Separate kernel tuners in .cpp with main and .hpp with settings

22 files changed, 1291 insertions, 1116 deletions

diff --git a/src/tuning/kernels/copy_fast.cpp b/src/tuning/kernels/copy_fast.cpp
index d046c9e9..a71dd083 100644
--- a/src/tuning/kernels/copy_fast.cpp
+++ b/src/tuning/kernels/copy_fast.cpp
@@ -11,86 +11,7 @@
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// 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)
-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 =
-#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;
-}
-
-// 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
+#include "tuning/kernels/copy_fast.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/copy_fast.hpp b/src/tuning/kernels/copy_fast.hpp
new file mode 100644
index 00000000..eab1c7dd
--- /dev/null
+++ b/src/tuning/kernels/copy_fast.hpp
@@ -0,0 +1,93 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the copy OpenCL kernels.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// 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)
+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 =
+#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;
+}
+
+// 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
diff --git a/src/tuning/kernels/copy_pad.cpp b/src/tuning/kernels/copy_pad.cpp
index 1b483e86..e32a5746 100644
--- a/src/tuning/kernels/copy_pad.cpp
+++ b/src/tuning/kernels/copy_pad.cpp
@@ -11,94 +11,7 @@
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// 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)
-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 =
-#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;
-}
-
-// 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
+#include "tuning/kernels/copy_pad.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/copy_pad.hpp b/src/tuning/kernels/copy_pad.hpp
new file mode 100644
index 00000000..b39c0318
--- /dev/null
+++ b/src/tuning/kernels/copy_pad.hpp
@@ -0,0 +1,101 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the pad OpenCL kernels.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// 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)
+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 =
+#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;
+}
+
+// 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
diff --git a/src/tuning/kernels/invert.cpp b/src/tuning/kernels/invert.cpp
index d846fdf7..ecd51199 100644
--- a/src/tuning/kernels/invert.cpp
+++ b/src/tuning/kernels/invert.cpp
@@ -11,99 +11,7 @@
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// Settings for this kernel (default command-line arguments)
-TunerDefaults GetTunerDefaults(const int) {
-  auto settings = TunerDefaults();
-  settings.options = {kArgN, kArgM, kArgK};
-  settings.default_n = 128; // dimension of input matrix 'n'
-  settings.default_m = 64; // block size
-  settings.default_k = 16; // current size
-  return settings;
-}
-
-// 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 = "invert";
-  settings.kernel_name = "TripleMatMul16Part1Lower";
-  settings.sources =
-"#define ROUTINE_INVERT"
-#include "../src/kernels/level3/invert_diagonal_blocks_part1.opencl"
-#include "../src/kernels/level3/invert_diagonal_blocks_part2.opencl"
-  ;
-
-  // Buffer sizes
-  settings.size_a = args.n * args.n + args.a_offset;
-  settings.size_b = Ceil(args.n, args.m) * args.m; // Ceil(n, block_size) * block_size
-
-  // 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
-  const auto num_pages = CeilDiv(args.n, args.k * 2); // CeilDiv(n, current_size*2)
-  settings.global_size = {args.k / 4, num_pages * (args.k / 16) * 4};
-  settings.global_size_ref = settings.global_size;
-  settings.local_size = {1, 1};
-  settings.local_size_ref = {4, 4};
-
-  // Transforms the thread configuration based on the parameters
-  settings.mul_local = {{"TMMWGSX", "TMMWGSY"}};
-  settings.div_global = {{}};
-
-  // Sets the tuning parameters and their possible values
-  // TODO: Make these actually tunable, apart from LOCALPAD
-  settings.parameters = {
-    {"INTERNAL_BLOCK_SIZE", {16}},
-    {"LOCALPAD", {0, 1}},
-    {"TMMWGSX", {4}},
-    {"TMMWGSY", {4}},
-  };
-
-  // Describes how to compute the performance metrics
-  settings.metric_amount = 1 * GetBytes(args.precision);
-  settings.performance_unit = "N/A";
-
-  return settings;
-}
-
-// Tests for valid arguments
-template <typename T>
-void TestValidArguments(const int, const Arguments<T> &args) {
-  if (!(args.k == 16)) {
-    throw std::runtime_error("'TripleMatMul16Part1Lower' requires 'k' to be 16");
-  }
-}
-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) {
-  const auto num_pages = CeilDiv(args.n, args.k * 2); // CeilDiv(n, current_size*2)
-  kernel.SetArgument(0, static_cast<int>(args.n)); // n
-  kernel.SetArgument(1, buffers[2]()); // 2 == A matrix
-  kernel.SetArgument(2, 0); // a_offset
-  kernel.SetArgument(3, static_cast<int>(args.n)); // a_ld
-  kernel.SetArgument(4, buffers[3]()); // 3 == B matrix
-  kernel.SetArgument(5, static_cast<int>(args.k)); // current_size
-  kernel.SetArgument(6, static_cast<int>(num_pages)); // num_pages
-  kernel.SetArgument(7, static_cast<int>(args.m)); // block_size
-}
-
-// =================================================================================================
-} // namespace clblast
+#include "tuning/kernels/invert.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/invert.hpp b/src/tuning/kernels/invert.hpp
new file mode 100644
index 00000000..0178a2aa
--- /dev/null
+++ b/src/tuning/kernels/invert.hpp
@@ -0,0 +1,106 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the invert OpenCL kernels.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// Settings for this kernel (default command-line arguments)
+TunerDefaults GetTunerDefaults(const int) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgN, kArgM, kArgK};
+  settings.default_n = 128; // dimension of input matrix 'n'
+  settings.default_m = 64; // block size
+  settings.default_k = 16; // current size
+  return settings;
+}
+
+// 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 = "invert";
+  settings.kernel_name = "TripleMatMul16Part1Lower";
+  settings.sources =
+"#define ROUTINE_INVERT"
+#include "../src/kernels/level3/invert_diagonal_blocks_part1.opencl"
+#include "../src/kernels/level3/invert_diagonal_blocks_part2.opencl"
+  ;
+
+  // Buffer sizes
+  settings.size_a = args.n * args.n + args.a_offset;
+  settings.size_b = Ceil(args.n, args.m) * args.m; // Ceil(n, block_size) * block_size
+
+  // 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
+  const auto num_pages = CeilDiv(args.n, args.k * 2); // CeilDiv(n, current_size*2)
+  settings.global_size = {args.k / 4, num_pages * (args.k / 16) * 4};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1, 1};
+  settings.local_size_ref = {4, 4};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"TMMWGSX", "TMMWGSY"}};
+  settings.div_global = {{}};
+
+  // Sets the tuning parameters and their possible values
+  // TODO: Make these actually tunable, apart from LOCALPAD
+  settings.parameters = {
+    {"INTERNAL_BLOCK_SIZE", {16}},
+    {"LOCALPAD", {0, 1}},
+    {"TMMWGSX", {4}},
+    {"TMMWGSY", {4}},
+  };
+
+  // Describes how to compute the performance metrics
+  settings.metric_amount = 1 * GetBytes(args.precision);
+  settings.performance_unit = "N/A";
+
+  return settings;
+}
+
+// Tests for valid arguments
+template <typename T>
+void TestValidArguments(const int, const Arguments<T> &args) {
+  if (!(args.k == 16)) {
+    throw std::runtime_error("'TripleMatMul16Part1Lower' requires 'k' to be 16");
+  }
+}
+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) {
+  const auto num_pages = CeilDiv(args.n, args.k * 2); // CeilDiv(n, current_size*2)
+  kernel.SetArgument(0, static_cast<int>(args.n)); // n
+  kernel.SetArgument(1, buffers[2]()); // 2 == A matrix
+  kernel.SetArgument(2, 0); // a_offset
+  kernel.SetArgument(3, static_cast<int>(args.n)); // a_ld
+  kernel.SetArgument(4, buffers[3]()); // 3 == B matrix
+  kernel.SetArgument(5, static_cast<int>(args.k)); // current_size
+  kernel.SetArgument(6, static_cast<int>(num_pages)); // num_pages
+  kernel.SetArgument(7, static_cast<int>(args.m)); // block_size
+}
+
+// =================================================================================================
+} // namespace clblast
diff --git a/src/tuning/kernels/transpose_fast.cpp b/src/tuning/kernels/transpose_fast.cpp
index 5b701a5b..2144f23f 100644
--- a/src/tuning/kernels/transpose_fast.cpp
+++ b/src/tuning/kernels/transpose_fast.cpp
@@ -11,86 +11,7 @@
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// 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)
-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 =
-#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;
-}
-
-// 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
+#include "tuning/kernels/transpose_fast.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/transpose_fast.hpp b/src/tuning/kernels/transpose_fast.hpp
new file mode 100644
index 00000000..47859f7a
--- /dev/null
+++ b/src/tuning/kernels/transpose_fast.hpp
@@ -0,0 +1,93 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the transpose OpenCL kernels.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// 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)
+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 =
+#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;
+}
+
+// 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
diff --git a/src/tuning/kernels/transpose_pad.cpp b/src/tuning/kernels/transpose_pad.cpp
index ed24fb04..ce39d857 100644
--- a/src/tuning/kernels/transpose_pad.cpp
+++ b/src/tuning/kernels/transpose_pad.cpp
@@ -11,93 +11,7 @@
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// 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)
-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 =
-#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;
-}
-
-// 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
+#include "tuning/kernels/transpose_pad.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/transpose_pad.hpp b/src/tuning/kernels/transpose_pad.hpp
new file mode 100644
index 00000000..c9ce983e
--- /dev/null
+++ b/src/tuning/kernels/transpose_pad.hpp
@@ -0,0 +1,100 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the pad-transpose OpenCL kernels.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// 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)
+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 =
+#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;
+}
+
+// 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
diff --git a/src/tuning/kernels/xaxpy.cpp b/src/tuning/kernels/xaxpy.cpp
index dd44018c..a290b6c1 100644
--- a/src/tuning/kernels/xaxpy.cpp
+++ b/src/tuning/kernels/xaxpy.cpp
@@ -11,88 +11,7 @@
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// 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)
-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 =
-#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;
-}
-
-// 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
+#include "tuning/kernels/xaxpy.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/xaxpy.hpp b/src/tuning/kernels/xaxpy.hpp
new file mode 100644
index 00000000..266a671b
--- /dev/null
+++ b/src/tuning/kernels/xaxpy.hpp
@@ -0,0 +1,95 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the xaxpy OpenCL kernels.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// 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)
+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 =
+#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;
+}
+
+// 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
diff --git a/src/tuning/kernels/xdot.cpp b/src/tuning/kernels/xdot.cpp
index 635d012a..10126392 100644
--- a/src/tuning/kernels/xdot.cpp
+++ b/src/tuning/kernels/xdot.cpp
@@ -12,94 +12,7 @@
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// 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)
-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 =
-#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
-
-  // 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};
-
-  // 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}},
-  };
-
-  // 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;
-}
-
-// 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
+#include "tuning/kernels/xdot.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/xdot.hpp b/src/tuning/kernels/xdot.hpp
new file mode 100644
index 00000000..456baea4
--- /dev/null
+++ b/src/tuning/kernels/xdot.hpp
@@ -0,0 +1,102 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the xdot OpenCL kernels. Note that the results are
+// not verified, since the result is not final and depends on the WGS2 parameter.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// 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)
+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 =
+#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
+
+  // 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};
+
+  // 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}},
+  };
+
+  // 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;
+}
+
+// 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
diff --git a/src/tuning/kernels/xgemm.cpp b/src/tuning/kernels/xgemm.cpp
index 5c242757..f388c76c 100644
--- a/src/tuning/kernels/xgemm.cpp
+++ b/src/tuning/kernels/xgemm.cpp
@@ -7,162 +7,11 @@
 // Author(s):
 //   Cedric Nugteren <www.cedricnugteren.nl>
 //
-// This file uses the auto-tuner to tune the xgemm OpenCL kernels. There are two variations:
-// - V==1: This tests some limited set of tuning parameters exhaustively.
-// - V==2: This tests a much larger set of tuning parameters by randomly sampling a subset.
+// This file uses the auto-tuner to tune the xgemm OpenCL kernels.
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// 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)
-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 =
-#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;
-}
-
-// 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 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
+#include "tuning/kernels/xgemm.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/xgemm.hpp b/src/tuning/kernels/xgemm.hpp
new file mode 100644
index 00000000..09ebe91d
--- /dev/null
+++ b/src/tuning/kernels/xgemm.hpp
@@ -0,0 +1,165 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the xgemm OpenCL kernels. There are two variations:
+// - V==1: This tests some limited set of tuning parameters exhaustively.
+// - V==2: This tests a much larger set of tuning parameters by randomly sampling a subset.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// 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)
+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 =
+#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;
+}
+
+// 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 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
diff --git a/src/tuning/kernels/xgemm_direct.cpp b/src/tuning/kernels/xgemm_direct.cpp
index 0bd2e94d..b059ad62 100644
--- a/src/tuning/kernels/xgemm_direct.cpp
+++ b/src/tuning/kernels/xgemm_direct.cpp
@@ -7,159 +7,11 @@
 // Author(s):
 //   Cedric Nugteren <www.cedricnugteren.nl>
 //
-// This file uses the auto-tuner to tune the direct xgemm kernels. There are two variations:
-// - V==1: This tests some limited set of tuning parameters exhaustively.
-// - V==2: This tests a much larger set of tuning parameters by randomly sampling a subset.
+// This file uses the auto-tuner to tune the direct xgemm kernels.
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// 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)
-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 =
-#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;
-}
-
-// 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
+#include "tuning/kernels/xgemm_direct.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/xgemm_direct.hpp b/src/tuning/kernels/xgemm_direct.hpp
new file mode 100644
index 00000000..7cc9f654
--- /dev/null
+++ b/src/tuning/kernels/xgemm_direct.hpp
@@ -0,0 +1,162 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the direct xgemm kernels. There are two variations:
+// - V==1: This tests some limited set of tuning parameters exhaustively.
+// - V==2: This tests a much larger set of tuning parameters by randomly sampling a subset.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// 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)
+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 =
+#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;
+}
+
+// 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
diff --git a/src/tuning/kernels/xgemv.cpp b/src/tuning/kernels/xgemv.cpp
index 965fc53d..165c5628 100644
--- a/src/tuning/kernels/xgemv.cpp
+++ b/src/tuning/kernels/xgemv.cpp
@@ -7,135 +7,11 @@
 // Author(s):
 //   Cedric Nugteren <www.cedricnugteren.nl>
 //
-// This file uses the auto-tuner to tune the xgemv OpenCL kernels. Three variants are tuned:
-// 1: The full version of the kernel
-// 2: The fast version for non-transposed matrices
-// 3: The fast version for transposed matrices
+// This file uses the auto-tuner to tune the xgemv OpenCL kernels.
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// 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;
-  settings.default_num_runs = 4;
-  return settings;
-}
-
-// 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 =
-#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;
-}
-
-// 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
+#include "tuning/kernels/xgemv.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/xgemv.hpp b/src/tuning/kernels/xgemv.hpp
new file mode 100644
index 00000000..87186195
--- /dev/null
+++ b/src/tuning/kernels/xgemv.hpp
@@ -0,0 +1,138 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the xgemv OpenCL kernels. Three variants are tuned:
+// 1: The full version of the kernel
+// 2: The fast version for non-transposed matrices
+// 3: The fast version for transposed matrices
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// 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;
+  settings.default_num_runs = 4;
+  return settings;
+}
+
+// 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 =
+#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;
+}
+
+// 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
diff --git a/src/tuning/kernels/xger.cpp b/src/tuning/kernels/xger.cpp
index a88fb5d6..8a90f340 100644
--- a/src/tuning/kernels/xger.cpp
+++ b/src/tuning/kernels/xger.cpp
@@ -11,95 +11,7 @@
 //
 // =================================================================================================
 
-#include <string>
-#include <vector>
-
-#include "utilities/utilities.hpp"
-#include "tuning/tuning.hpp"
-
-namespace clblast {
-// =================================================================================================
-
-// 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)
-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 =
-#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;
-}
-
-// 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
+#include "tuning/kernels/xger.hpp"
 
 // Shortcuts to the clblast namespace
 using half = clblast::half;
diff --git a/src/tuning/kernels/xger.hpp b/src/tuning/kernels/xger.hpp
new file mode 100644
index 00000000..a88fb5d6
--- /dev/null
+++ b/src/tuning/kernels/xger.hpp
@@ -0,0 +1,122 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the xger OpenCL kernels.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// 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)
+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 =
+#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;
+}
+
+// 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
+
+// Shortcuts to the clblast namespace
+using half = clblast::half;
+using float2 = clblast::float2;
+using double2 = clblast::double2;
+
+// Main function (not within the clblast namespace)
+int main(int argc, char *argv[]) {
+  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+  switch(clblast::GetPrecision(command_line_args)) {
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::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;
+}
+
+// =================================================================================================