Merge pull request #19 from CNugteren/basic_level2_routines

Level-2 routines: HEMV and SYMV

5 files changed, 368 insertions, 64 deletions

diff --git a/src/clblast.cc b/src/clblast.cc
index eddb8022..12c7b880 100644
--- a/src/clblast.cc
+++ b/src/clblast.cc
@@ -22,6 +22,8 @@
 
 // BLAS level-2 includes
 #include "internal/routines/level2/xgemv.h"
+#include "internal/routines/level2/xhemv.h"
+#include "internal/routines/level2/xsymv.h"
 
 // BLAS level-3 includes
 #include "internal/routines/level3/xgemm.h"
@@ -36,6 +38,7 @@
 namespace clblast {
 // =================================================================================================
 // BLAS level-1 (vector-vector) routines
+// =================================================================================================
 
 // AXPY
 template <typename T>
@@ -75,6 +78,7 @@ template StatusCode Axpy<double2>(const size_t, const double2,
 
 // =================================================================================================
 // BLAS level-2 (matrix-vector) routines
+// =================================================================================================
 
 // GEMV
 template <typename T>
@@ -125,7 +129,84 @@ template StatusCode Gemv<double2>(const Layout, const Transpose,
                                   cl_command_queue*, cl_event*);
 
 // =================================================================================================
+
+// HEMV
+template <typename T>
+StatusCode Hemv(const Layout layout, const Triangle triangle,
+                const size_t n, const T alpha,
+                const cl_mem a_buffer, const size_t a_offset, const size_t a_ld,
+                const cl_mem x_buffer, const size_t x_offset, const size_t x_inc, const T beta,
+                cl_mem y_buffer, const size_t y_offset, const size_t y_inc,
+                cl_command_queue* queue, cl_event* event) {
+
+  auto queue_cpp = Queue(*queue);
+  auto event_cpp = Event(*event);
+  auto routine = Xhemv<T>(queue_cpp, event_cpp);
+
+  // Compiles the routine's device kernels
+  auto status = routine.SetUp();
+  if (status != StatusCode::kSuccess) { return status; }
+
+  // Runs the routine
+  return routine.DoHemv(layout, triangle, n, alpha,
+                        Buffer<T>(a_buffer), a_offset, a_ld,
+                        Buffer<T>(x_buffer), x_offset, x_inc, beta,
+                        Buffer<T>(y_buffer), y_offset, y_inc);
+}
+template StatusCode Hemv<float2>(const Layout, const Triangle,
+                                 const size_t, const float2,
+                                 const cl_mem, const size_t, const size_t,
+                                 const cl_mem, const size_t, const size_t, const float2,
+                                 cl_mem, const size_t, const size_t,
+                                 cl_command_queue*, cl_event*);
+template StatusCode Hemv<double2>(const Layout, const Triangle,
+                                  const size_t, const double2,
+                                  const cl_mem, const size_t, const size_t,
+                                  const cl_mem, const size_t, const size_t, const double2,
+                                  cl_mem, const size_t, const size_t,
+                                  cl_command_queue*, cl_event*);
+
+// =================================================================================================
+
+// SYMV
+template <typename T>
+StatusCode Symv(const Layout layout, const Triangle triangle,
+                const size_t n, const T alpha,
+                const cl_mem a_buffer, const size_t a_offset, const size_t a_ld,
+                const cl_mem x_buffer, const size_t x_offset, const size_t x_inc, const T beta,
+                cl_mem y_buffer, const size_t y_offset, const size_t y_inc,
+                cl_command_queue* queue, cl_event* event) {
+
+  auto queue_cpp = Queue(*queue);
+  auto event_cpp = Event(*event);
+  auto routine = Xsymv<T>(queue_cpp, event_cpp);
+
+  // Compiles the routine's device kernels
+  auto status = routine.SetUp();
+  if (status != StatusCode::kSuccess) { return status; }
+
+  // Runs the routine
+  return routine.DoSymv(layout, triangle, n, alpha,
+                        Buffer<T>(a_buffer), a_offset, a_ld,
+                        Buffer<T>(x_buffer), x_offset, x_inc, beta,
+                        Buffer<T>(y_buffer), y_offset, y_inc);
+}
+template StatusCode Symv<float>(const Layout, const Triangle,
+                                const size_t, const float,
+                                const cl_mem, const size_t, const size_t,
+                                const cl_mem, const size_t, const size_t, const float,
+                                cl_mem, const size_t, const size_t,
+                                cl_command_queue*, cl_event*);
+template StatusCode Symv<double>(const Layout, const Triangle,
+                                 const size_t, const double,
+                                 const cl_mem, const size_t, const size_t,
+                                 const cl_mem, const size_t, const size_t, const double,
+                                 cl_mem, const size_t, const size_t,
+                                 cl_command_queue*, cl_event*);
+
+// =================================================================================================
 // BLAS level-3 (matrix-matrix) routines
+// =================================================================================================
 
 // GEMM
 template <typename T>
diff --git a/src/kernels/xgemv.opencl b/src/kernels/xgemv.opencl
index 65061717..1e12dd78 100644
--- a/src/kernels/xgemv.opencl
+++ b/src/kernels/xgemv.opencl
@@ -52,6 +52,53 @@ R"(
 
 // =================================================================================================
 
+// Data-widths for the 'fast' kernel
+#if VW2 == 1
+  typedef real realVF;
+#elif VW2 == 2
+  typedef real2 realVF;
+#elif VW2 == 4
+  typedef real4 realVF;
+#elif VW2 == 8
+  typedef real8 realVF;
+#elif VW2 == 16
+  typedef real16 realVF;
+#endif
+
+// Data-widths for the 'fast' kernel with rotated matrix
+#if VW3 == 1
+  typedef real realVFR;
+#elif VW3 == 2
+  typedef real2 realVFR;
+#elif VW3 == 4
+  typedef real4 realVFR;
+#elif VW3 == 8
+  typedef real8 realVFR;
+#elif VW3 == 16
+  typedef real16 realVFR;
+#endif
+
+// =================================================================================================
+// Defines how to load the input matrix in the regular case
+
+// Loads a scalar input value
+inline real LoadMatrixA(const __global real* restrict agm, const int x, const int y,
+                        const int a_ld, const int a_offset) {
+  return agm[x + a_ld*y + a_offset];
+}
+// Loads a vector input value (1/2)
+inline realVF LoadMatrixAVF(const __global realVF* restrict agm, const int x, const int y,
+                            const int a_ld) {
+  return agm[x + a_ld*y];
+}
+// Loads a vector input value (2/2): as before, but different data-type
+inline realVFR LoadMatrixAVFR(const __global realVFR* restrict agm, const int x, const int y,
+                              const int a_ld) {
+  return agm[x + a_ld*y];
+}
+
+// =================================================================================================
+
 // Full version of the kernel
 __attribute__((reqd_work_group_size(WGS1, 1, 1)))
 __kernel void Xgemv(const int m, const int n, const real alpha, const real beta,
@@ -96,7 +143,7 @@ __kernel void Xgemv(const int m, const int n, const real alpha, const real beta,
           #pragma unroll
           for (int kl=0; kl<WGS1; ++kl) {
             const int k = kwg + kl;
-            real value = agm[gid + a_ld*k + a_offset];
+            real value = LoadMatrixA(agm, gid, k, a_ld, a_offset);
             if (do_conjugate == 1) { COMPLEX_CONJUGATE(value); }
             MultiplyAdd(acc[w], xlm[kl], value);
           }
@@ -105,7 +152,7 @@ __kernel void Xgemv(const int m, const int n, const real alpha, const real beta,
           #pragma unroll
           for (int kl=0; kl<WGS1; ++kl) {
             const int k = kwg + kl;
-            real value = agm[k + a_ld*gid + a_offset];
+            real value = LoadMatrixA(agm, k, gid, a_ld, a_offset);
             if (do_conjugate == 1) { COMPLEX_CONJUGATE(value); }
             MultiplyAdd(acc[w], xlm[kl], value);
           }
@@ -127,7 +174,7 @@ __kernel void Xgemv(const int m, const int n, const real alpha, const real beta,
       if (a_rotated == 0) { // Not rotated
         #pragma unroll
         for (int k=n_floor; k<n; ++k) {
-          real value = agm[gid + a_ld*k + a_offset];
+          real value = LoadMatrixA(agm, gid, k, a_ld, a_offset);
           if (do_conjugate == 1) { COMPLEX_CONJUGATE(value); }
           MultiplyAdd(acc[w], xgm[k*x_inc + x_offset], value);
         }
@@ -135,7 +182,7 @@ __kernel void Xgemv(const int m, const int n, const real alpha, const real beta,
       else { // Transposed
         #pragma unroll
         for (int k=n_floor; k<n; ++k) {
-          real value = agm[k + a_ld*gid + a_offset];
+          real value = LoadMatrixA(agm, k, gid, a_ld, a_offset);
           if (do_conjugate == 1) { COMPLEX_CONJUGATE(value); }
           MultiplyAdd(acc[w], xgm[k*x_inc + x_offset], value);
         }
@@ -150,19 +197,6 @@ __kernel void Xgemv(const int m, const int n, const real alpha, const real beta,
 
 // =================================================================================================
 
-// Data-widths for the 'fast' kernel
-#if VW2 == 1
-  typedef real realVF;
-#elif VW2 == 2
-  typedef real2 realVF;
-#elif VW2 == 4
-  typedef real4 realVF;
-#elif VW2 == 8
-  typedef real8 realVF;
-#elif VW2 == 16
-  typedef real16 realVF;
-#endif
-
 // Faster version of the kernel, assuming that:
 // --> 'm' and 'n' are multiples of WGS2
 // --> 'a_offset' is 0
@@ -203,42 +237,43 @@ __kernel void XgemvFast(const int m, const int n, const real alpha, const real b
       #pragma unroll
       for (int w=0; w<WPT2/VW2; ++w) {
         const int gid = (WPT2/VW2)*get_global_id(0) + w;
+        realVF avec = LoadMatrixAVF(agm, gid, k, a_ld/VW2);
         #if VW2 == 1
-          MultiplyAdd(acc[VW2*w+0], xlm[kl], agm[gid + (a_ld/VW2)*k]);
+          MultiplyAdd(acc[VW2*w+0], xlm[kl], avec);
         #elif VW2 == 2
-          MultiplyAdd(acc[VW2*w+0], xlm[kl], agm[gid + (a_ld/VW2)*k].x);
-          MultiplyAdd(acc[VW2*w+1], xlm[kl], agm[gid + (a_ld/VW2)*k].y);
+          MultiplyAdd(acc[VW2*w+0], xlm[kl], avec.x);
+          MultiplyAdd(acc[VW2*w+1], xlm[kl], avec.y);
         #elif VW2 == 4
-          MultiplyAdd(acc[VW2*w+0], xlm[kl], agm[gid + (a_ld/VW2)*k].x);
-          MultiplyAdd(acc[VW2*w+1], xlm[kl], agm[gid + (a_ld/VW2)*k].y);
-          MultiplyAdd(acc[VW2*w+2], xlm[kl], agm[gid + (a_ld/VW2)*k].z);
-          MultiplyAdd(acc[VW2*w+3], xlm[kl], agm[gid + (a_ld/VW2)*k].w);
+          MultiplyAdd(acc[VW2*w+0], xlm[kl], avec.x);
+          MultiplyAdd(acc[VW2*w+1], xlm[kl], avec.y);
+          MultiplyAdd(acc[VW2*w+2], xlm[kl], avec.z);
+          MultiplyAdd(acc[VW2*w+3], xlm[kl], avec.w);
         #elif VW2 == 8
-          MultiplyAdd(acc[VW2*w+0], xlm[kl], agm[gid + (a_ld/VW2)*k].s0);
-          MultiplyAdd(acc[VW2*w+1], xlm[kl], agm[gid + (a_ld/VW2)*k].s1);
-          MultiplyAdd(acc[VW2*w+2], xlm[kl], agm[gid + (a_ld/VW2)*k].s2);
-          MultiplyAdd(acc[VW2*w+3], xlm[kl], agm[gid + (a_ld/VW2)*k].s3);
-          MultiplyAdd(acc[VW2*w+4], xlm[kl], agm[gid + (a_ld/VW2)*k].s4);
-          MultiplyAdd(acc[VW2*w+5], xlm[kl], agm[gid + (a_ld/VW2)*k].s5);
-          MultiplyAdd(acc[VW2*w+6], xlm[kl], agm[gid + (a_ld/VW2)*k].s6);
-          MultiplyAdd(acc[VW2*w+7], xlm[kl], agm[gid + (a_ld/VW2)*k].s7);
+          MultiplyAdd(acc[VW2*w+0], xlm[kl], avec.s0);
+          MultiplyAdd(acc[VW2*w+1], xlm[kl], avec.s1);
+          MultiplyAdd(acc[VW2*w+2], xlm[kl], avec.s2);
+          MultiplyAdd(acc[VW2*w+3], xlm[kl], avec.s3);
+          MultiplyAdd(acc[VW2*w+4], xlm[kl], avec.s4);
+          MultiplyAdd(acc[VW2*w+5], xlm[kl], avec.s5);
+          MultiplyAdd(acc[VW2*w+6], xlm[kl], avec.s6);
+          MultiplyAdd(acc[VW2*w+7], xlm[kl], avec.s7);
         #elif VW2 == 16
-          MultiplyAdd(acc[VW2*w+0], xlm[kl], agm[gid + (a_ld/VW2)*k].s0);
-          MultiplyAdd(acc[VW2*w+1], xlm[kl], agm[gid + (a_ld/VW2)*k].s1);
-          MultiplyAdd(acc[VW2*w+2], xlm[kl], agm[gid + (a_ld/VW2)*k].s2);
-          MultiplyAdd(acc[VW2*w+3], xlm[kl], agm[gid + (a_ld/VW2)*k].s3);
-          MultiplyAdd(acc[VW2*w+4], xlm[kl], agm[gid + (a_ld/VW2)*k].s4);
-          MultiplyAdd(acc[VW2*w+5], xlm[kl], agm[gid + (a_ld/VW2)*k].s5);
-          MultiplyAdd(acc[VW2*w+6], xlm[kl], agm[gid + (a_ld/VW2)*k].s6);
-          MultiplyAdd(acc[VW2*w+7], xlm[kl], agm[gid + (a_ld/VW2)*k].s7);
-          MultiplyAdd(acc[VW2*w+8], xlm[kl], agm[gid + (a_ld/VW2)*k].s8);
-          MultiplyAdd(acc[VW2*w+9], xlm[kl], agm[gid + (a_ld/VW2)*k].s9);
-          MultiplyAdd(acc[VW2*w+10], xlm[kl], agm[gid + (a_ld/VW2)*k].sA);
-          MultiplyAdd(acc[VW2*w+11], xlm[kl], agm[gid + (a_ld/VW2)*k].sB);
-          MultiplyAdd(acc[VW2*w+12], xlm[kl], agm[gid + (a_ld/VW2)*k].sC);
-          MultiplyAdd(acc[VW2*w+13], xlm[kl], agm[gid + (a_ld/VW2)*k].sD);
-          MultiplyAdd(acc[VW2*w+14], xlm[kl], agm[gid + (a_ld/VW2)*k].sE);
-          MultiplyAdd(acc[VW2*w+15], xlm[kl], agm[gid + (a_ld/VW2)*k].sF);
+          MultiplyAdd(acc[VW2*w+0], xlm[kl], avec.s0);
+          MultiplyAdd(acc[VW2*w+1], xlm[kl], avec.s1);
+          MultiplyAdd(acc[VW2*w+2], xlm[kl], avec.s2);
+          MultiplyAdd(acc[VW2*w+3], xlm[kl], avec.s3);
+          MultiplyAdd(acc[VW2*w+4], xlm[kl], avec.s4);
+          MultiplyAdd(acc[VW2*w+5], xlm[kl], avec.s5);
+          MultiplyAdd(acc[VW2*w+6], xlm[kl], avec.s6);
+          MultiplyAdd(acc[VW2*w+7], xlm[kl], avec.s7);
+          MultiplyAdd(acc[VW2*w+8], xlm[kl], avec.s8);
+          MultiplyAdd(acc[VW2*w+9], xlm[kl], avec.s9);
+          MultiplyAdd(acc[VW2*w+10], xlm[kl], avec.sA);
+          MultiplyAdd(acc[VW2*w+11], xlm[kl], avec.sB);
+          MultiplyAdd(acc[VW2*w+12], xlm[kl], avec.sC);
+          MultiplyAdd(acc[VW2*w+13], xlm[kl], avec.sD);
+          MultiplyAdd(acc[VW2*w+14], xlm[kl], avec.sE);
+          MultiplyAdd(acc[VW2*w+15], xlm[kl], avec.sF);
         #endif
       }
     }
@@ -258,19 +293,6 @@ __kernel void XgemvFast(const int m, const int n, const real alpha, const real b
 
 // =================================================================================================
 
-// Data-widths for the 'fast' kernel with rotated matrix
-#if VW3 == 1
-  typedef real realVFR;
-#elif VW3 == 2
-  typedef real2 realVFR;
-#elif VW3 == 4
-  typedef real4 realVFR;
-#elif VW3 == 8
-  typedef real8 realVFR;
-#elif VW3 == 16
-  typedef real16 realVFR;
-#endif
-
 // Faster version of the kernel, assuming that:
 // --> 'm' and 'n' are multiples of WGS3
 // --> 'a_offset' is 0
@@ -311,7 +333,7 @@ __kernel void XgemvFastRot(const int m, const int n, const real alpha, const rea
       #pragma unroll
       for (int w=0; w<WPT3; ++w) {
         const int gid = WPT3*get_global_id(0) + w;
-        realVFR avec = agm[k + (a_ld/VW3)*gid];
+        realVFR avec = LoadMatrixAVFR(agm, k, gid, a_ld/VW3);
         #if VW3 == 1
           MultiplyAdd(acc[w], xlm[VW3*kl+0], avec);
         #elif VW3 == 2
diff --git a/src/routines/level2/xgemv.cc b/src/routines/level2/xgemv.cc
index 75219b63..f95a9957 100644
--- a/src/routines/level2/xgemv.cc
+++ b/src/routines/level2/xgemv.cc
@@ -29,9 +29,10 @@ template <> const Precision Xgemv<double2>::precision_ = Precision::kComplexDoub
 
 // Constructor: forwards to base class constructor
 template <typename T>
-Xgemv<T>::Xgemv(Queue &queue, Event &event):
-    Routine<T>(queue, event, "GEMV", {"Xgemv"}, precision_) {
+Xgemv<T>::Xgemv(Queue &queue, Event &event, const std::string &name):
+    Routine<T>(queue, event, name, {"Pad", "Xgemv"}, precision_) {
   source_string_ =
+    #include "../../kernels/pad.opencl" // For {Herm,Symm}{Upper,Lower}ToSquared (for HEMV/SYMV)
     #include "../../kernels/xgemv.opencl"
   ;
 }
diff --git a/src/routines/level2/xhemv.cc b/src/routines/level2/xhemv.cc
new file mode 100644
index 00000000..2d92e45f
--- /dev/null
+++ b/src/routines/level2/xhemv.cc
@@ -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 implements the Xhemv class (see the header for information about the class).
+//
+// =================================================================================================
+
+#include "internal/routines/level2/xhemv.h"
+
+#include <string>
+#include <vector>
+
+namespace clblast {
+// =================================================================================================
+
+// Constructor: forwards to base class constructor
+template <typename T>
+Xhemv<T>::Xhemv(Queue &queue, Event &event, const std::string &name):
+    Xgemv<T>(queue, event, name) {
+}
+
+// =================================================================================================
+
+// The main routine
+template <typename T>
+StatusCode Xhemv<T>::DoHemv(const Layout layout, const Triangle triangle,
+                            const size_t n,
+                            const T alpha,
+                            const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
+                            const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
+                            const T beta,
+                            const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc) {
+
+  // Makes sure all dimensions are larger than zero
+  if (n == 0) { return StatusCode::kInvalidDimension; }
+
+  // Checks for validity of the squared A matrix
+  auto status = TestMatrixA(n, n, a_buffer, a_offset, a_ld, sizeof(T));
+  if (ErrorIn(status)) { return status; }
+
+  // Determines which kernel to run based on the layout (the Xgemv kernel assumes column-major as
+  // default) and on whether we are dealing with an upper or lower triangle of the hermitian matrix
+  bool is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) ||
+                   (triangle == Triangle::kLower && layout == Layout::kRowMajor));
+  auto kernel_name = (is_upper) ? "HermUpperToSquared" : "HermLowerToSquared";
+
+  // Temporary buffer for a copy of the hermitian matrix
+  try {
+    auto temp_herm = Buffer<T>(context_, n*n);
+
+    // Creates a general matrix from the hermitian matrix to be able to run the regular Xgemv
+    // routine afterwards
+    try {
+      auto& program = GetProgramFromCache();
+      auto kernel = Kernel(program, kernel_name);
+
+      // Sets the arguments for the hermitian-to-squared kernel
+      kernel.SetArgument(0, static_cast<int>(n));
+      kernel.SetArgument(1, static_cast<int>(a_ld));
+      kernel.SetArgument(2, static_cast<int>(a_offset));
+      kernel.SetArgument(3, a_buffer());
+      kernel.SetArgument(4, static_cast<int>(n));
+      kernel.SetArgument(5, static_cast<int>(n));
+      kernel.SetArgument(6, static_cast<int>(0));
+      kernel.SetArgument(7, temp_herm());
+
+      // Uses the common padding kernel's thread configuration. This is allowed, since the
+      // hermitian-to-squared kernel uses the same parameters.
+      auto global = std::vector<size_t>{Ceil(CeilDiv(n, db_["PAD_WPTX"]), db_["PAD_DIMX"]),
+                                        Ceil(CeilDiv(n, db_["PAD_WPTY"]), db_["PAD_DIMY"])};
+      auto local = std::vector<size_t>{db_["PAD_DIMX"], db_["PAD_DIMY"]};
+      status = RunKernel(kernel, global, local);
+      if (ErrorIn(status)) { return status; }
+
+      // Runs the regular Xgemv code
+      status = DoGemv(layout, Transpose::kNo, n, n, alpha,
+                      temp_herm, 0, n,
+                      x_buffer, x_offset, x_inc, beta,
+                      y_buffer, y_offset, y_inc);
+
+      // Return the status of the Xgemv routine
+      return status;
+    } catch (...) { return StatusCode::kInvalidKernel; }
+  } catch (...) { return StatusCode::kTempBufferAllocFailure; }
+}
+
+// =================================================================================================
+
+// Compiles the templated class
+template class Xhemv<float2>;
+template class Xhemv<double2>;
+
+// =================================================================================================
+} // namespace clblast
diff --git a/src/routines/level2/xsymv.cc b/src/routines/level2/xsymv.cc
new file mode 100644
index 00000000..2ccb51f6
--- /dev/null
+++ b/src/routines/level2/xsymv.cc
@@ -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 implements the Xsymv class (see the header for information about the class).
+//
+// =================================================================================================
+
+#include "internal/routines/level2/xsymv.h"
+
+#include <string>
+#include <vector>
+
+namespace clblast {
+// =================================================================================================
+
+// Constructor: forwards to base class constructor
+template <typename T>
+Xsymv<T>::Xsymv(Queue &queue, Event &event, const std::string &name):
+    Xgemv<T>(queue, event, name) {
+}
+
+// =================================================================================================
+
+// The main routine
+template <typename T>
+StatusCode Xsymv<T>::DoSymv(const Layout layout, const Triangle triangle,
+                            const size_t n,
+                            const T alpha,
+                            const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
+                            const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
+                            const T beta,
+                            const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc) {
+
+  // Makes sure all dimensions are larger than zero
+  if (n == 0) { return StatusCode::kInvalidDimension; }
+
+  // Checks for validity of the squared A matrix
+  auto status = TestMatrixA(n, n, a_buffer, a_offset, a_ld, sizeof(T));
+  if (ErrorIn(status)) { return status; }
+
+  // Determines which kernel to run based on the layout (the Xgemv kernel assumes column-major as
+  // default) and on whether we are dealing with an upper or lower triangle of the symmetric matrix
+  bool is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) ||
+                   (triangle == Triangle::kLower && layout == Layout::kRowMajor));
+  auto kernel_name = (is_upper) ? "SymmUpperToSquared" : "SymmLowerToSquared";
+
+  // Temporary buffer for a copy of the symmetric matrix
+  try {
+    auto temp_symm = Buffer<T>(context_, n*n);
+
+    // Creates a general matrix from the symmetric matrix to be able to run the regular Xgemv
+    // routine afterwards
+    try {
+      auto& program = GetProgramFromCache();
+      auto kernel = Kernel(program, kernel_name);
+
+      // Sets the arguments for the symmetric-to-squared kernel
+      kernel.SetArgument(0, static_cast<int>(n));
+      kernel.SetArgument(1, static_cast<int>(a_ld));
+      kernel.SetArgument(2, static_cast<int>(a_offset));
+      kernel.SetArgument(3, a_buffer());
+      kernel.SetArgument(4, static_cast<int>(n));
+      kernel.SetArgument(5, static_cast<int>(n));
+      kernel.SetArgument(6, static_cast<int>(0));
+      kernel.SetArgument(7, temp_symm());
+
+      // Uses the common padding kernel's thread configuration. This is allowed, since the
+      // symmetric-to-squared kernel uses the same parameters.
+      auto global = std::vector<size_t>{Ceil(CeilDiv(n, db_["PAD_WPTX"]), db_["PAD_DIMX"]),
+                                        Ceil(CeilDiv(n, db_["PAD_WPTY"]), db_["PAD_DIMY"])};
+      auto local = std::vector<size_t>{db_["PAD_DIMX"], db_["PAD_DIMY"]};
+      status = RunKernel(kernel, global, local);
+      if (ErrorIn(status)) { return status; }
+
+      // Runs the regular Xgemv code
+      status = DoGemv(layout, Transpose::kNo, n, n, alpha,
+                      temp_symm, 0, n,
+                      x_buffer, x_offset, x_inc, beta,
+                      y_buffer, y_offset, y_inc);
+
+      // Return the status of the Xgemv routine
+      return status;
+    } catch (...) { return StatusCode::kInvalidKernel; }
+  } catch (...) { return StatusCode::kTempBufferAllocFailure; }
+}
+
+// =================================================================================================
+
+// Compiles the templated class
+template class Xsymv<float>;
+template class Xsymv<double>;
+
+// =================================================================================================
+} // namespace clblast