From b98af44fcf89b9946e1de438b1f5527e6bf28905 Mon Sep 17 00:00:00 2001
From: Ivan Shapovalov <intelfx@intelfx.name>
Date: Sat, 22 Oct 2016 05:14:19 +0300
Subject: treewide: use C++ exceptions properly

Since the codebase is designed around proper C++ idioms such as RAII, it
makes sense to only use C++ exceptions internally instead of mixing
exceptions and error codes. The exceptions are now caught at top level
to preserve compatibility with the existing error code-based API.

Note that we deliberately do not catch C++ runtime errors (such as
`std::bad_alloc`) nor logic errors (aka failed assertions) because no
actual handling can ever happen for such errors.

However, in the C interface we do catch _all_ exceptions (...) and
convert them into a wild-card error code.
---
 src/routines/level2/xger.cpp | 77 +++++++++++++++++++-------------------------
 1 file changed, 34 insertions(+), 43 deletions(-)

(limited to 'src/routines/level2/xger.cpp')

diff --git a/src/routines/level2/xger.cpp b/src/routines/level2/xger.cpp
index 29cffe0c..f22131bb 100644
--- a/src/routines/level2/xger.cpp
+++ b/src/routines/level2/xger.cpp
@@ -33,15 +33,15 @@ Xger<T>::Xger(Queue &queue, EventPointer event, const std::string &name):
 
 // The main routine
 template <typename T>
-StatusCode Xger<T>::DoGer(const Layout layout,
-                          const size_t m, const size_t n,
-                          const T alpha,
-                          const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                          const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
-                          const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld) {
+void Xger<T>::DoGer(const Layout layout,
+                    const size_t m, const size_t n,
+                    const T alpha,
+                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
+                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
+                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld) {
 
   // Makes sure all dimensions are larger than zero
-  if (m == 0 || n == 0) { return StatusCode::kInvalidDimension; }
+  if (m == 0 || n == 0) { throw BLASError(StatusCode::kInvalidDimension); }
 
   // Computes whether or not the matrix has an alternative layout (row or column-major).
   const auto a_is_rowmajor = (layout == Layout::kRowMajor);
@@ -49,44 +49,35 @@ StatusCode Xger<T>::DoGer(const Layout layout,
   const auto a_two = (a_is_rowmajor) ? m : n;
 
   // Tests the matrix and the vectors for validity
-  auto status = TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld);
-  if (ErrorIn(status)) { return status; }
-  status = TestVectorX(m, x_buffer, x_offset, x_inc);
-  if (ErrorIn(status)) { return status; }
-  status = TestVectorY(n, y_buffer, y_offset, y_inc);
-  if (ErrorIn(status)) { return status; }
+  TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld);
+  TestVectorX(m, x_buffer, x_offset, x_inc);
+  TestVectorY(n, y_buffer, y_offset, y_inc);
 
   // Retrieves the kernel from the compiled binary
-  try {
-    const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-    auto kernel = Kernel(program, "Xger");
-
-    // Sets the kernel arguments
-    kernel.SetArgument(0, static_cast<int>(a_one));
-    kernel.SetArgument(1, static_cast<int>(a_two));
-    kernel.SetArgument(2, GetRealArg(alpha));
-    kernel.SetArgument(3, x_buffer());
-    kernel.SetArgument(4, static_cast<int>(x_offset));
-    kernel.SetArgument(5, static_cast<int>(x_inc));
-    kernel.SetArgument(6, y_buffer());
-    kernel.SetArgument(7, static_cast<int>(y_offset));
-    kernel.SetArgument(8, static_cast<int>(y_inc));
-    kernel.SetArgument(9, a_buffer());
-    kernel.SetArgument(10, static_cast<int>(a_offset));
-    kernel.SetArgument(11, static_cast<int>(a_ld));
-    kernel.SetArgument(12, static_cast<int>(a_is_rowmajor));
-
-    // Launches the kernel
-    auto a_one_ceiled = Ceil(CeilDiv(a_one, db_["WPT"]), db_["WGS1"]);
-    auto a_two_ceiled = Ceil(CeilDiv(a_two, db_["WPT"]), db_["WGS2"]);
-    auto global = std::vector<size_t>{a_one_ceiled, a_two_ceiled};
-    auto local = std::vector<size_t>{db_["WGS1"], db_["WGS2"]};
-    status = RunKernel(kernel, queue_, device_, global, local, event_);
-    if (ErrorIn(status)) { return status; }
-
-    // Succesfully finished the computation
-    return StatusCode::kSuccess;
-  } catch (...) { return StatusCode::kInvalidKernel; }
+  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
+  auto kernel = Kernel(program, "Xger");
+
+  // Sets the kernel arguments
+  kernel.SetArgument(0, static_cast<int>(a_one));
+  kernel.SetArgument(1, static_cast<int>(a_two));
+  kernel.SetArgument(2, GetRealArg(alpha));
+  kernel.SetArgument(3, x_buffer());
+  kernel.SetArgument(4, static_cast<int>(x_offset));
+  kernel.SetArgument(5, static_cast<int>(x_inc));
+  kernel.SetArgument(6, y_buffer());
+  kernel.SetArgument(7, static_cast<int>(y_offset));
+  kernel.SetArgument(8, static_cast<int>(y_inc));
+  kernel.SetArgument(9, a_buffer());
+  kernel.SetArgument(10, static_cast<int>(a_offset));
+  kernel.SetArgument(11, static_cast<int>(a_ld));
+  kernel.SetArgument(12, static_cast<int>(a_is_rowmajor));
+
+  // Launches the kernel
+  auto a_one_ceiled = Ceil(CeilDiv(a_one, db_["WPT"]), db_["WGS1"]);
+  auto a_two_ceiled = Ceil(CeilDiv(a_two, db_["WPT"]), db_["WGS2"]);
+  auto global = std::vector<size_t>{a_one_ceiled, a_two_ceiled};
+  auto local = std::vector<size_t>{db_["WGS1"], db_["WGS2"]};
+  RunKernel(kernel, queue_, device_, global, local, event_);
 }
 
 // =================================================================================================
-- 
cgit v1.2.3


From 56f300607b1d0b81ab3269894fda5a066c46cdeb Mon Sep 17 00:00:00 2001
From: Ivan Shapovalov <intelfx@intelfx.name>
Date: Tue, 18 Oct 2016 04:53:06 +0300
Subject: Routine: get rid of ::SetUp()

Since we now use C++ exceptions inside the implementation (and exceptions
can be thrown from constructors), there is no need for a separate
Routine::SetUp() function.

For this, we also change the way how the kernel source string is constructed.
The kernel-specific source code is now passed to the Routine ctor via
an initializer_list of C strings to avoid unnecessary data copying
while also working around C1091 of MSVC 2013.
---
 scripts/generator/generator/cpp.py |   1 -
 src/clblast.cpp                    | 134 +++++++++++++------------------------
 src/routine.cpp                    |  39 +++++------
 src/routine.hpp                    |  10 ++-
 src/routines/level1/xamax.cpp      |   5 +-
 src/routines/level1/xasum.cpp      |   5 +-
 src/routines/level1/xaxpy.cpp      |   5 +-
 src/routines/level1/xcopy.cpp      |   5 +-
 src/routines/level1/xdot.cpp       |   5 +-
 src/routines/level1/xnrm2.cpp      |   5 +-
 src/routines/level1/xscal.cpp      |   5 +-
 src/routines/level1/xswap.cpp      |   5 +-
 src/routines/level2/xgemv.cpp      |   5 +-
 src/routines/level2/xger.cpp       |   5 +-
 src/routines/level2/xher.cpp       |   5 +-
 src/routines/level2/xher2.cpp      |   5 +-
 src/routines/level3/xgemm.cpp      |   9 +--
 src/routines/level3/xher2k.cpp     |   5 +-
 src/routines/level3/xherk.cpp      |   5 +-
 src/routines/level3/xsyr2k.cpp     |   5 +-
 src/routines/level3/xsyrk.cpp      |   5 +-
 src/routines/levelx/xomatcopy.cpp  |   5 +-
 22 files changed, 104 insertions(+), 174 deletions(-)

(limited to 'src/routines/level2/xger.cpp')

diff --git a/scripts/generator/generator/cpp.py b/scripts/generator/generator/cpp.py
index 507b0d28..a0d43667 100644
--- a/scripts/generator/generator/cpp.py
+++ b/scripts/generator/generator/cpp.py
@@ -52,7 +52,6 @@ def clblast_cc(routine):
         result += "  try {" + NL
         result += "    auto queue_cpp = Queue(*queue);" + NL
         result += "    auto routine = X" + routine.name + "<" + routine.template.template + ">(queue_cpp, event);" + NL
-        result += "    routine.SetUp();" + NL
         result += "    routine.Do" + routine.name.capitalize() + "("
         result += ("," + NL + indent1).join([a for a in routine.arguments_clcudaapi()])
         result += ");" + NL
diff --git a/src/clblast.cpp b/src/clblast.cpp
index 8e7b042f..4bb4e0b3 100644
--- a/src/clblast.cpp
+++ b/src/clblast.cpp
@@ -171,7 +171,6 @@ StatusCode Swap(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xswap<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSwap(n,
                    Buffer<T>(x_buffer), x_offset, x_inc,
                    Buffer<T>(y_buffer), y_offset, y_inc);
@@ -208,7 +207,6 @@ StatusCode Scal(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xscal<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoScal(n,
                    alpha,
                    Buffer<T>(x_buffer), x_offset, x_inc);
@@ -245,7 +243,6 @@ StatusCode Copy(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xcopy<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoCopy(n,
                    Buffer<T>(x_buffer), x_offset, x_inc,
                    Buffer<T>(y_buffer), y_offset, y_inc);
@@ -283,7 +280,6 @@ StatusCode Axpy(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xaxpy<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoAxpy(n,
                    alpha,
                    Buffer<T>(x_buffer), x_offset, x_inc,
@@ -327,7 +323,6 @@ StatusCode Dot(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xdot<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoDot(n,
                   Buffer<T>(dot_buffer), dot_offset,
                   Buffer<T>(x_buffer), x_offset, x_inc,
@@ -361,7 +356,6 @@ StatusCode Dotu(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xdotu<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoDotu(n,
                    Buffer<T>(dot_buffer), dot_offset,
                    Buffer<T>(x_buffer), x_offset, x_inc,
@@ -390,7 +384,6 @@ StatusCode Dotc(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xdotc<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoDotc(n,
                    Buffer<T>(dot_buffer), dot_offset,
                    Buffer<T>(x_buffer), x_offset, x_inc,
@@ -418,7 +411,6 @@ StatusCode Nrm2(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xnrm2<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoNrm2(n,
                    Buffer<T>(nrm2_buffer), nrm2_offset,
                    Buffer<T>(x_buffer), x_offset, x_inc);
@@ -455,7 +447,6 @@ StatusCode Asum(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xasum<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoAsum(n,
                    Buffer<T>(asum_buffer), asum_offset,
                    Buffer<T>(x_buffer), x_offset, x_inc);
@@ -492,7 +483,6 @@ StatusCode Sum(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xsum<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSum(n,
                   Buffer<T>(sum_buffer), sum_offset,
                   Buffer<T>(x_buffer), x_offset, x_inc);
@@ -529,7 +519,6 @@ StatusCode Amax(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xamax<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoAmax(n,
                    Buffer<unsigned int>(imax_buffer), imax_offset,
                    Buffer<T>(x_buffer), x_offset, x_inc);
@@ -566,7 +555,6 @@ StatusCode Max(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xmax<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoMax(n,
                   Buffer<unsigned int>(imax_buffer), imax_offset,
                   Buffer<T>(x_buffer), x_offset, x_inc);
@@ -603,7 +591,6 @@ StatusCode Min(const size_t n,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xmin<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoMin(n,
                   Buffer<unsigned int>(imin_buffer), imin_offset,
                   Buffer<T>(x_buffer), x_offset, x_inc);
@@ -648,7 +635,6 @@ StatusCode Gemv(const Layout layout, const Transpose a_transpose,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xgemv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoGemv(layout, a_transpose,
                    m, n,
                    alpha,
@@ -713,7 +699,6 @@ StatusCode Gbmv(const Layout layout, const Transpose a_transpose,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xgbmv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoGbmv(layout, a_transpose,
                    m, n, kl, ku,
                    alpha,
@@ -778,7 +763,6 @@ StatusCode Hemv(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xhemv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoHemv(layout, triangle,
                    n,
                    alpha,
@@ -819,7 +803,6 @@ StatusCode Hbmv(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xhbmv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoHbmv(layout, triangle,
                    n, k,
                    alpha,
@@ -860,7 +843,6 @@ StatusCode Hpmv(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xhpmv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoHpmv(layout, triangle,
                    n,
                    alpha,
@@ -901,7 +883,6 @@ StatusCode Symv(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xsymv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSymv(layout, triangle,
                    n,
                    alpha,
@@ -950,7 +931,6 @@ StatusCode Sbmv(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xsbmv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSbmv(layout, triangle,
                    n, k,
                    alpha,
@@ -999,7 +979,6 @@ StatusCode Spmv(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xspmv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSpmv(layout, triangle,
                    n,
                    alpha,
@@ -1045,7 +1024,6 @@ StatusCode Trmv(const Layout layout, const Triangle triangle, const Transpose a_
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xtrmv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoTrmv(layout, triangle, a_transpose, diagonal,
                    n,
                    Buffer<T>(a_buffer), a_offset, a_ld,
@@ -1089,7 +1067,6 @@ StatusCode Tbmv(const Layout layout, const Triangle triangle, const Transpose a_
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xtbmv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoTbmv(layout, triangle, a_transpose, diagonal,
                    n, k,
                    Buffer<T>(a_buffer), a_offset, a_ld,
@@ -1133,7 +1110,6 @@ StatusCode Tpmv(const Layout layout, const Triangle triangle, const Transpose a_
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xtpmv<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoTpmv(layout, triangle, a_transpose, diagonal,
                    n,
                    Buffer<T>(ap_buffer), ap_offset,
@@ -1269,7 +1245,6 @@ StatusCode Ger(const Layout layout,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xger<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoGer(layout,
                   m, n,
                   alpha,
@@ -1313,7 +1288,6 @@ StatusCode Geru(const Layout layout,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xgeru<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoGeru(layout,
                    m, n,
                    alpha,
@@ -1350,7 +1324,6 @@ StatusCode Gerc(const Layout layout,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xgerc<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoGerc(layout,
                    m, n,
                    alpha,
@@ -1386,7 +1359,6 @@ StatusCode Her(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xher<std::complex<T>,T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoHer(layout, triangle,
                   n,
                   alpha,
@@ -1419,7 +1391,6 @@ StatusCode Hpr(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xhpr<std::complex<T>,T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoHpr(layout, triangle,
                   n,
                   alpha,
@@ -1453,7 +1424,6 @@ StatusCode Her2(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xher2<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoHer2(layout, triangle,
                    n,
                    alpha,
@@ -1490,7 +1460,6 @@ StatusCode Hpr2(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xhpr2<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoHpr2(layout, triangle,
                    n,
                    alpha,
@@ -1526,7 +1495,6 @@ StatusCode Syr(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xsyr<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSyr(layout, triangle,
                   n,
                   alpha,
@@ -1565,7 +1533,6 @@ StatusCode Spr(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xspr<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSpr(layout, triangle,
                   n,
                   alpha,
@@ -1605,7 +1572,6 @@ StatusCode Syr2(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xsyr2<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSyr2(layout, triangle,
                    n,
                    alpha,
@@ -1649,7 +1615,6 @@ StatusCode Spr2(const Layout layout, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xspr2<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSpr2(layout, triangle,
                    n,
                    alpha,
@@ -1698,7 +1663,6 @@ StatusCode Gemm(const Layout layout, const Transpose a_transpose, const Transpos
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xgemm<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoGemm(layout, a_transpose, b_transpose,
                    m, n, k,
                    alpha,
@@ -1763,7 +1727,6 @@ StatusCode Symm(const Layout layout, const Side side, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xsymm<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSymm(layout, side, triangle,
                    m, n,
                    alpha,
@@ -1828,7 +1791,6 @@ StatusCode Hemm(const Layout layout, const Side side, const Triangle triangle,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xhemm<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoHemm(layout, side, triangle,
                    m, n,
                    alpha,
@@ -1868,7 +1830,6 @@ StatusCode Syrk(const Layout layout, const Triangle triangle, const Transpose a_
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xsyrk<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSyrk(layout, triangle, a_transpose,
                    n, k,
                    alpha,
@@ -1926,7 +1887,6 @@ StatusCode Herk(const Layout layout, const Triangle triangle, const Transpose a_
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xherk<std::complex<T>,T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoHerk(layout, triangle, a_transpose,
                    n, k,
                    alpha,
@@ -1964,7 +1924,6 @@ StatusCode Syr2k(const Layout layout, const Triangle triangle, const Transpose a
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xsyr2k<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoSyr2k(layout, triangle, ab_transpose,
                     n, k,
                     alpha,
@@ -2029,7 +1988,6 @@ StatusCode Her2k(const Layout layout, const Triangle triangle, const Transpose a
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xher2k<T,U>(queue_cpp, event);
-    routine.SetUp();
     routine.DoHer2k(layout, triangle, ab_transpose,
                     n, k,
                     alpha,
@@ -2068,7 +2026,6 @@ StatusCode Trmm(const Layout layout, const Side side, const Triangle triangle, c
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xtrmm<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoTrmm(layout, side, triangle, a_transpose, diagonal,
                    m, n,
                    alpha,
@@ -2164,7 +2121,6 @@ StatusCode Omatcopy(const Layout layout, const Transpose a_transpose,
   try {
     auto queue_cpp = Queue(*queue);
     auto routine = Xomatcopy<T>(queue_cpp, event);
-    routine.SetUp();
     routine.DoOmatcopy(layout, a_transpose,
                        m, n,
                        alpha,
@@ -2225,57 +2181,57 @@ StatusCode FillCache(const cl_device_id device) {
     auto queue = Queue(context, device_cpp);
 
     // Runs all the level 1 set-up functions
-    Xswap<float>(queue, nullptr).SetUp(); Xswap<double>(queue, nullptr).SetUp(); Xswap<float2>(queue, nullptr).SetUp(); Xswap<double2>(queue, nullptr).SetUp();
-    Xswap<float>(queue, nullptr).SetUp(); Xswap<double>(queue, nullptr).SetUp(); Xswap<float2>(queue, nullptr).SetUp(); Xswap<double2>(queue, nullptr).SetUp();
-    Xscal<float>(queue, nullptr).SetUp(); Xscal<double>(queue, nullptr).SetUp(); Xscal<float2>(queue, nullptr).SetUp(); Xscal<double2>(queue, nullptr).SetUp();
-    Xcopy<float>(queue, nullptr).SetUp(); Xcopy<double>(queue, nullptr).SetUp(); Xcopy<float2>(queue, nullptr).SetUp(); Xcopy<double2>(queue, nullptr).SetUp();
-    Xaxpy<float>(queue, nullptr).SetUp(); Xaxpy<double>(queue, nullptr).SetUp(); Xaxpy<float2>(queue, nullptr).SetUp(); Xaxpy<double2>(queue, nullptr).SetUp();
-    Xdot<float>(queue, nullptr).SetUp(); Xdot<double>(queue, nullptr).SetUp();
-    Xdotu<float2>(queue, nullptr).SetUp(); Xdotu<double2>(queue, nullptr).SetUp();
-    Xdotc<float2>(queue, nullptr).SetUp(); Xdotc<double2>(queue, nullptr).SetUp();
-    Xnrm2<float>(queue, nullptr).SetUp(); Xnrm2<double>(queue, nullptr).SetUp(); Xnrm2<float2>(queue, nullptr).SetUp(); Xnrm2<double2>(queue, nullptr).SetUp();
-    Xasum<float>(queue, nullptr).SetUp(); Xasum<double>(queue, nullptr).SetUp(); Xasum<float2>(queue, nullptr).SetUp(); Xasum<double2>(queue, nullptr).SetUp();
-    Xsum<float>(queue, nullptr).SetUp(); Xsum<double>(queue, nullptr).SetUp(); Xsum<float2>(queue, nullptr).SetUp(); Xsum<double2>(queue, nullptr).SetUp();
-    Xamax<float>(queue, nullptr).SetUp(); Xamax<double>(queue, nullptr).SetUp(); Xamax<float2>(queue, nullptr).SetUp(); Xamax<double2>(queue, nullptr).SetUp();
-    Xmax<float>(queue, nullptr).SetUp(); Xmax<double>(queue, nullptr).SetUp(); Xmax<float2>(queue, nullptr).SetUp(); Xmax<double2>(queue, nullptr).SetUp();
-    Xmin<float>(queue, nullptr).SetUp(); Xmin<double>(queue, nullptr).SetUp(); Xmin<float2>(queue, nullptr).SetUp(); Xmin<double2>(queue, nullptr).SetUp();
+    Xswap<float>(queue, nullptr); Xswap<double>(queue, nullptr); Xswap<float2>(queue, nullptr); Xswap<double2>(queue, nullptr);
+    Xswap<float>(queue, nullptr); Xswap<double>(queue, nullptr); Xswap<float2>(queue, nullptr); Xswap<double2>(queue, nullptr);
+    Xscal<float>(queue, nullptr); Xscal<double>(queue, nullptr); Xscal<float2>(queue, nullptr); Xscal<double2>(queue, nullptr);
+    Xcopy<float>(queue, nullptr); Xcopy<double>(queue, nullptr); Xcopy<float2>(queue, nullptr); Xcopy<double2>(queue, nullptr);
+    Xaxpy<float>(queue, nullptr); Xaxpy<double>(queue, nullptr); Xaxpy<float2>(queue, nullptr); Xaxpy<double2>(queue, nullptr);
+    Xdot<float>(queue, nullptr); Xdot<double>(queue, nullptr);
+    Xdotu<float2>(queue, nullptr); Xdotu<double2>(queue, nullptr);
+    Xdotc<float2>(queue, nullptr); Xdotc<double2>(queue, nullptr);
+    Xnrm2<float>(queue, nullptr); Xnrm2<double>(queue, nullptr); Xnrm2<float2>(queue, nullptr); Xnrm2<double2>(queue, nullptr);
+    Xasum<float>(queue, nullptr); Xasum<double>(queue, nullptr); Xasum<float2>(queue, nullptr); Xasum<double2>(queue, nullptr);
+    Xsum<float>(queue, nullptr); Xsum<double>(queue, nullptr); Xsum<float2>(queue, nullptr); Xsum<double2>(queue, nullptr);
+    Xamax<float>(queue, nullptr); Xamax<double>(queue, nullptr); Xamax<float2>(queue, nullptr); Xamax<double2>(queue, nullptr);
+    Xmax<float>(queue, nullptr); Xmax<double>(queue, nullptr); Xmax<float2>(queue, nullptr); Xmax<double2>(queue, nullptr);
+    Xmin<float>(queue, nullptr); Xmin<double>(queue, nullptr); Xmin<float2>(queue, nullptr); Xmin<double2>(queue, nullptr);
 
     // Runs all the level 2 set-up functions
-    Xgemv<float>(queue, nullptr).SetUp(); Xgemv<double>(queue, nullptr).SetUp(); Xgemv<float2>(queue, nullptr).SetUp(); Xgemv<double2>(queue, nullptr).SetUp();
-    Xgbmv<float>(queue, nullptr).SetUp(); Xgbmv<double>(queue, nullptr).SetUp(); Xgbmv<float2>(queue, nullptr).SetUp(); Xgbmv<double2>(queue, nullptr).SetUp();
-    Xhemv<float2>(queue, nullptr).SetUp(); Xhemv<double2>(queue, nullptr).SetUp();
-    Xhbmv<float2>(queue, nullptr).SetUp(); Xhbmv<double2>(queue, nullptr).SetUp();
-    Xhpmv<float2>(queue, nullptr).SetUp(); Xhpmv<double2>(queue, nullptr).SetUp();
-    Xsymv<float>(queue, nullptr).SetUp(); Xsymv<double>(queue, nullptr).SetUp();
-    Xsbmv<float>(queue, nullptr).SetUp(); Xsbmv<double>(queue, nullptr).SetUp();
-    Xspmv<float>(queue, nullptr).SetUp(); Xspmv<double>(queue, nullptr).SetUp();
-    Xtrmv<float>(queue, nullptr).SetUp(); Xtrmv<double>(queue, nullptr).SetUp(); Xtrmv<float2>(queue, nullptr).SetUp(); Xtrmv<double2>(queue, nullptr).SetUp();
-    Xtbmv<float>(queue, nullptr).SetUp(); Xtbmv<double>(queue, nullptr).SetUp(); Xtbmv<float2>(queue, nullptr).SetUp(); Xtbmv<double2>(queue, nullptr).SetUp();
-    Xtpmv<float>(queue, nullptr).SetUp(); Xtpmv<double>(queue, nullptr).SetUp(); Xtpmv<float2>(queue, nullptr).SetUp(); Xtpmv<double2>(queue, nullptr).SetUp();
-    Xger<float>(queue, nullptr).SetUp(); Xger<double>(queue, nullptr).SetUp();
-    Xgeru<float2>(queue, nullptr).SetUp(); Xgeru<double2>(queue, nullptr).SetUp();
-    Xgerc<float2>(queue, nullptr).SetUp(); Xgerc<double2>(queue, nullptr).SetUp();
-    Xher<float2,float>(queue, nullptr).SetUp(); Xher<double2,double>(queue, nullptr).SetUp();
-    Xhpr<float2,float>(queue, nullptr).SetUp(); Xhpr<double2,double>(queue, nullptr).SetUp();
-    Xher2<float2>(queue, nullptr).SetUp(); Xher2<double2>(queue, nullptr).SetUp();
-    Xhpr2<float2>(queue, nullptr).SetUp(); Xhpr2<double2>(queue, nullptr).SetUp();
-    Xsyr<float>(queue, nullptr).SetUp(); Xsyr<double>(queue, nullptr).SetUp();
-    Xspr<float>(queue, nullptr).SetUp(); Xspr<double>(queue, nullptr).SetUp();
-    Xsyr2<float>(queue, nullptr).SetUp(); Xsyr2<double>(queue, nullptr).SetUp();
-    Xspr2<float>(queue, nullptr).SetUp(); Xspr2<double>(queue, nullptr).SetUp();
+    Xgemv<float>(queue, nullptr); Xgemv<double>(queue, nullptr); Xgemv<float2>(queue, nullptr); Xgemv<double2>(queue, nullptr);
+    Xgbmv<float>(queue, nullptr); Xgbmv<double>(queue, nullptr); Xgbmv<float2>(queue, nullptr); Xgbmv<double2>(queue, nullptr);
+    Xhemv<float2>(queue, nullptr); Xhemv<double2>(queue, nullptr);
+    Xhbmv<float2>(queue, nullptr); Xhbmv<double2>(queue, nullptr);
+    Xhpmv<float2>(queue, nullptr); Xhpmv<double2>(queue, nullptr);
+    Xsymv<float>(queue, nullptr); Xsymv<double>(queue, nullptr);
+    Xsbmv<float>(queue, nullptr); Xsbmv<double>(queue, nullptr);
+    Xspmv<float>(queue, nullptr); Xspmv<double>(queue, nullptr);
+    Xtrmv<float>(queue, nullptr); Xtrmv<double>(queue, nullptr); Xtrmv<float2>(queue, nullptr); Xtrmv<double2>(queue, nullptr);
+    Xtbmv<float>(queue, nullptr); Xtbmv<double>(queue, nullptr); Xtbmv<float2>(queue, nullptr); Xtbmv<double2>(queue, nullptr);
+    Xtpmv<float>(queue, nullptr); Xtpmv<double>(queue, nullptr); Xtpmv<float2>(queue, nullptr); Xtpmv<double2>(queue, nullptr);
+    Xger<float>(queue, nullptr); Xger<double>(queue, nullptr);
+    Xgeru<float2>(queue, nullptr); Xgeru<double2>(queue, nullptr);
+    Xgerc<float2>(queue, nullptr); Xgerc<double2>(queue, nullptr);
+    Xher<float2,float>(queue, nullptr); Xher<double2,double>(queue, nullptr);
+    Xhpr<float2,float>(queue, nullptr); Xhpr<double2,double>(queue, nullptr);
+    Xher2<float2>(queue, nullptr); Xher2<double2>(queue, nullptr);
+    Xhpr2<float2>(queue, nullptr); Xhpr2<double2>(queue, nullptr);
+    Xsyr<float>(queue, nullptr); Xsyr<double>(queue, nullptr);
+    Xspr<float>(queue, nullptr); Xspr<double>(queue, nullptr);
+    Xsyr2<float>(queue, nullptr); Xsyr2<double>(queue, nullptr);
+    Xspr2<float>(queue, nullptr); Xspr2<double>(queue, nullptr);
 
     // Runs all the level 3 set-up functions
-    Xgemm<float>(queue, nullptr).SetUp(); Xgemm<double>(queue, nullptr).SetUp(); Xgemm<float2>(queue, nullptr).SetUp(); Xgemm<double2>(queue, nullptr).SetUp();
-    Xsymm<float>(queue, nullptr).SetUp(); Xsymm<double>(queue, nullptr).SetUp(); Xsymm<float2>(queue, nullptr).SetUp(); Xsymm<double2>(queue, nullptr).SetUp();
-    Xhemm<float2>(queue, nullptr).SetUp(); Xhemm<double2>(queue, nullptr).SetUp();
-    Xsyrk<float>(queue, nullptr).SetUp(); Xsyrk<double>(queue, nullptr).SetUp(); Xsyrk<float2>(queue, nullptr).SetUp(); Xsyrk<double2>(queue, nullptr).SetUp();
-    Xherk<float2,float>(queue, nullptr).SetUp(); Xherk<double2,double>(queue, nullptr).SetUp();
-    Xsyr2k<float>(queue, nullptr).SetUp(); Xsyr2k<double>(queue, nullptr).SetUp(); Xsyr2k<float2>(queue, nullptr).SetUp(); Xsyr2k<double2>(queue, nullptr).SetUp();
-    Xher2k<float2,float>(queue, nullptr).SetUp(); Xher2k<double2,double>(queue, nullptr).SetUp();
-    Xtrmm<float>(queue, nullptr).SetUp(); Xtrmm<double>(queue, nullptr).SetUp(); Xtrmm<float2>(queue, nullptr).SetUp(); Xtrmm<double2>(queue, nullptr).SetUp();
+    Xgemm<float>(queue, nullptr); Xgemm<double>(queue, nullptr); Xgemm<float2>(queue, nullptr); Xgemm<double2>(queue, nullptr);
+    Xsymm<float>(queue, nullptr); Xsymm<double>(queue, nullptr); Xsymm<float2>(queue, nullptr); Xsymm<double2>(queue, nullptr);
+    Xhemm<float2>(queue, nullptr); Xhemm<double2>(queue, nullptr);
+    Xsyrk<float>(queue, nullptr); Xsyrk<double>(queue, nullptr); Xsyrk<float2>(queue, nullptr); Xsyrk<double2>(queue, nullptr);
+    Xherk<float2,float>(queue, nullptr); Xherk<double2,double>(queue, nullptr);
+    Xsyr2k<float>(queue, nullptr); Xsyr2k<double>(queue, nullptr); Xsyr2k<float2>(queue, nullptr); Xsyr2k<double2>(queue, nullptr);
+    Xher2k<float2,float>(queue, nullptr); Xher2k<double2,double>(queue, nullptr);
+    Xtrmm<float>(queue, nullptr); Xtrmm<double>(queue, nullptr); Xtrmm<float2>(queue, nullptr); Xtrmm<double2>(queue, nullptr);
 
     // Runs all the level 3 set-up functions
-    Xomatcopy<float>(queue, nullptr).SetUp(); Xomatcopy<double>(queue, nullptr).SetUp(); Xomatcopy<float2>(queue, nullptr).SetUp(); Xomatcopy<double2>(queue, nullptr).SetUp();
+    Xomatcopy<float>(queue, nullptr); Xomatcopy<double>(queue, nullptr); Xomatcopy<float2>(queue, nullptr); Xomatcopy<double2>(queue, nullptr);
 
   } catch (...) { return DispatchException(); }
   return StatusCode::kSuccess;
diff --git a/src/routine.cpp b/src/routine.cpp
index 5e3a9dfe..acafb0d2 100644
--- a/src/routine.cpp
+++ b/src/routine.cpp
@@ -21,10 +21,11 @@
 namespace clblast {
 // =================================================================================================
 
-// Constructor: not much here, because no status codes can be returned
+// The constructor does all heavy work, errors are returned as exceptions
 Routine::Routine(Queue &queue, EventPointer event, const std::string &name,
                  const std::vector<std::string> &routines, const Precision precision,
-                 const std::vector<const Database::DatabaseEntry*> &userDatabase):
+                 const std::vector<const Database::DatabaseEntry*> &userDatabase,
+                 std::initializer_list<const char *> source):
     precision_(precision),
     routine_name_(name),
     queue_(queue),
@@ -33,12 +34,6 @@ Routine::Routine(Queue &queue, EventPointer event, const std::string &name,
     device_(queue_.GetDevice()),
     device_name_(device_.Name()),
     db_(queue_, routines, precision_, userDatabase) {
-}
-
-// =================================================================================================
-
-// Separate set-up function to allow for status codes to be returned
-void Routine::SetUp() {
 
   // Queries the cache to see whether or not the program (context-specific) is already there
   if (ProgramIsInCache(context_, precision_, routine_name_)) { return; }
@@ -77,37 +72,39 @@ void Routine::SetUp() {
     }
   }
 
-  // Loads the common header (typedefs and defines and such)
-  std::string common_header =
-    #include "kernels/common.opencl"
-  ;
-
   // Collects the parameters for this device in the form of defines, and adds the precision
-  auto defines = db_.GetDefines();
-  defines += "#define PRECISION "+ToString(static_cast<int>(precision_))+"\n";
+  auto source_string = db_.GetDefines();
+  source_string += "#define PRECISION "+ToString(static_cast<int>(precision_))+"\n";
 
   // Adds the name of the routine as a define
-  defines += "#define ROUTINE_"+routine_name_+"\n";
+  source_string += "#define ROUTINE_"+routine_name_+"\n";
 
   // For specific devices, use the non-IEE754 compilant OpenCL mad() instruction. This can improve
   // performance, but might result in a reduced accuracy.
   if (device_.IsAMD() && device_.IsGPU()) {
-    defines += "#define USE_CL_MAD 1\n";
+    source_string += "#define USE_CL_MAD 1\n";
   }
 
   // For specific devices, use staggered/shuffled workgroup indices.
   if (device_.IsAMD() && device_.IsGPU()) {
-    defines += "#define USE_STAGGERED_INDICES 1\n";
+    source_string += "#define USE_STAGGERED_INDICES 1\n";
   }
 
   // For specific devices add a global synchronisation barrier to the GEMM kernel to optimize
   // performance through better cache behaviour
   if (device_.IsARM() && device_.IsGPU()) {
-    defines += "#define GLOBAL_MEM_FENCE 1\n";
+    source_string += "#define GLOBAL_MEM_FENCE 1\n";
   }
 
-  // Combines everything together into a single source string
-  const auto source_string = defines + common_header + source_string_;
+  // Loads the common header (typedefs and defines and such)
+  source_string +=
+    #include "kernels/common.opencl"
+  ;
+
+  // Adds routine-specific code to the constructed source string
+  for (const char *s: source) {
+    source_string += s;
+  }
 
   // Prints details of the routine to compile in case of debugging in verbose mode
   #ifdef VERBOSE
diff --git a/src/routine.hpp b/src/routine.hpp
index c2bcdaff..f4ad435e 100644
--- a/src/routine.hpp
+++ b/src/routine.hpp
@@ -34,21 +34,19 @@ class Routine {
 
   // Base class constructor. The user database is an optional extra database to override the
   // built-in database.
+  // All heavy preparation work is done inside this constructor.
   explicit Routine(Queue &queue, EventPointer event, const std::string &name,
                    const std::vector<std::string> &routines, const Precision precision,
-                   const std::vector<const Database::DatabaseEntry*> &userDatabase = {});
-
-  // Set-up phase of the kernel
-  void SetUp();
+                   const std::vector<const Database::DatabaseEntry*> &userDatabase,
+                   std::initializer_list<const char *> source);
 
  protected:
 
   // Non-static variable for the precision
   const Precision precision_;
 
-  // The routine's name and its kernel-source in string form
+  // The routine's name
   const std::string routine_name_;
-  std::string source_string_;
 
   // The OpenCL objects, accessible only from derived classes
   Queue queue_;
diff --git a/src/routines/level1/xamax.cpp b/src/routines/level1/xamax.cpp
index 8307188b..e9efa1a7 100644
--- a/src/routines/level1/xamax.cpp
+++ b/src/routines/level1/xamax.cpp
@@ -22,10 +22,9 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xamax<T>::Xamax(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xdot"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xdot"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level1/xamax.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level1/xasum.cpp b/src/routines/level1/xasum.cpp
index 9dde7a87..a242a5fa 100644
--- a/src/routines/level1/xasum.cpp
+++ b/src/routines/level1/xasum.cpp
@@ -22,10 +22,9 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xasum<T>::Xasum(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xdot"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xdot"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level1/xasum.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level1/xaxpy.cpp b/src/routines/level1/xaxpy.cpp
index cbcbb3cd..5436c5b7 100644
--- a/src/routines/level1/xaxpy.cpp
+++ b/src/routines/level1/xaxpy.cpp
@@ -22,11 +22,10 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xaxpy<T>::Xaxpy(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xaxpy"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xaxpy"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level1/level1.opencl"
     #include "../../kernels/level1/xaxpy.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level1/xcopy.cpp b/src/routines/level1/xcopy.cpp
index 3bfbada6..d86200c0 100644
--- a/src/routines/level1/xcopy.cpp
+++ b/src/routines/level1/xcopy.cpp
@@ -22,11 +22,10 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xcopy<T>::Xcopy(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xaxpy"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xaxpy"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level1/level1.opencl"
     #include "../../kernels/level1/xcopy.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level1/xdot.cpp b/src/routines/level1/xdot.cpp
index e0d297f8..9d718913 100644
--- a/src/routines/level1/xdot.cpp
+++ b/src/routines/level1/xdot.cpp
@@ -22,10 +22,9 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xdot<T>::Xdot(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xdot"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xdot"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level1/xdot.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level1/xnrm2.cpp b/src/routines/level1/xnrm2.cpp
index eb795498..373820a4 100644
--- a/src/routines/level1/xnrm2.cpp
+++ b/src/routines/level1/xnrm2.cpp
@@ -22,10 +22,9 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xnrm2<T>::Xnrm2(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xdot"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xdot"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level1/xnrm2.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level1/xscal.cpp b/src/routines/level1/xscal.cpp
index ed126879..17410f01 100644
--- a/src/routines/level1/xscal.cpp
+++ b/src/routines/level1/xscal.cpp
@@ -22,11 +22,10 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xscal<T>::Xscal(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xaxpy"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xaxpy"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level1/level1.opencl"
     #include "../../kernels/level1/xscal.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level1/xswap.cpp b/src/routines/level1/xswap.cpp
index 2f2c0370..c9b97dc9 100644
--- a/src/routines/level1/xswap.cpp
+++ b/src/routines/level1/xswap.cpp
@@ -22,11 +22,10 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xswap<T>::Xswap(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xaxpy"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xaxpy"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level1/level1.opencl"
     #include "../../kernels/level1/xswap.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level2/xgemv.cpp b/src/routines/level2/xgemv.cpp
index 97dcd8ef..7b4c2e8f 100644
--- a/src/routines/level2/xgemv.cpp
+++ b/src/routines/level2/xgemv.cpp
@@ -22,11 +22,10 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xgemv<T>::Xgemv(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Pad", "Xgemv", "XgemvFast", "XgemvFastRot"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Pad", "Xgemv", "XgemvFast", "XgemvFastRot"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level2/xgemv.opencl"
     #include "../../kernels/level2/xgemv_fast.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level2/xger.cpp b/src/routines/level2/xger.cpp
index f22131bb..d16ebd11 100644
--- a/src/routines/level2/xger.cpp
+++ b/src/routines/level2/xger.cpp
@@ -22,11 +22,10 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xger<T>::Xger(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xger"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xger"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level2/level2.opencl"
     #include "../../kernels/level2/xger.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level2/xher.cpp b/src/routines/level2/xher.cpp
index 4cf27702..6c334e63 100644
--- a/src/routines/level2/xher.cpp
+++ b/src/routines/level2/xher.cpp
@@ -21,11 +21,10 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T, typename U>
 Xher<T,U>::Xher(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xger"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xger"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level2/level2.opencl"
     #include "../../kernels/level2/xher.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level2/xher2.cpp b/src/routines/level2/xher2.cpp
index c93585de..11e2c871 100644
--- a/src/routines/level2/xher2.cpp
+++ b/src/routines/level2/xher2.cpp
@@ -21,11 +21,10 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xher2<T>::Xher2(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Xger"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Xger"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level2/level2.opencl"
     #include "../../kernels/level2/xher2.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level3/xgemm.cpp b/src/routines/level3/xgemm.cpp
index a6f7c286..4f70dc7a 100644
--- a/src/routines/level3/xgemm.cpp
+++ b/src/routines/level3/xgemm.cpp
@@ -24,8 +24,7 @@ template <typename T>
 Xgemm<T>::Xgemm(Queue &queue, EventPointer event, const std::string &name):
     Routine(queue, event, name,
             {"Copy","Pad","Transpose","Padtranspose","Xgemm","XgemmDirect","KernelSelection"},
-            PrecisionValue<T>()) {
-  source_string_ =
+            PrecisionValue<T>(), {}, {
     #include "../../kernels/level3/level3.opencl"
     #include "../../kernels/level3/copy_fast.opencl"
     #include "../../kernels/level3/copy_pad.opencl"
@@ -37,13 +36,11 @@ Xgemm<T>::Xgemm(Queue &queue, EventPointer event, const std::string &name):
     #include "../../kernels/level3/xgemm_direct_part1.opencl"
     #include "../../kernels/level3/xgemm_direct_part2.opencl"
     #include "../../kernels/level3/xgemm_direct_part3.opencl"
-  ;
-  auto source_string_part_2 = // separated in two parts to prevent C1091 in MSVC 2013
+    , // separated in two parts to prevent C1091 in MSVC 2013
     #include "../../kernels/level3/xgemm_part1.opencl"
     #include "../../kernels/level3/xgemm_part2.opencl"
     #include "../../kernels/level3/xgemm_part3.opencl"
-  ;
-  source_string_ += source_string_part_2;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level3/xher2k.cpp b/src/routines/level3/xher2k.cpp
index a326dfbe..7244c848 100644
--- a/src/routines/level3/xher2k.cpp
+++ b/src/routines/level3/xher2k.cpp
@@ -22,8 +22,7 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T, typename U>
 Xher2k<T,U>::Xher2k(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level3/level3.opencl"
     #include "../../kernels/level3/copy_fast.opencl"
     #include "../../kernels/level3/copy_pad.opencl"
@@ -32,7 +31,7 @@ Xher2k<T,U>::Xher2k(Queue &queue, EventPointer event, const std::string &name):
     #include "../../kernels/level3/xgemm_part1.opencl"
     #include "../../kernels/level3/xgemm_part2.opencl"
     #include "../../kernels/level3/xgemm_part3.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level3/xherk.cpp b/src/routines/level3/xherk.cpp
index 6e36714e..865c6c37 100644
--- a/src/routines/level3/xherk.cpp
+++ b/src/routines/level3/xherk.cpp
@@ -22,8 +22,7 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T, typename U>
 Xherk<T,U>::Xherk(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level3/level3.opencl"
     #include "../../kernels/level3/copy_fast.opencl"
     #include "../../kernels/level3/copy_pad.opencl"
@@ -32,7 +31,7 @@ Xherk<T,U>::Xherk(Queue &queue, EventPointer event, const std::string &name):
     #include "../../kernels/level3/xgemm_part1.opencl"
     #include "../../kernels/level3/xgemm_part2.opencl"
     #include "../../kernels/level3/xgemm_part3.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level3/xsyr2k.cpp b/src/routines/level3/xsyr2k.cpp
index b10ee586..826854a8 100644
--- a/src/routines/level3/xsyr2k.cpp
+++ b/src/routines/level3/xsyr2k.cpp
@@ -22,8 +22,7 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xsyr2k<T>::Xsyr2k(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level3/level3.opencl"
     #include "../../kernels/level3/copy_fast.opencl"
     #include "../../kernels/level3/copy_pad.opencl"
@@ -32,7 +31,7 @@ Xsyr2k<T>::Xsyr2k(Queue &queue, EventPointer event, const std::string &name):
     #include "../../kernels/level3/xgemm_part1.opencl"
     #include "../../kernels/level3/xgemm_part2.opencl"
     #include "../../kernels/level3/xgemm_part3.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/level3/xsyrk.cpp b/src/routines/level3/xsyrk.cpp
index 93fd4666..9aa8ca2d 100644
--- a/src/routines/level3/xsyrk.cpp
+++ b/src/routines/level3/xsyrk.cpp
@@ -22,8 +22,7 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xsyrk<T>::Xsyrk(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level3/level3.opencl"
     #include "../../kernels/level3/copy_fast.opencl"
     #include "../../kernels/level3/copy_pad.opencl"
@@ -32,7 +31,7 @@ Xsyrk<T>::Xsyrk(Queue &queue, EventPointer event, const std::string &name):
     #include "../../kernels/level3/xgemm_part1.opencl"
     #include "../../kernels/level3/xgemm_part2.opencl"
     #include "../../kernels/level3/xgemm_part3.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
diff --git a/src/routines/levelx/xomatcopy.cpp b/src/routines/levelx/xomatcopy.cpp
index e053c076..875ca7d2 100644
--- a/src/routines/levelx/xomatcopy.cpp
+++ b/src/routines/levelx/xomatcopy.cpp
@@ -22,14 +22,13 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xomatcopy<T>::Xomatcopy(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose"}, PrecisionValue<T>()) {
-  source_string_ =
+    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose"}, PrecisionValue<T>(), {}, {
     #include "../../kernels/level3/level3.opencl"
     #include "../../kernels/level3/copy_fast.opencl"
     #include "../../kernels/level3/copy_pad.opencl"
     #include "../../kernels/level3/transpose_fast.opencl"
     #include "../../kernels/level3/transpose_pad.opencl"
-  ;
+    }) {
 }
 
 // =================================================================================================
-- 
cgit v1.2.3