+      {"SA", {0, 1}},
+      {"SB", {0, 1}},
+    };
   }
 
-  // Sets the kernel's arguments
-  static void SetArguments(Kernel &kernel, const Arguments<T> &args,
-                           std::vector<Buffer<T>>& buffers) {
-    kernel.SetArgument(0, static_cast<int>(args.m));
-    kernel.SetArgument(1, static_cast<int>(args.n));
-    kernel.SetArgument(2, static_cast<int>(args.k));
-    kernel.SetArgument(3, GetRealArg(args.alpha));
-    kernel.SetArgument(4, GetRealArg(args.beta));
-    kernel.SetArgument(5, buffers[2]()); // 2 == A matrix
-    kernel.SetArgument(6, buffers[3]()); // 3 == B matrix
-    kernel.SetArgument(7, buffers[4]()); // 4 == C matrix
-    kernel.SetArgument(8, 0);
-    kernel.SetArgument(9, 0);
+  // Describes how to compute the performance metrics
+  settings.metric_amount = 2 * args.m * args.n * args.k;
+  settings.performance_unit = "GFLOPS";
+
+  return settings;
+}
+
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int V, const Arguments<T> &) { }
+std::vector<Constraint> SetConstraints(const int V) {
+  auto constraints = std::vector<Constraint>();
+  auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
+  auto MultipleOfXMulY = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]*v[2]); };
+  auto MultipleOfXMulYDivZ = [] (std::vector<size_t> v) { return IsMultiple(v[0], (v[1]*v[2])/v[3]); };
+  // Requirement for unrolling the KWG loop
+  constraints.push_back({MultipleOfX, {"KWG", "KWI"}});
+  // Required for integer MWI and NWI
+  constraints.push_back({MultipleOfXMulY, {"MWG", "MDIMC", "VWM"}});
+  constraints.push_back({MultipleOfXMulY, {"NWG", "NDIMC", "VWN"}});
+  // Required for integer MWIA and NWIB
+  constraints.push_back({MultipleOfXMulY, {"MWG", "MDIMA", "VWM"}});
+  constraints.push_back({MultipleOfXMulY, {"NWG", "NDIMB", "VWN"}});
+  // KWG has to be a multiple of KDIMA = ((MDIMC*NDIMC)/(MDIMA)) and KDIMB = (...)
+  constraints.push_back({MultipleOfXMulYDivZ, {"KWG", "MDIMC", "NDIMC", "MDIMA"}});
+  constraints.push_back({MultipleOfXMulYDivZ, {"KWG", "MDIMC", "NDIMC", "NDIMB"}});
+
+  // Extra constraints for variation 1 to limit the set of options significantly
+  if (V==1) {
+    auto IsEqual = [] (std::vector<size_t> v) { return v[0] == v[1]; };
+    constraints.push_back({IsEqual, {"MDIMC", "MDIMA"}});
+    constraints.push_back({IsEqual, {"NDIMC", "NDIMB"}});
+    constraints.push_back({IsEqual, {"SA", "SB"}});
   }
-};
+  return constraints;
+}
+
+// Sets the kernel's arguments
+template <typename T>
+void SetArguments(const int, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  kernel.SetArgument(0, static_cast<int>(args.m));
+  kernel.SetArgument(1, static_cast<int>(args.n));
+  kernel.SetArgument(2, static_cast<int>(args.k));
+  kernel.SetArgument(3, GetRealArg(args.alpha));
+  kernel.SetArgument(4, GetRealArg(args.beta));
+  kernel.SetArgument(5, buffers[2]()); // 2 == A matrix
+  kernel.SetArgument(6, buffers[3]()); // 3 == B matrix
+  kernel.SetArgument(7, buffers[4]()); // 4 == C matrix
+  kernel.SetArgument(8, 0);
+  kernel.SetArgument(9, 0);
+}
 
 // =================================================================================================
 } // namespace clblast
@@ -178,11 +174,11 @@ template <int V>
 void StartVariation(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<clblast::TuneXgemm<half,V>, half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneXgemm<float,V>, float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneXgemm<double,V>, double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneXgemm<float2,V>, float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneXgemm<double2,V>, double2>(argc, argv); break;
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break;
+    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break;
+    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break;
   }
 }
 
diff --git a/src/tuning/kernels/xgemm_direct.cpp b/src/tuning/kernels/xgemm_direct.cpp
index 60a983b4..0bd2e94d 100644
--- a/src/tuning/kernels/xgemm_direct.cpp
+++ b/src/tuning/kernels/xgemm_direct.cpp
@@ -22,145 +22,141 @@
 namespace clblast {
 // =================================================================================================
 
-// See comment at top of file for a description of the class
-template <typename T, int V>
-class TuneXgemmDirect {
- public:
-
-  // Settings for this kernel (default command-line arguments)
-  static TunerDefaults GetTunerDefaults() {
-    auto settings = TunerDefaults();
-    settings.options = {kArgM, kArgN, kArgK, kArgAlpha, kArgBeta, kArgFraction,
-                        kArgHeuristicSelection, kArgPsoSwarmSize,
-                        kArgPsoInfGlobal, kArgPsoInfLocal, kArgPsoInfRandom};
-    settings.default_m = 256;
-    settings.default_n = 256;
-    settings.default_k = 256;
-    settings.default_fraction = (V==1) ? 1.0 : 64.0; // test all or sample randomly
-    settings.default_num_runs = 4;
-    return settings;
-  }
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int V) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgM, kArgN, kArgK, kArgAlpha, kArgBeta, kArgFraction,
+                      kArgHeuristicSelection, kArgPsoSwarmSize,
+                      kArgPsoInfGlobal, kArgPsoInfLocal, kArgPsoInfRandom};
+  settings.default_m = 256;
+  settings.default_n = 256;
+  settings.default_k = 256;
+  settings.default_fraction = (V==1) ? 1.0 : 64.0; // test all or sample randomly
+  settings.default_num_runs = 4;
+  return settings;
+}
 
-  // Settings for this kernel (general)
-  static TunerSettings GetTunerSettings(const Arguments<T> &args) {
-    auto settings = TunerSettings();
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings GetTunerSettings(const int V, const Arguments<T> &args) {
+  auto settings = TunerSettings();
 
-    // Identification of the kernel
-    settings.kernel_family = (V==1) ? "xgemm_direct_1" : "xgemm_direct_2";
-    settings.kernel_name = "XgemmDirectTN";
-    settings.sources =
+  // Identification of the kernel
+  settings.kernel_family = (V==1) ? "xgemm_direct_1" : "xgemm_direct_2";
+  settings.kernel_name = "XgemmDirectTN";
+  settings.sources =
 #include "../src/kernels/level3/xgemm_direct_part1.opencl"
 #include "../src/kernels/level3/xgemm_direct_part2.opencl"
 #include "../src/kernels/level3/xgemm_direct_part3.opencl"
-    ;
-
-    // Buffer sizes
-    settings.size_a = args.m * args.k;
-    settings.size_b = args.n * args.k;
-    settings.size_c = args.m * args.n;
-
-    // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
-    settings.inputs = {2, 3, 4};
-    settings.outputs = {4};
-
-    // Sets the base thread configuration
-    settings.global_size = {args.m, args.n};
-    settings.global_size_ref = settings.global_size;
-    settings.local_size = {1, 1};
-    settings.local_size_ref = {8, 8};
-
-    // Transforms the thread configuration based on the parameters
-    settings.mul_local = {{"MDIMCD", "NDIMCD"}};
-    settings.mul_global = {{"MDIMCD", "NDIMCD"}};
-    settings.div_global = {{"WGD", "WGD"}};
-
-    // Sets the tuning parameters and their possible values
-    if (V==1) { // limited subset of tuning parameters - but explorable exhaustively
-      settings.parameters = {
-        {"WGD", {8, 16, 32}},
-        {"MDIMCD", {8, 16, 32}},
-        {"NDIMCD", {8, 16, 32}},
-        {"MDIMAD", {8, 16, 32}},
-        {"NDIMBD", {8, 16, 32}},
-        {"KWID", {2}},
-        {"VWMD", {1, 2, 4, 8}},
-        {"VWND", {1, 2, 4, 8}},
-        {"PADA", {1}},
-        {"PADB", {1}},
-      };
-    }
-    else { // a lot more tuning parameters - has to be sampled randomly, too much to test all
-      settings.parameters = {
-        {"WGD", {8, 16, 32, 64}},
-        {"MDIMCD", {8, 16, 32}},
-        {"NDIMCD", {8, 16, 32}},
-        {"MDIMAD", {8, 16, 32}},
-        {"NDIMBD", {8, 16, 32}},
-        {"KWID", {2, 8, 16}},
-        {"VWMD", {1, 2, 4, 8}},
-        {"VWND", {1, 2, 4, 8}},
-        {"PADA", {0, 1}},
-        {"PADB", {0, 1}},
-      };
-    }
-
-    // Describes how to compute the performance metrics
-    settings.metric_amount = 2 * args.m * args.n * args.k;
-    settings.performance_unit = "GFLOPS";
-
-    return settings;
+  ;
+
+  // Buffer sizes
+  settings.size_a = args.m * args.k;
+  settings.size_b = args.n * args.k;
+  settings.size_c = args.m * args.n;
+
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {2, 3, 4};
+  settings.outputs = {4};
+
+  // Sets the base thread configuration
+  settings.global_size = {args.m, args.n};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1, 1};
+  settings.local_size_ref = {8, 8};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"MDIMCD", "NDIMCD"}};
+  settings.mul_global = {{"MDIMCD", "NDIMCD"}};
+  settings.div_global = {{"WGD", "WGD"}};
+
+  // Sets the tuning parameters and their possible values
+  if (V==1) { // limited subset of tuning parameters - but explorable exhaustively
+    settings.parameters = {
+      {"WGD", {8, 16, 32}},
+      {"MDIMCD", {8, 16, 32}},
+      {"NDIMCD", {8, 16, 32}},
+      {"MDIMAD", {8, 16, 32}},
+      {"NDIMBD", {8, 16, 32}},
+      {"KWID", {2}},
+      {"VWMD", {1, 2, 4, 8}},
+      {"VWND", {1, 2, 4, 8}},
+      {"PADA", {1}},
+      {"PADB", {1}},
+    };
   }
-
-  // Tests for valid arguments
-  static void TestValidArguments(const Arguments<T> &) { }
-  static std::vector<Constraint> SetConstraints() {
-    auto constraints = std::vector<Constraint>();
-    auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
-    auto MultipleOfXMulY = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]*v[2]); };
-    auto MultipleOfXMulYDivZ = [] (std::vector<size_t> v) { return IsMultiple(v[0], (v[1]*v[2])/v[3]); };
-    // Requirement for unrolling the WGD loop
-    constraints.push_back({MultipleOfX, {"WGD", "KWID"}});
-    // Required for integer MWID and NWID
-    constraints.push_back({MultipleOfXMulY, {"WGD", "MDIMCD", "VWMD"}});
-    constraints.push_back({MultipleOfXMulY, {"WGD", "NDIMCD", "VWND"}});
-    // Required for integer MWIAD and NWIBD
-    constraints.push_back({MultipleOfXMulY, {"WGD", "MDIMAD", "VWMD"}});
-    constraints.push_back({MultipleOfXMulY, {"WGD", "NDIMBD", "VWND"}});
-    // WGD has to be a multiple of KDIMAD = ((MDIMCD*NDIMCD)/(MDIMAD)) and KDIMBD = (...)
-    constraints.push_back({MultipleOfXMulYDivZ, {"WGD", "MDIMCD", "NDIMCD", "MDIMAD"}});
-    constraints.push_back({MultipleOfXMulYDivZ, {"WGD", "MDIMCD", "NDIMCD", "NDIMBD"}});
-
-    // Extra constraints for variation 1 to limit the set of options significantly
-    if (V==1) {
-      auto IsEqual = [] (std::vector<size_t> v) { return v[0] == v[1]; };
-      constraints.push_back({IsEqual, {"MDIMCD", "MDIMAD"}});
-      constraints.push_back({IsEqual, {"NDIMCD", "NDIMBD"}});
-    }
-    return constraints;
+  else { // a lot more tuning parameters - has to be sampled randomly, too much to test all
+    settings.parameters = {
+      {"WGD", {8, 16, 32, 64}},
+      {"MDIMCD", {8, 16, 32}},
+      {"NDIMCD", {8, 16, 32}},
+      {"MDIMAD", {8, 16, 32}},
+      {"NDIMBD", {8, 16, 32}},
+      {"KWID", {2, 8, 16}},
+      {"VWMD", {1, 2, 4, 8}},
+      {"VWND", {1, 2, 4, 8}},
+      {"PADA", {0, 1}},
+      {"PADB", {0, 1}},
+    };
   }
 
-  // Sets the kernel's arguments
-  static void SetArguments(Kernel &kernel, const Arguments<T> &args,
-                           std::vector<Buffer<T>>& buffers) {
-    kernel.SetArgument(0, static_cast<int>(args.m));
-    kernel.SetArgument(1, static_cast<int>(args.n));
-    kernel.SetArgument(2, static_cast<int>(args.k));
-    kernel.SetArgument(3, GetRealArg(args.alpha));
-    kernel.SetArgument(4, GetRealArg(args.beta));
-    kernel.SetArgument(5, buffers[2]()); // 2 == A matrix
-    kernel.SetArgument(6, 0); // a_offset
-    kernel.SetArgument(7, static_cast<int>(args.k)); // a_ld
-    kernel.SetArgument(8, buffers[3]()); // 3 == B matrix
-    kernel.SetArgument(9, 0); // b_offset
-    kernel.SetArgument(10, static_cast<int>(args.n)); // b_ld
-    kernel.SetArgument(11, buffers[4]()); // 4 == C matrix
-    kernel.SetArgument(12, 0); // c_offset
-    kernel.SetArgument(13, static_cast<int>(args.n)); // c_ld
-    kernel.SetArgument(14, 1); // c_do_transpose
-    kernel.SetArgument(15, 0); // a_conjugate
-    kernel.SetArgument(16, 0); // b_conjugate
+  // Describes how to compute the performance metrics
+  settings.metric_amount = 2 * args.m * args.n * args.k;
+  settings.performance_unit = "GFLOPS";
+
+  return settings;
+}
+
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int, const Arguments<T> &) { }
+std::vector<Constraint> SetConstraints(const int V) {
+  auto constraints = std::vector<Constraint>();
+  auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
+  auto MultipleOfXMulY = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]*v[2]); };
+  auto MultipleOfXMulYDivZ = [] (std::vector<size_t> v) { return IsMultiple(v[0], (v[1]*v[2])/v[3]); };
+  // Requirement for unrolling the WGD loop
+  constraints.push_back({MultipleOfX, {"WGD", "KWID"}});
+  // Required for integer MWID and NWID
+  constraints.push_back({MultipleOfXMulY, {"WGD", "MDIMCD", "VWMD"}});
+  constraints.push_back({MultipleOfXMulY, {"WGD", "NDIMCD", "VWND"}});
+  // Required for integer MWIAD and NWIBD
+  constraints.push_back({MultipleOfXMulY, {"WGD", "MDIMAD", "VWMD"}});
+  constraints.push_back({MultipleOfXMulY, {"WGD", "NDIMBD", "VWND"}});
+  // WGD has to be a multiple of KDIMAD = ((MDIMCD*NDIMCD)/(MDIMAD)) and KDIMBD = (...)
+  constraints.push_back({MultipleOfXMulYDivZ, {"WGD", "MDIMCD", "NDIMCD", "MDIMAD"}});
+  constraints.push_back({MultipleOfXMulYDivZ, {"WGD", "MDIMCD", "NDIMCD", "NDIMBD"}});
+
+  // Extra constraints for variation 1 to limit the set of options significantly
+  if (V==1) {
+    auto IsEqual = [] (std::vector<size_t> v) { return v[0] == v[1]; };
+    constraints.push_back({IsEqual, {"MDIMCD", "MDIMAD"}});
+    constraints.push_back({IsEqual, {"NDIMCD", "NDIMBD"}});
   }
-};
+  return constraints;
+}
+
+// Sets the kernel's arguments
+template <typename T>
+void SetArguments(const int, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  kernel.SetArgument(0, static_cast<int>(args.m));
+  kernel.SetArgument(1, static_cast<int>(args.n));
+  kernel.SetArgument(2, static_cast<int>(args.k));
+  kernel.SetArgument(3, GetRealArg(args.alpha));
+  kernel.SetArgument(4, GetRealArg(args.beta));
+  kernel.SetArgument(5, buffers[2]()); // 2 == A matrix
+  kernel.SetArgument(6, 0); // a_offset
+  kernel.SetArgument(7, static_cast<int>(args.k)); // a_ld
+  kernel.SetArgument(8, buffers[3]()); // 3 == B matrix
+  kernel.SetArgument(9, 0); // b_offset
+  kernel.SetArgument(10, static_cast<int>(args.n)); // b_ld
+  kernel.SetArgument(11, buffers[4]()); // 4 == C matrix
+  kernel.SetArgument(12, 0); // c_offset
+  kernel.SetArgument(13, static_cast<int>(args.n)); // c_ld
+  kernel.SetArgument(14, 1); // c_do_transpose
+  kernel.SetArgument(15, 0); // a_conjugate
+  kernel.SetArgument(16, 0); // b_conjugate
+}
 
 // =================================================================================================
 } // namespace clblast
@@ -175,11 +171,11 @@ template <int V>
 void StartVariation(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<clblast::TuneXgemmDirect<half,V>, half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneXgemmDirect<float,V>, float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneXgemmDirect<double,V>, double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneXgemmDirect<float2,V>, float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneXgemmDirect<double2,V>, double2>(argc, argv); break;
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break;
+    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break;
+    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break;
   }
 }
 
diff --git a/src/tuning/kernels/xgemv.cpp b/src/tuning/kernels/xgemv.cpp
index 3eadd32b..c2ee1df3 100644
--- a/src/tuning/kernels/xgemv.cpp
+++ b/src/tuning/kernels/xgemv.cpp
@@ -23,119 +23,115 @@
 namespace clblast {
 // =================================================================================================
 
-// See comment at top of file for a description of the class
-template <typename T, int V>
-class TuneXgemv {
- public:
-
-  // Settings for this kernel (default command-line arguments)
-  static TunerDefaults GetTunerDefaults() {
-    auto settings = TunerDefaults();
-    settings.options = {kArgM, kArgN, kArgAlpha, kArgBeta};
-    settings.default_m = 2048;
-    settings.default_n = 2048;
-    return settings;
-  }
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgM, kArgN, kArgAlpha, kArgBeta};
+  settings.default_m = 2048;
+  settings.default_n = 2048;
+  return settings;
+}
 
-  // Settings for this kernel (general)
-  static TunerSettings GetTunerSettings(const Arguments<T> &args) {
-    auto settings = TunerSettings();
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings GetTunerSettings(const int V, const Arguments<T> &args) {
+  auto settings = TunerSettings();
 
-    // Identification of the kernel
-    settings.kernel_family = (V==1) ? "xgemv" : ((V==2) ? "xgemv_fast" : "xgemv_fast_rot");
-    settings.kernel_name = (V==1) ? "Xgemv" : ((V==2) ? "XgemvFast" : "XgemvFastRot");
-    settings.sources =
+  // Identification of the kernel
+  settings.kernel_family = (V==1) ? "xgemv" : ((V==2) ? "xgemv_fast" : "xgemv_fast_rot");
+  settings.kernel_name = (V==1) ? "Xgemv" : ((V==2) ? "XgemvFast" : "XgemvFastRot");
+  settings.sources =
 #include "../src/kernels/level2/xgemv.opencl"
 #include "../src/kernels/level2/xgemv_fast.opencl"
-    ;
-
-    // Buffer sizes
-    settings.size_x = args.n;
-    settings.size_y = args.m;
-    settings.size_a = args.m * args.n;
-
-    // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
-    settings.inputs = {0, 1, 2};
-    settings.outputs = {1};
-
-    // Sets the base thread configuration
-    settings.global_size = {args.m};
-    settings.global_size_ref = settings.global_size;
-    settings.local_size = {1};
-    settings.local_size_ref = {64};
-
-    // Transforms the thread configuration based on the parameters
-    settings.mul_local = {{"WGS"+std::to_string(V)}};
-    settings.div_global = (V==1 || V==2) ? TransformVector{{"WPT"+std::to_string(V)}} : TransformVector{};
-
-    // Sets the tuning parameters and their possible values
-    if (V==1) {
-      settings.parameters = {
-        {"WGS"+std::to_string(V), {32, 64, 128, 256}},
-        {"WPT"+std::to_string(V), {1, 2, 4}},
-      };
-    }
-    if (V==2) {
-      settings.parameters = {
-        {"WGS"+std::to_string(V), {16, 32, 64, 128, 256}},
-        {"WPT"+std::to_string(V), {1, 2, 4}},
-        {"VW"+std::to_string(V), {1, 2, 4, 8}},
-      };
-    }
-    if (V==3) {
-      settings.parameters = {
-        {"WGS"+std::to_string(V), {16, 32, 64, 128}},
-        {"WPT"+std::to_string(V), {1, 2, 4, 8, 16, 32}},
-        {"VW"+std::to_string(V), {1, 2, 4, 8}},
-      };
-    }
-
-    // Describes how to compute the performance metrics
-    settings.metric_amount = (args.m*args.n + 2*args.m + args.n) * GetBytes(args.precision);
-    settings.performance_unit = "GB/s";
-
-    return settings;
+  ;
+
+  // Buffer sizes
+  settings.size_x = args.n;
+  settings.size_y = args.m;
+  settings.size_a = args.m * args.n;
+
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {0, 1, 2};
+  settings.outputs = {1};
+
+  // Sets the base thread configuration
+  settings.global_size = {args.m};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1};
+  settings.local_size_ref = {64};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"WGS"+std::to_string(V)}};
+  settings.div_global = (V==1 || V==2) ? TransformVector{{"WPT"+std::to_string(V)}} : TransformVector{};
+
+  // Sets the tuning parameters and their possible values
+  if (V==1) {
+    settings.parameters = {
+      {"WGS"+std::to_string(V), {32, 64, 128, 256}},
+      {"WPT"+std::to_string(V), {1, 2, 4}},
+    };
   }
-
-  // Tests for valid arguments
-  static void TestValidArguments(const Arguments<T> &) { }
-  static std::vector<Constraint> SetConstraints() {
-    auto constraints = std::vector<Constraint>();
-    if (V==2 || V==3) {
-      auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
-      constraints.push_back({MultipleOfX, {"WPT"+std::to_string(V), "VW"+std::to_string(V)}});
-    }
-    if (V==3) {
-      auto LargerOrEqual = [] (std::vector<size_t> v) { return v[0] >= v[1]; };
-      constraints.push_back({LargerOrEqual, {"WGS"+std::to_string(V), "WPT"+std::to_string(V)}});
-    }
-    return constraints;
+  if (V==2) {
+    settings.parameters = {
+      {"WGS"+std::to_string(V), {16, 32, 64, 128, 256}},
+      {"WPT"+std::to_string(V), {1, 2, 4}},
+      {"VW"+std::to_string(V), {1, 2, 4, 8}},
+    };
   }
+  if (V==3) {
+    settings.parameters = {
+      {"WGS"+std::to_string(V), {16, 32, 64, 128}},
+      {"WPT"+std::to_string(V), {1, 2, 4, 8, 16, 32}},
+      {"VW"+std::to_string(V), {1, 2, 4, 8}},
+    };
+  }
+
+  // Describes how to compute the performance metrics
+  settings.metric_amount = (args.m*args.n + 2*args.m + args.n) * GetBytes(args.precision);
+  settings.performance_unit = "GB/s";
 
-  // Sets the kernel's arguments
-  static void SetArguments(Kernel &kernel, const Arguments<T> &args,
-                           std::vector<Buffer<T>>& buffers) {
-    auto a_rotated = (V==3) ? 1 : 0;
-    kernel.SetArgument(0, static_cast<int>(args.m));
-    kernel.SetArgument(1, static_cast<int>(args.n));
-    kernel.SetArgument(2, GetRealArg(args.alpha));
-    kernel.SetArgument(3, GetRealArg(args.beta));
-    kernel.SetArgument(4, a_rotated);
-    kernel.SetArgument(5, buffers[2]()); // 2 == A matrix
-    kernel.SetArgument(6, 0);
-    kernel.SetArgument(7, static_cast<int>(args.m));
-    kernel.SetArgument(8, buffers[0]()); // 0 == X vector
-    kernel.SetArgument(9, 0);
-    kernel.SetArgument(10, 1);
-    kernel.SetArgument(11, buffers[1]()); // 1 == Y vector
-    kernel.SetArgument(12, 0);
-    kernel.SetArgument(13, 1);
-    kernel.SetArgument(14, 0); // Conjugate transpose
-    kernel.SetArgument(15, 0); // Additional parameter
-    kernel.SetArgument(16, 0); // Banded 'kl'
-    kernel.SetArgument(17, 0); // Banded 'ku'
+  return settings;
+}
+
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int, const Arguments<T> &) { }
+std::vector<Constraint> SetConstraints(const int V) {
+  auto constraints = std::vector<Constraint>();
+  if (V==2 || V==3) {
+    auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
+    constraints.push_back({MultipleOfX, {"WPT"+std::to_string(V), "VW"+std::to_string(V)}});
+  }
+  if (V==3) {
+    auto LargerOrEqual = [] (std::vector<size_t> v) { return v[0] >= v[1]; };
+    constraints.push_back({LargerOrEqual, {"WGS"+std::to_string(V), "WPT"+std::to_string(V)}});
   }
-};
+  return constraints;
+}
+
+// Sets the kernel's arguments
+template <typename T>
+void SetArguments(const int V, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  auto a_rotated = (V==3) ? 1 : 0;
+  kernel.SetArgument(0, static_cast<int>(args.m));
+  kernel.SetArgument(1, static_cast<int>(args.n));
+  kernel.SetArgument(2, GetRealArg(args.alpha));
+  kernel.SetArgument(3, GetRealArg(args.beta));
+  kernel.SetArgument(4, a_rotated);
+  kernel.SetArgument(5, buffers[2]()); // 2 == A matrix
+  kernel.SetArgument(6, 0);
+  kernel.SetArgument(7, static_cast<int>(args.m));
+  kernel.SetArgument(8, buffers[0]()); // 0 == X vector
+  kernel.SetArgument(9, 0);
+  kernel.SetArgument(10, 1);
+  kernel.SetArgument(11, buffers[1]()); // 1 == Y vector
+  kernel.SetArgument(12, 0);
+  kernel.SetArgument(13, 1);
+  kernel.SetArgument(14, 0); // Conjugate transpose
+  kernel.SetArgument(15, 0); // Additional parameter
+  kernel.SetArgument(16, 0); // Banded 'kl'
+  kernel.SetArgument(17, 0); // Banded 'ku'
+}
 
 // =================================================================================================
 } // namespace clblast
@@ -150,11 +146,11 @@ template <int V>
 void StartVariation(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<clblast::TuneXgemv<half,V>, half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneXgemv<float,V>, float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneXgemv<double,V>, double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneXgemv<float2,V>, float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneXgemv<double2,V>, double2>(argc, argv); break;
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break;
+    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break;
+    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, V, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break;
   }
 }
 
diff --git a/src/tuning/kernels/xger.cpp b/src/tuning/kernels/xger.cpp
index 745e553f..a88fb5d6 100644
--- a/src/tuning/kernels/xger.cpp
+++ b/src/tuning/kernels/xger.cpp
@@ -20,87 +20,83 @@
 namespace clblast {
 // =================================================================================================
 
-// See comment at top of file for a description of the class
-template <typename T>
-class TuneXger {
- public:
-
-  // Settings for this kernel (default command-line arguments)
-  static TunerDefaults GetTunerDefaults() {
-    auto settings = TunerDefaults();
-    settings.options = {kArgM, kArgN, kArgAlpha};
-    settings.default_m = 1024;
-    settings.default_n = 1024;
-    return settings;
-  }
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgM, kArgN, kArgAlpha};
+  settings.default_m = 1024;
+  settings.default_n = 1024;
+  return settings;
+}
 
-  // Settings for this kernel (general)
-  static TunerSettings GetTunerSettings(const Arguments<T> &args) {
-    auto settings = TunerSettings();
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings GetTunerSettings(const int, const Arguments<T> &args) {
+  auto settings = TunerSettings();
 
-    // Identification of the kernel
-    settings.kernel_family = "xger";
-    settings.kernel_name = "Xger";
-    settings.sources =
+  // Identification of the kernel
+  settings.kernel_family = "xger";
+  settings.kernel_name = "Xger";
+  settings.sources =
 #include "../src/kernels/level2/level2.opencl"
 #include "../src/kernels/level2/xger.opencl"
-    ;
-
-    // Buffer sizes
-    settings.size_x = args.m;
-    settings.size_y = args.n;
-    settings.size_a = args.m * args.n;
-
-    // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
-    settings.inputs = {0, 1, 2};
-    settings.outputs = {2};
-
-    // Sets the base thread configuration
-    settings.global_size = {args.m, args.n};
-    settings.global_size_ref = settings.global_size;
-    settings.local_size = {1, 1};
-    settings.local_size_ref = {8, 8};
-
-    // Transforms the thread configuration based on the parameters
-    settings.mul_local = {{"WGS1", "WGS2"}};
-    settings.div_global = {{"WPT", "WPT"}};
-
-    // Sets the tuning parameters and their possible values
-    settings.parameters = {
-      {"WGS1", {4, 8, 16, 32, 64, 128, 256, 512}},
-      {"WGS2", {1, 2, 4, 8, 16, 32, 64, 128, 256}},
-      {"WPT", {1, 2, 4}},
-    };
-
-    // Describes how to compute the performance metrics
-    settings.metric_amount = (2*args.m*args.n + args.m + args.n) * GetBytes(args.precision);
-    settings.performance_unit = "GB/s";
-
-    return settings;
-  }
+  ;
+
+  // Buffer sizes
+  settings.size_x = args.m;
+  settings.size_y = args.n;
+  settings.size_a = args.m * args.n;
+
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {0, 1, 2};
+  settings.outputs = {2};
+
+  // Sets the base thread configuration
+  settings.global_size = {args.m, args.n};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1, 1};
+  settings.local_size_ref = {8, 8};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"WGS1", "WGS2"}};
+  settings.div_global = {{"WPT", "WPT"}};
+
+  // Sets the tuning parameters and their possible values
+  settings.parameters = {
+    {"WGS1", {4, 8, 16, 32, 64, 128, 256, 512}},
+    {"WGS2", {1, 2, 4, 8, 16, 32, 64, 128, 256}},
+    {"WPT", {1, 2, 4}},
+  };
+
+  // Describes how to compute the performance metrics
+  settings.metric_amount = (2*args.m*args.n + args.m + args.n) * GetBytes(args.precision);
+  settings.performance_unit = "GB/s";
+
+  return settings;
+}
 
-  // Tests for valid arguments
-  static void TestValidArguments(const Arguments<T> &) { }
-  static std::vector<Constraint> SetConstraints() { return {}; }
-
-  // Sets the kernel's arguments
-  static void SetArguments(Kernel &kernel, const Arguments<T> &args,
-                           std::vector<Buffer<T>>& buffers) {
-    kernel.SetArgument(0, static_cast<int>(args.m));
-    kernel.SetArgument(1, static_cast<int>(args.n));
-    kernel.SetArgument(2, GetRealArg(args.alpha));
-    kernel.SetArgument(3, buffers[0]()); // 0 == X vector
-    kernel.SetArgument(4, 0); // x_offset
-    kernel.SetArgument(5, 1); // x_increment
-    kernel.SetArgument(6, buffers[1]()); // 1 == Y vector
-    kernel.SetArgument(7, 0); // y_offset
-    kernel.SetArgument(8, 1); // y_increment
-    kernel.SetArgument(9, buffers[2]()); // 2 == A matrix
-    kernel.SetArgument(10, 0); // a_offset
-    kernel.SetArgument(11, static_cast<int>(args.m)); // a_ld
-    kernel.SetArgument(12, 0); // a_is_rowmajor
-  }
-};
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int, const Arguments<T> &) { }
+std::vector<Constraint> SetConstraints(const int) { return {}; }
+
+// Sets the kernel's arguments
+template <typename T>
+void SetArguments(const int, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  kernel.SetArgument(0, static_cast<int>(args.m));
+  kernel.SetArgument(1, static_cast<int>(args.n));
+  kernel.SetArgument(2, GetRealArg(args.alpha));
+  kernel.SetArgument(3, buffers[0]()); // 0 == X vector
+  kernel.SetArgument(4, 0); // x_offset
+  kernel.SetArgument(5, 1); // x_increment
+  kernel.SetArgument(6, buffers[1]()); // 1 == Y vector
+  kernel.SetArgument(7, 0); // y_offset
+  kernel.SetArgument(8, 1); // y_increment
+  kernel.SetArgument(9, buffers[2]()); // 2 == A matrix
+  kernel.SetArgument(10, 0); // a_offset
+  kernel.SetArgument(11, static_cast<int>(args.m)); // a_ld
+  kernel.SetArgument(12, 0); // a_is_rowmajor
+}
 
 // =================================================================================================
 } // namespace clblast
@@ -114,11 +110,11 @@ using double2 = clblast::double2;
 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<clblast::TuneXger<half>, half>(argc, argv); break;
-    case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneXger<float>, float>(argc, argv); break;
-    case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneXger<double>, double>(argc, argv); break;
-    case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneXger<float2>, float2>(argc, argv); break;
-    case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneXger<double2>, double2>(argc, argv); break;
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<half>, clblast::TestValidArguments<half>, clblast::SetConstraints, clblast::SetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float>, clblast::TestValidArguments<float>, clblast::SetConstraints, clblast::SetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double>, clblast::TestValidArguments<double>, clblast::SetConstraints, clblast::SetArguments<double>); break;
+    case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<float2>, clblast::TestValidArguments<float2>, clblast::SetConstraints, clblast::SetArguments<float2>); break;
+    case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::GetTunerDefaults, clblast::GetTunerSettings<double2>, clblast::TestValidArguments<double2>, clblast::SetConstraints, clblast::SetArguments<double2>); break;
   }
   return 0;
 }
diff --git a/src/tuning/tuning.cpp b/src/tuning/tuning.cpp
index 935ab257..c8532b36 100644
--- a/src/tuning/tuning.cpp
+++ b/src/tuning/tuning.cpp
@@ -85,5 +85,293 @@ void print_separator(const size_t parameters_size) {
   printf("-x----------------x--------------x--------x-------------------x\n");
 }
 
+// =================================================================================================
+
+template <typename T>
+void Tuner(int argc, char* argv[], const int V,
+           GetTunerDefaultsFunc GetTunerDefaults,
+           GetTunerSettingsFunc<T> GetTunerSettings,
+           TestValidArgumentsFunc<T> TestValidArguments,
+           SetConstraintsFunc SetConstraints,
+           SetArgumentsFunc<T> SetArguments) {
+  constexpr auto kSeed = 42; // fixed seed for reproducibility
+
+  // Sets the parameters and platform/device for which to tune (command-line options)
+  const TunerDefaults defaults = GetTunerDefaults(V);
+  auto command_line_args = RetrieveCommandLineArguments(argc, argv);
+  auto help = std::string{"* Options given/available:\n"};
+  auto args = Arguments<T>{};
+  args.platform_id = GetArgument(command_line_args, help, kArgPlatform, ConvertArgument(std::getenv("CLBLAST_PLATFORM"), size_t{0}));
+  args.device_id   = GetArgument(command_line_args, help, kArgDevice, ConvertArgument(std::getenv("CLBLAST_DEVICE"), size_t{0}));
+  args.precision   = GetArgument(command_line_args, help, kArgPrecision, Precision::kSingle);
+  for (auto &o: defaults.options) {
+    if (o == kArgM)        { args.m        = GetArgument(command_line_args, help, kArgM, defaults.default_m); }
+    if (o == kArgN)        { args.n        = GetArgument(command_line_args, help, kArgN, defaults.default_n); }
+    if (o == kArgK)        { args.k        = GetArgument(command_line_args, help, kArgK, defaults.default_k); }
+    if (o == kArgAlpha)    { args.alpha    = GetArgument(command_line_args, help, kArgAlpha, GetScalar<T>()); }
+    if (o == kArgBeta)     { args.beta     = GetArgument(command_line_args, help, kArgBeta, GetScalar<T>()); }
+    if (o == kArgBatchCount) { args.batch_count = GetArgument(command_line_args, help, kArgBatchCount, defaults.default_batch_count); }
+  }
+  args.fraction = GetArgument(command_line_args, help, kArgFraction, defaults.default_fraction);
+  args.num_runs = GetArgument(command_line_args, help, kArgNumRuns, defaults.default_num_runs);
+  const auto max_l2_norm = GetArgument(command_line_args, help, kArgMaxL2Norm, 1.0e-4);
+  printf("%s\n", help.c_str());
+  const TunerSettings settings = GetTunerSettings(V, args);
+
+  // Tests validity of the given arguments
+  TestValidArguments(V, args);
+
+  // Initializes OpenCL
+  const auto platform = Platform(args.platform_id);
+  const auto device = Device(platform, args.device_id);
+  const auto context = Context(device);
+
+  // Tests for validity of the precision and retrieves properties
+  if (!PrecisionSupported<T>(device)) {
+    printf("* Unsupported precision, skipping this tuning run\n\n");
+    return;
+  }
+  const auto device_type = GetDeviceType(device);
+  const auto device_vendor = GetDeviceVendor(device);
+  const auto device_architecture = GetDeviceArchitecture(device);
+  const auto device_name = GetDeviceName(device);
+
+  // Creates input buffers with random data
+  const auto buffer_sizes = std::vector<size_t>{
+      settings.size_x, settings.size_y,
+      settings.size_a, settings.size_b, settings.size_c,
+      settings.size_temp
+  };
+  std::mt19937 mt(kSeed);
+  std::uniform_real_distribution<double> dist(kTestDataLowerLimit, kTestDataUpperLimit);
+  auto source_buffers = std::vector<std::vector<T>>();
+  auto reference_buffers = std::vector<std::vector<T>>();
+  auto result_buffers = std::vector<std::vector<T>>();
+  auto device_buffers = std::vector<Buffer<T>>();
+  for (const auto size : buffer_sizes) {
+    auto host_buffer = std::vector<T>(size);
+    PopulateVector(host_buffer, mt, dist);
+    source_buffers.push_back(host_buffer);
+    reference_buffers.push_back(std::vector<T>(size));
+    result_buffers.push_back(std::vector<T>(size));
+    device_buffers.push_back(Buffer<T>(context, size));
+  }
+
+  // Sets the tunable parameters and their possible values
+  auto configurations = SetConfigurations(settings.parameters, SetConstraints(V));
+  printf("* Found %s%zu configuration(s)%s\n",
+         kPrintMessage.c_str(), configurations.size(), kPrintEnd.c_str());
+
+  // Select the search method (full search or a random fraction)
+  if (args.fraction != 0.0 && args.fraction != 1.0) {
+    const auto new_size = static_cast<size_t>(configurations.size() / args.fraction);
+    auto rng = std::default_random_engine{};
+    std::shuffle(std::begin(configurations), std::end(configurations), rng);
+    configurations.resize(new_size);
+    printf("* Exploring a random subset of %s%zu configuration(s)%s\n",
+           kPrintMessage.c_str(), configurations.size(), kPrintEnd.c_str());
+  }
+
+  // Prints information about the parameters
+  printf("* Parameters explored: ");
+  for (const auto& parameter : settings.parameters) { printf("%s ", parameter.first.c_str()); }
+  printf("\n");
+
+  // Prints the header of the table
+  printf("\n");
+  printf("|   ID | total |");
+  for (auto i = size_t{0}; i < settings.parameters.size() - 1; ++i) { printf("     "); }
+  printf("param |       compiles |         time | %6s |            status |\n", settings.performance_unit.c_str());
+  print_separator(settings.parameters.size());
+
+  // First runs a reference example to compare against
+  try {
+    auto queue = Queue(context, device);
+    printf("|  ref |     - |");
+    for (auto i = size_t{0}; i < settings.parameters.size() - 1; ++i) { printf("     "); }
+    printf("    - |");
+
+
+    // Sets the input
+    for (const auto id : settings.inputs) {
+      device_buffers[id].Write(queue, buffer_sizes[id], source_buffers[id]);
+    }
+
+    // Compiles the kernel
+    auto compiler_options = std::vector<std::string>();
+    const auto program = CompileFromSource(settings.sources, args.precision, settings.kernel_name,
+                                           device, context, compiler_options, 0);
+    auto kernel = Kernel(program, settings.kernel_name);
+    SetArguments(V, kernel, args, device_buffers);
+    printf("             %sOK%s |", kPrintSuccess.c_str(), kPrintEnd.c_str());
+
+    // Runs the kernel
+    const auto time_ms = TimeKernel(args.num_runs, kernel, queue, device,
+                                    settings.global_size_ref, settings.local_size_ref);
+    printf("      - |");
+    if (time_ms == -1.0) { throw std::runtime_error("Error in reference implementation"); }
+
+    // Saves the result
+    for (const auto id : settings.outputs) {
+      device_buffers[id].Read(queue, buffer_sizes[id], reference_buffers[id]);
+    }
+    printf("      %sreference OK%s |\n", kPrintSuccess.c_str(), kPrintEnd.c_str());
+  }
+  catch (...) {
+    const auto status_code = DispatchExceptionCatchAll(true);
+    printf("* Exception caught with status %d while running the reference, aborting\n",
+           static_cast<int>(status_code));
+    return;
+  }
+  print_separator(settings.parameters.size());
+
+  // Starts the tuning process
+  auto results = std::vector<TuningResult>();
+  for (auto config_id = size_t{0}; config_id < configurations.size(); ++config_id) {
+    try {
+      auto queue = Queue(context, device);
+
+      auto configuration = configurations[config_id];
+      printf("| %4zu | %5zu |", config_id + 1, configurations.size());
+      for (const auto& parameter : settings.parameters) {
+        printf("%5zu", configuration.at(parameter.first));
+      }
+      printf(" |");
+
+      // Sets the input
+      for (const auto id : settings.inputs) {
+        device_buffers[id].Write(queue, buffer_sizes[id], source_buffers[id]);
+      }
+
+      // Sets the thread configuration
+      const auto global = SetThreadConfiguration(configuration, settings.global_size,
+                                                 settings.mul_global, settings.div_global);
+      const auto local = SetThreadConfiguration(configuration, settings.local_size,
+                                                settings.mul_local, settings.div_local);
+
+      // Sets the parameters for this configuration
+      auto kernel_source = std::string{""};
+      for (const auto &parameter : configuration) {
+        kernel_source += "#define " + parameter.first + " " + ToString(parameter.second) + "\n";
+      }
+      kernel_source += settings.sources;
+
+      // Compiles the kernel
+      const auto start_time = std::chrono::steady_clock::now();
+      auto compiler_options = std::vector<std::string>();
+      const auto program = CompileFromSource(kernel_source, args.precision, settings.kernel_name,
+                                             device, context, compiler_options, 0, true);
+      auto kernel = Kernel(program, settings.kernel_name);
+      const auto elapsed_time = std::chrono::steady_clock::now() - start_time;
+      const auto timing = std::chrono::duration<double,std::milli>(elapsed_time).count();
+      printf("   %sOK%s  %5.0lf ms |", kPrintSuccess.c_str(), kPrintEnd.c_str(), timing);
+
+      // Runs the kernel
+      SetArguments(V, kernel, args, device_buffers);
+      const auto time_ms = TimeKernel(args.num_runs, kernel, queue, device, global, local);
+
+      // Kernel run was not successful
+      if (time_ms == -1.0) {
+        printf("      - |");
+        printf("   %sinvalid config.%s |", kPrintError.c_str(), kPrintEnd.c_str());
+        printf(" <-- skipping\n");
+        continue;
+      }
+
+      // Compares the results
+      auto l2_error = 0.0;
+      for (const auto id : settings.outputs) {
+        device_buffers[id].Read(queue, buffer_sizes[id], result_buffers[id]);
+        for (auto index = size_t{0}; index<buffer_sizes[id]; ++index) {
+          const auto diff = SquaredDifference(result_buffers[id][index], reference_buffers[id][index]);
+          l2_error += diff;
+        }
+        l2_error /= static_cast<double>(buffer_sizes[id]);
+        if (std::isnan(l2_error) || l2_error > max_l2_norm) {
+          printf("      - |");
+          printf(" %sL2 error %8.2e%s |", kPrintError.c_str(), l2_error, kPrintEnd.c_str());
+          throw std::runtime_error("L2 error too large");
+        }
+      }
+
+      // All was OK
+      configuration["PRECISION"] = static_cast<size_t>(args.precision);
+      results.push_back(TuningResult{settings.kernel_name, time_ms, configuration});
+      printf(" %6.1lf |", settings.metric_amount / (time_ms * 1.0e6));
+      printf("     %sresults match%s |\n", kPrintSuccess.c_str(), kPrintEnd.c_str());
+    }
+    catch (CLCudaAPIBuildError) {
+      const auto status_code = DispatchExceptionCatchAll(true);
+      printf("  %scompilation error: %5d%s     |",
+             kPrintError.c_str(), static_cast<int>(status_code), kPrintEnd.c_str());
+      printf("      - |                 - | <-- skipping\n");
+    }
+    catch (...) {
+      const auto status_code = DispatchExceptionCatchAll(true);
+      if (status_code != StatusCode::kUnknownError) {
+        printf("   %serror code %d%s |",
+               kPrintError.c_str(), static_cast<int>(status_code), kPrintEnd.c_str());
+      }
+      printf(" <-- skipping\n");
+    }
+  }
+
+  // Completed the tuning process
+  print_separator(settings.parameters.size());
+  printf("\n");
+  if (results.size() == 0) { return; }
+
+  // Computes the best results
+  auto comparison = [](const TuningResult& lhs, const TuningResult& rhs) { return lhs.score < rhs.score; };
+  const auto best_configuration = std::min_element(results.begin(), results.end(), comparison);
+  const auto best_time_ms = best_configuration->score;
+  if (best_time_ms == 0.0) { return; }
+
+  // Also prints the performance of the best-case in terms of GB/s or GFLOPS
+  printf("\n");
+  printf("* Found best result %.2lf ms", best_time_ms);
+  printf(": %.1lf %s\n", settings.metric_amount / (best_time_ms * 1.0e6),
+         settings.performance_unit.c_str());
+  printf("* Best parameters: ");
+  auto best_string = std::string{""};
+  auto i = size_t{0};
+  for (const auto config : best_configuration->config) {
+    best_string += "" + config.first + "=" + ToString(config.second);
+    if (i < best_configuration->config.size() - 1) { best_string += " "; }
+    ++i;
+  }
+  printf("%s\n\n", best_string.c_str());
+
+  // Outputs the results as JSON to disk, including some meta-data
+  auto precision_string = std::to_string(static_cast<size_t>(args.precision));
+  auto metadata = std::vector<std::pair<std::string,std::string>>{
+    {"kernel_family", settings.kernel_family},
+    {"precision", precision_string},
+    {"best_kernel", best_configuration->name},
+    {"best_time", ToString(best_configuration->score)},
+    {"best_parameters", best_string}
+  };
+  for (auto &o: defaults.options) {
+    if (o == kArgM)     { metadata.push_back({"arg_m", ToString(args.m)}); }
+    if (o == kArgN)     { metadata.push_back({"arg_n", ToString(args.n)}); }
+    if (o == kArgK)     { metadata.push_back({"arg_k", ToString(args.k)}); }
+    if (o == kArgAlpha) { metadata.push_back({"arg_alpha", ToString(args.alpha)}); }
+    if (o == kArgBeta)  { metadata.push_back({"arg_beta", ToString(args.beta)}); }
+    if (o == kArgBatchCount) { metadata.push_back({"arg_batch_count", ToString(args.batch_count)}); }
+  }
+  PrintTimingsToFileAsJSON("clblast_" + settings.kernel_family + "_" + precision_string + ".json",
+                           device, platform, metadata, results);
+
+  printf("* Completed tuning process\n");
+  printf("\n");
+}
+
+// Compiles the above function
+template void Tuner<half>(int argc, char* argv[], const int V, GetTunerDefaultsFunc GetTunerDefaults, GetTunerSettingsFunc<half> GetTunerSettings, TestValidArgumentsFunc<half> TestValidArguments, SetConstraintsFunc SetConstraints, SetArgumentsFunc<half> SetArguments);
+template void Tuner<float>(int argc, char* argv[], const int V, GetTunerDefaultsFunc GetTunerDefaults, GetTunerSettingsFunc<float> GetTunerSettings, TestValidArgumentsFunc<float> TestValidArguments, SetConstraintsFunc SetConstraints, SetArgumentsFunc<float> SetArguments);
+template void Tuner<double>(int argc, char* argv[], const int V, GetTunerDefaultsFunc GetTunerDefaults, GetTunerSettingsFunc<double> GetTunerSettings, TestValidArgumentsFunc<double> TestValidArguments, SetConstraintsFunc SetConstraints, SetArgumentsFunc<double> SetArguments);
+template void Tuner<float2>(int argc, char* argv[], const int V, GetTunerDefaultsFunc GetTunerDefaults, GetTunerSettingsFunc<float2> GetTunerSettings, TestValidArgumentsFunc<float2> TestValidArguments, SetConstraintsFunc SetConstraints, SetArgumentsFunc<float2> SetArguments);
+template void Tuner<double2>(int argc, char* argv[], const int V, GetTunerDefaultsFunc GetTunerDefaults, GetTunerSettingsFunc<double2> GetTunerSettings, TestValidArgumentsFunc<double2> TestValidArguments, SetConstraintsFunc SetConstraints, SetArgumentsFunc<double2> SetArguments);
+
 // =================================================================================================
 } // namespace clblast
diff --git a/src/tuning/tuning.hpp b/src/tuning/tuning.hpp
index ac6968dc..22210c7d 100644
--- a/src/tuning/tuning.hpp
+++ b/src/tuning/tuning.hpp
@@ -22,6 +22,7 @@
 #include <algorithm>
 #include <iostream>
 #include <chrono>
+#include <functional>
 
 #include "utilities/utilities.hpp"
 #include "utilities/compile.hpp"
@@ -116,282 +117,25 @@ void print_separator(const size_t parameters_size);
 
 // =================================================================================================
 
+using GetTunerDefaultsFunc = std::function<TunerDefaults(const int V)>;
+template <typename T>
+using GetTunerSettingsFunc = std::function<TunerSettings(const int V, const Arguments<T> &args)>;
+template <typename T>
+using TestValidArgumentsFunc = std::function<void(const int V, const Arguments<T> &args)>;
+using SetConstraintsFunc = std::function<std::vector<Constraint>(const int V)>;
+template <typename T>
+using SetArgumentsFunc = std::function<void(const int V, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers)>;
+
 // Function to get command-line argument, set-up the input buffers, configure the tuner, and collect
 // the results. Used for all types of kernel families. Note that this is a header-only function so
 // that it is automatically compiled for the various kernels (given as the 'C' template argument).
-template <typename C, typename T>
-void Tuner(int argc, char* argv[]) {
-  constexpr auto kSeed = 42; // fixed seed for reproducibility
-
-  // Sets the parameters and platform/device for which to tune (command-line options)
-  const TunerDefaults defaults = C::GetTunerDefaults();
-  auto command_line_args = RetrieveCommandLineArguments(argc, argv);
-  auto help = std::string{"* Options given/available:\n"};
-  auto args = Arguments<T>{};
-  args.platform_id = GetArgument(command_line_args, help, kArgPlatform, ConvertArgument(std::getenv("CLBLAST_PLATFORM"), size_t{0}));
-  args.device_id   = GetArgument(command_line_args, help, kArgDevice, ConvertArgument(std::getenv("CLBLAST_DEVICE"), size_t{0}));
-  args.precision   = GetArgument(command_line_args, help, kArgPrecision, Precision::kSingle);
-  for (auto &o: defaults.options) {
-    if (o == kArgM)        { args.m        = GetArgument(command_line_args, help, kArgM, defaults.default_m); }
-    if (o == kArgN)        { args.n        = GetArgument(command_line_args, help, kArgN, defaults.default_n); }
-    if (o == kArgK)        { args.k        = GetArgument(command_line_args, help, kArgK, defaults.default_k); }
-    if (o == kArgAlpha)    { args.alpha    = GetArgument(command_line_args, help, kArgAlpha, GetScalar<T>()); }
-    if (o == kArgBeta)     { args.beta     = GetArgument(command_line_args, help, kArgBeta, GetScalar<T>()); }
-    if (o == kArgBatchCount) { args.batch_count = GetArgument(command_line_args, help, kArgBatchCount, defaults.default_batch_count); }
-  }
-  args.fraction = GetArgument(command_line_args, help, kArgFraction, defaults.default_fraction);
-  args.num_runs = GetArgument(command_line_args, help, kArgNumRuns, defaults.default_num_runs);
-  const auto max_l2_norm = GetArgument(command_line_args, help, kArgMaxL2Norm, 1.0e-4);
-  printf("%s\n", help.c_str());
-  const TunerSettings settings = C::GetTunerSettings(args);
-
-  // Tests validity of the given arguments
-  C::TestValidArguments(args);
-
-  // Initializes OpenCL
-  const auto platform = Platform(args.platform_id);
-  const auto device = Device(platform, args.device_id);
-  const auto context = Context(device);
-
-  // Tests for validity of the precision and retrieves properties
-  if (!PrecisionSupported<T>(device)) {
-    printf("* Unsupported precision, skipping this tuning run\n\n");
-    return;
-  }
-  const auto device_type = GetDeviceType(device);
-  const auto device_vendor = GetDeviceVendor(device);
-  const auto device_architecture = GetDeviceArchitecture(device);
-  const auto device_name = GetDeviceName(device);
-
-  // Creates input buffers with random data
-  const auto buffer_sizes = std::vector<size_t>{
-      settings.size_x, settings.size_y,
-      settings.size_a, settings.size_b, settings.size_c,
-      settings.size_temp
-  };
-  std::mt19937 mt(kSeed);
-  std::uniform_real_distribution<double> dist(kTestDataLowerLimit, kTestDataUpperLimit);
-  auto source_buffers = std::vector<std::vector<T>>();
-  auto reference_buffers = std::vector<std::vector<T>>();
-  auto result_buffers = std::vector<std::vector<T>>();
-  auto device_buffers = std::vector<Buffer<T>>();
-  for (const auto size : buffer_sizes) {
-    auto host_buffer = std::vector<T>(size);
-    PopulateVector(host_buffer, mt, dist);
-    source_buffers.push_back(host_buffer);
-    reference_buffers.push_back(std::vector<T>(size));
-    result_buffers.push_back(std::vector<T>(size));
-    device_buffers.push_back(Buffer<T>(context, size));
-  }
-
-  // Sets the tunable parameters and their possible values
-  auto configurations = SetConfigurations(settings.parameters, C::SetConstraints());
-  printf("* Found %s%zu configuration(s)%s\n",
-         kPrintMessage.c_str(), configurations.size(), kPrintEnd.c_str());
-
-  // Select the search method (full search or a random fraction)
-  if (args.fraction != 0.0 && args.fraction != 1.0) {
-    const auto new_size = static_cast<size_t>(configurations.size() / args.fraction);
-    auto rng = std::default_random_engine{};
-    std::shuffle(std::begin(configurations), std::end(configurations), rng);
-    configurations.resize(new_size);
-    printf("* Exploring a random subset of %s%zu configuration(s)%s\n",
-           kPrintMessage.c_str(), configurations.size(), kPrintEnd.c_str());
-  }
-
-  // Prints information about the parameters
-  printf("* Parameters explored: ");
-  for (const auto& parameter : settings.parameters) { printf("%s ", parameter.first.c_str()); }
-  printf("\n");
-
-  // Prints the header of the table
-  printf("\n");
-  printf("|   ID | total |");
-  for (auto i = size_t{0}; i < settings.parameters.size() - 1; ++i) { printf("     "); }
-  printf("param |       compiles |         time | %6s |            status |\n", settings.performance_unit.c_str());
-  print_separator(settings.parameters.size());
-
-  // First runs a reference example to compare against
-  try {
-    auto queue = Queue(context, device);
-    printf("|  ref |     - |");
-    for (auto i = size_t{0}; i < settings.parameters.size() - 1; ++i) { printf("     "); }
-    printf("    - |");
-
-
-    // Sets the input
-    for (const auto id : settings.inputs) {
-      device_buffers[id].Write(queue, buffer_sizes[id], source_buffers[id]);
-    }
-
-    // Compiles the kernel
-    auto compiler_options = std::vector<std::string>();
-    const auto program = CompileFromSource(settings.sources, args.precision, settings.kernel_name,
-                                           device, context, compiler_options, 0);
-    auto kernel = Kernel(program, settings.kernel_name);
-    C::SetArguments(kernel, args, device_buffers);
-    printf("             %sOK%s |", kPrintSuccess.c_str(), kPrintEnd.c_str());
-
-    // Runs the kernel
-    const auto time_ms = TimeKernel(args.num_runs, kernel, queue, device,
-                                    settings.global_size_ref, settings.local_size_ref);
-    printf("      - |");
-    if (time_ms == -1.0) { throw std::runtime_error("Error in reference implementation"); }
-
-    // Saves the result
-    for (const auto id : settings.outputs) {
-      device_buffers[id].Read(queue, buffer_sizes[id], reference_buffers[id]);
-    }
-    printf("      %sreference OK%s |\n", kPrintSuccess.c_str(), kPrintEnd.c_str());
-  }
-  catch (...) {
-    const auto status_code = DispatchExceptionCatchAll(true);
-    printf("* Exception caught with status %d while running the reference, aborting\n",
-           static_cast<int>(status_code));
-    return;
-  }
-  print_separator(settings.parameters.size());
-
-  // Starts the tuning process
-  auto results = std::vector<TuningResult>();
-  for (auto config_id = size_t{0}; config_id < configurations.size(); ++config_id) {
-    try {
-      auto queue = Queue(context, device);
-
-      auto configuration = configurations[config_id];
-      printf("| %4zu | %5zu |", config_id + 1, configurations.size());
-      for (const auto& parameter : settings.parameters) {
-        printf("%5zu", configuration.at(parameter.first));
-      }
-      printf(" |");
-
-      // Sets the input
-      for (const auto id : settings.inputs) {
-        device_buffers[id].Write(queue, buffer_sizes[id], source_buffers[id]);
-      }
-
-      // Sets the thread configuration
-      const auto global = SetThreadConfiguration(configuration, settings.global_size,
-                                                 settings.mul_global, settings.div_global);
-      const auto local = SetThreadConfiguration(configuration, settings.local_size,
-                                                settings.mul_local, settings.div_local);
-
-      // Sets the parameters for this configuration
-      auto kernel_source = std::string{""};
-      for (const auto &parameter : configuration) {
-        kernel_source += "#define " + parameter.first + " " + ToString(parameter.second) + "\n";
-      }
-      kernel_source += settings.sources;
-
-      // Compiles the kernel
-      const auto start_time = std::chrono::steady_clock::now();
-      auto compiler_options = std::vector<std::string>();
-      const auto program = CompileFromSource(kernel_source, args.precision, settings.kernel_name,
-                                             device, context, compiler_options, 0, true);
-      auto kernel = Kernel(program, settings.kernel_name);
-      const auto elapsed_time = std::chrono::steady_clock::now() - start_time;
-      const auto timing = std::chrono::duration<double,std::milli>(elapsed_time).count();
-      printf("   %sOK%s  %5.0lf ms |", kPrintSuccess.c_str(), kPrintEnd.c_str(), timing);
-
-      // Runs the kernel
-      C::SetArguments(kernel, args, device_buffers);
-      const auto time_ms = TimeKernel(args.num_runs, kernel, queue, device, global, local);
-
-      // Kernel run was not successful
-      if (time_ms == -1.0) {
-        printf("      - |");
-        printf("   %sinvalid config.%s |", kPrintError.c_str(), kPrintEnd.c_str());
-        printf(" <-- skipping\n");
-        continue;
-      }
-
-      // Compares the results
-      auto l2_error = 0.0;
-      for (const auto id : settings.outputs) {
-        device_buffers[id].Read(queue, buffer_sizes[id], result_buffers[id]);
-        for (auto index = size_t{0}; index<buffer_sizes[id]; ++index) {
-          const auto diff = SquaredDifference(result_buffers[id][index], reference_buffers[id][index]);
-          l2_error += diff;
-        }
-        l2_error /= static_cast<double>(buffer_sizes[id]);
-        if (std::isnan(l2_error) || l2_error > max_l2_norm) {
-          printf("      - |");
-          printf(" %sL2 error %8.2e%s |", kPrintError.c_str(), l2_error, kPrintEnd.c_str());
-          throw std::runtime_error("L2 error too large");
-        }
-      }
-
-      // All was OK
-      configuration["PRECISION"] = static_cast<size_t>(args.precision);
-      results.push_back(TuningResult{settings.kernel_name, time_ms, configuration});
-      printf(" %6.1lf |", settings.metric_amount / (time_ms * 1.0e6));
-      printf("     %sresults match%s |\n", kPrintSuccess.c_str(), kPrintEnd.c_str());
-    }
-    catch (CLCudaAPIBuildError) {
-      const auto status_code = DispatchExceptionCatchAll(true);
-      printf("  %scompilation error: %5d%s     |",
-             kPrintError.c_str(), static_cast<int>(status_code), kPrintEnd.c_str());
-      printf("      - |                 - | <-- skipping\n");
-    }
-    catch (...) {
-      const auto status_code = DispatchExceptionCatchAll(true);
-      if (status_code != StatusCode::kUnknownError) {
-        printf("   %serror code %d%s |",
-               kPrintError.c_str(), static_cast<int>(status_code), kPrintEnd.c_str());
-      }
-      printf(" <-- skipping\n");
-    }
-  }
-
-  // Completed the tuning process
-  print_separator(settings.parameters.size());
-  printf("\n");
-  if (results.size() == 0) { return; }
-
-  // Computes the best results
-  auto comparison = [](const TuningResult& lhs, const TuningResult& rhs) { return lhs.score < rhs.score; };
-  const auto best_configuration = std::min_element(results.begin(), results.end(), comparison);
-  const auto best_time_ms = best_configuration->score;
-  if (best_time_ms == 0.0) { return; }
-
-  // Also prints the performance of the best-case in terms of GB/s or GFLOPS
-  printf("\n");
-  printf("* Found best result %.2lf ms", best_time_ms);
-  printf(": %.1lf %s\n", settings.metric_amount / (best_time_ms * 1.0e6),
-         settings.performance_unit.c_str());
-  printf("* Best parameters: ");
-  auto best_string = std::string{""};
-  auto i = size_t{0};
-  for (const auto config : best_configuration->config) {
-    best_string += "" + config.first + "=" + ToString(config.second);
-    if (i < best_configuration->config.size() - 1) { best_string += " "; }
-    ++i;
-  }
-  printf("%s\n\n", best_string.c_str());
-
-  // Outputs the results as JSON to disk, including some meta-data
-  auto precision_string = std::to_string(static_cast<size_t>(args.precision));
-  auto metadata = std::vector<std::pair<std::string,std::string>>{
-    {"kernel_family", settings.kernel_family},
-    {"precision", precision_string},
-    {"best_kernel", best_configuration->name},
-    {"best_time", ToString(best_configuration->score)},
-    {"best_parameters", best_string}
-  };
-  for (auto &o: defaults.options) {
-    if (o == kArgM)     { metadata.push_back({"arg_m", ToString(args.m)}); }
-    if (o == kArgN)     { metadata.push_back({"arg_n", ToString(args.n)}); }
-    if (o == kArgK)     { metadata.push_back({"arg_k", ToString(args.k)}); }
-    if (o == kArgAlpha) { metadata.push_back({"arg_alpha", ToString(args.alpha)}); }
-    if (o == kArgBeta)  { metadata.push_back({"arg_beta", ToString(args.beta)}); }
-    if (o == kArgBatchCount) { metadata.push_back({"arg_batch_count", ToString(args.batch_count)}); }
-  }
-  PrintTimingsToFileAsJSON("clblast_" + settings.kernel_family + "_" + precision_string + ".json",
-                           device, platform, metadata, results);
-
-  printf("* Completed tuning process\n");
-  printf("\n");
-}
+template <typename T>
+void Tuner(int argc, char* argv[], const int V,
+           GetTunerDefaultsFunc GetTunerDefaults,
+           GetTunerSettingsFunc<T> GetTunerSettings,
+           TestValidArgumentsFunc<T> TestValidArguments,
+           SetConstraintsFunc SetConstraints,
+           SetArgumentsFunc<T> SetArguments);
 
 // =================================================================================================
 } // namespace clblast
-- 
cgit v1.2.3