Added the HEMM routine, tester, and client

10 files changed, 640 insertions, 1 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 96e6573e..1b2c5657 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -98,7 +98,7 @@ set(SAMPLE_PROGRAMS sgemm)
 set(ROUTINES
   xaxpy
   xgemv
-  xgemm xsymm xsyrk xherk xsyr2k xher2k xtrmm)
+  xgemm xsymm xhemm xsyrk xherk xsyr2k xher2k xtrmm)
 
 # ==================================================================================================
 
diff --git a/include/clblast.h b/include/clblast.h
index e6c49dbf..80ea1707 100644
--- a/include/clblast.h
+++ b/include/clblast.h
@@ -130,6 +130,17 @@ StatusCode Symm(const Layout layout, const Side side, const Triangle triangle,
                 cl_mem c_buffer, const size_t c_offset, const size_t c_ld,
                 cl_command_queue* queue, cl_event* event);
 
+// Templated-precision hermitian matrix-matrix multiplication: CHEMM/ZHEMM
+template <typename T>
+StatusCode Hemm(const Layout layout, const Side side, const Triangle triangle,
+                const size_t m, 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 b_buffer, const size_t b_offset, const size_t b_ld,
+                const T beta,
+                cl_mem c_buffer, const size_t c_offset, const size_t c_ld,
+                cl_command_queue* queue, cl_event* event);
+
 // Templated-precision rank-K update of a symmetric matrix: SSYRK/DSYRK/CSYRK/ZSYRK
 template <typename T>
 StatusCode Syrk(const Layout layout, const Triangle triangle, const Transpose a_transpose,
diff --git a/include/internal/routines/xhemm.h b/include/internal/routines/xhemm.h
new file mode 100644
index 00000000..1b1a0dfa
--- /dev/null
+++ b/include/internal/routines/xhemm.h
@@ -0,0 +1,58 @@
+
+// =================================================================================================
+// 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 Xhemm routine. It is based on the generalized matrix multiplication
+// routine (Xgemm). The implementation is very similar to the Xsymm routine.
+//
+// =================================================================================================
+
+#ifndef CLBLAST_ROUTINES_XHEMM_H_
+#define CLBLAST_ROUTINES_XHEMM_H_
+
+#include "internal/routines/xgemm.h"
+
+namespace clblast {
+// =================================================================================================
+
+// See comment at top of file for a description of the class
+template <typename T>
+class Xhemm: public Xgemm<T> {
+ public:
+
+  // Uses several variables from the Routine class
+  using Routine::db_;
+  using Routine::context_;
+
+  // Uses several helper functions from the Routine class
+  using Routine::RunKernel;
+  using Routine::ErrorIn;
+  using Routine::TestMatrixA;
+  using Routine::GetProgramFromCache;
+
+  // Uses the regular Xgemm routine
+  using Xgemm<T>::DoGemm;
+
+  // Constructor
+  Xhemm(CommandQueue &queue, Event &event);
+
+  // Templated-precision implementation of the routine
+  StatusCode DoHemm(const Layout layout, const Side side, const Triangle triangle,
+                    const size_t m, const size_t n,
+                    const T alpha,
+                    const Buffer &a_buffer, const size_t a_offset, const size_t a_ld,
+                    const Buffer &b_buffer, const size_t b_offset, const size_t b_ld,
+                    const T beta,
+                    const Buffer &c_buffer, const size_t c_offset, const size_t c_ld);
+};
+
+// =================================================================================================
+} // namespace clblast
+
+// CLBLAST_ROUTINES_XHEMM_H_
+#endif
diff --git a/src/clblast.cc b/src/clblast.cc
index 66202adb..23046b01 100644
--- a/src/clblast.cc
+++ b/src/clblast.cc
@@ -26,6 +26,7 @@
 // BLAS level-3 includes
 #include "internal/routines/xgemm.h"
 #include "internal/routines/xsymm.h"
+#include "internal/routines/xhemm.h"
 #include "internal/routines/xsyrk.h"
 #include "internal/routines/xherk.h"
 #include "internal/routines/xsyr2k.h"
@@ -250,6 +251,54 @@ template StatusCode Symm<double2>(const Layout, const Side, const Triangle,
 
 // =================================================================================================
 
+// HEMM
+template <typename T>
+StatusCode Hemm(const Layout layout, const Side side, const Triangle triangle,
+                const size_t m, 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 b_buffer, const size_t b_offset, const size_t b_ld, const T beta,
+                cl_mem c_buffer, const size_t c_offset, const size_t c_ld,
+                cl_command_queue* queue, cl_event* event) {
+  auto queue_cpp = CommandQueue(*queue);
+  auto event_cpp = Event(*event);
+  auto routine = Xhemm<T>(queue_cpp, event_cpp);
+
+  // Loads the kernel source-code as an include (C++11 raw string literal)
+  std::string common_source1 =
+  #include "kernels/copy.opencl"
+  std::string common_source2 =
+  #include "kernels/pad.opencl"
+  std::string common_source3 =
+  #include "kernels/transpose.opencl"
+  std::string common_source4 =
+  #include "kernels/padtranspose.opencl"
+  std::string kernel_source =
+  #include "kernels/xgemm.opencl"
+  auto status = routine.SetUp(common_source1 + common_source2 + common_source3 + common_source4 +
+                              kernel_source);
+  if (status != StatusCode::kSuccess) { return status; }
+
+  // Runs the routine
+  return routine.DoHemm(layout, side, triangle, m, n, alpha,
+                        Buffer(a_buffer), a_offset, a_ld,
+                        Buffer(b_buffer), b_offset, b_ld, beta,
+                        Buffer(c_buffer), c_offset, c_ld);
+}
+template StatusCode Hemm<float2>(const Layout, const Side, const Triangle,
+                                 const size_t, 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 Hemm<double2>(const Layout, const Side, const Triangle,
+                                  const size_t, 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*);
+
+// =================================================================================================
+
 // SYRK
 template <typename T>
 StatusCode Syrk(const Layout layout, const Triangle triangle, const Transpose a_transpose,
diff --git a/src/kernels/pad.opencl b/src/kernels/pad.opencl
index f8a89d24..2791db30 100644
--- a/src/kernels/pad.opencl
+++ b/src/kernels/pad.opencl
@@ -186,6 +186,89 @@ __kernel void SymmUpperToSquared(const int src_dim,
 }
 
 // =================================================================================================
+#if PRECISION == 3232 || PRECISION == 6464
+
+// Kernel to populate a squared hermitian matrix, given that the triangle which holds the data is
+// stored as the lower-triangle of the input matrix. This uses the padding kernel's parameters.
+__attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
+__kernel void HermLowerToSquared(const int src_dim,
+                                 const int src_ld, const int src_offset,
+                                 __global const real* restrict src,
+                                 const int dest_dim,
+                                 const int dest_ld, const int dest_offset,
+                                 __global real* dest) {
+
+  // Loops over the work per thread in both dimensions
+  #pragma unroll
+  for (int w_one=0; w_one<PAD_WPTX; ++w_one) {
+    const int id_one = (get_group_id(0)*PAD_WPTX + w_one) * PAD_DIMX + get_local_id(0);
+    #pragma unroll
+    for (int w_two=0; w_two<PAD_WPTY; ++w_two) {
+      const int id_two = (get_group_id(1)*PAD_WPTY + w_two) * PAD_DIMY + get_local_id(1);
+      if (id_two < dest_dim && id_one < dest_dim) {
+
+        // Loads data from the lower-hermitian matrix
+        real result;
+        SetToZero(result);
+        if (id_two < src_dim && id_one < src_dim) {
+          if (id_two <= id_one) {
+            result = src[id_two*src_ld + id_one + src_offset];
+            if (id_one == id_two) { result.y = ZERO; }
+          }
+          else {
+            result = src[id_one*src_ld + id_two + src_offset];
+            COMPLEX_CONJUGATE(result);
+          }
+        }
+
+        // Stores the result in the destination matrix
+        dest[id_two*dest_ld + id_one + dest_offset] = result;
+      }
+    }
+  }
+}
+
+// Same as above, but now the matrix' data is stored in the upper-triangle
+__attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
+__kernel void HermUpperToSquared(const int src_dim,
+                                 const int src_ld, const int src_offset,
+                                 __global const real* restrict src,
+                                 const int dest_dim,
+                                 const int dest_ld, const int dest_offset,
+                                 __global real* dest) {
+
+  // Loops over the work per thread in both dimensions
+  #pragma unroll
+  for (int w_one=0; w_one<PAD_WPTX; ++w_one) {
+    const int id_one = (get_group_id(0)*PAD_WPTX + w_one) * PAD_DIMX + get_local_id(0);
+    #pragma unroll
+    for (int w_two=0; w_two<PAD_WPTY; ++w_two) {
+      const int id_two = (get_group_id(1)*PAD_WPTY + w_two) * PAD_DIMY + get_local_id(1);
+      if (id_two < dest_dim && id_one < dest_dim) {
+
+        // Loads data from the upper-hermitian matrix
+        real result;
+        SetToZero(result);
+        if (id_two < src_dim && id_one < src_dim) {
+          if (id_one <= id_two) {
+            result = src[id_two*src_ld + id_one + src_offset];
+            if (id_one == id_two) { result.y = ZERO; }
+          }
+          else {
+            result = src[id_one*src_ld + id_two + src_offset];
+            COMPLEX_CONJUGATE(result);
+          }
+        }
+
+        // Stores the result in the destination matrix
+        dest[id_two*dest_ld + id_one + dest_offset] = result;
+      }
+    }
+  }
+}
+
+#endif
+// =================================================================================================
 
 // Kernel to populate a squared triangular matrix, given that the triangle which holds the data is
 // stored as the lower-triangle of the input matrix. This uses the padding kernel's parameters.
diff --git a/src/routines/xhemm.cc b/src/routines/xhemm.cc
new file mode 100644
index 00000000..73f769ed
--- /dev/null
+++ b/src/routines/xhemm.cc
@@ -0,0 +1,130 @@
+
+// =================================================================================================
+// 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 Xhemm class (see the header for information about the class).
+//
+// =================================================================================================
+
+#include "internal/routines/xhemm.h"
+
+#include <string>
+#include <vector>
+
+namespace clblast {
+// =================================================================================================
+
+// Constructor: forwards to base class constructor
+template <typename T>
+Xhemm<T>::Xhemm(CommandQueue &queue, Event &event):
+    Xgemm<T>(queue, event) {
+}
+
+// =================================================================================================
+
+// The main routine
+template <typename T>
+StatusCode Xhemm<T>::DoHemm(const Layout layout, const Side side, const Triangle triangle,
+                            const size_t m, const size_t n,
+                            const T alpha,
+                            const Buffer &a_buffer, const size_t a_offset, const size_t a_ld,
+                            const Buffer &b_buffer, const size_t b_offset, const size_t b_ld,
+                            const T beta,
+                            const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) {
+
+  // Makes sure all dimensions are larger than zero
+  if ((m == 0) || (n == 0) ) { return StatusCode::kInvalidDimension; }
+
+  // Computes the k dimension. This is based on whether or not the hermitian matrix is A (on the
+  // left) or B (on the right) in the Xgemm routine.
+  auto k = (side == Side::kLeft) ? m : n;
+
+  // Checks for validity of the squared A matrix
+  auto status = TestMatrixA(k, k, a_buffer, a_offset, a_ld, sizeof(T));
+  if (ErrorIn(status)) { return status; }
+
+  // Determines which kernel to run based on the layout (the Xgemm 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(context_, CL_MEM_READ_WRITE, k*k*sizeof(T));
+
+    // Creates a general matrix from the hermitian matrix to be able to run the regular Xgemm
+    // 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>(k));
+      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>(k));
+      kernel.SetArgument(5, static_cast<int>(k));
+      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(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]),
+                                        Ceil(CeilDiv(k, 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 Xgemm code with either "C := AB+C" or ...
+      if (side == Side::kLeft) {
+        status = DoGemm(layout, Transpose::kNo, Transpose::kNo,
+                        m, n, k,
+                        alpha,
+                        temp_herm, 0, k,
+                        b_buffer, b_offset, b_ld,
+                        beta,
+                        c_buffer, c_offset, c_ld);
+      }
+
+      // ... with "C := BA+C". Note that A and B are now reversed.
+      else {
+        status = DoGemm(layout, Transpose::kNo, Transpose::kNo,
+                        m, n, k,
+                        alpha,
+                        b_buffer, b_offset, b_ld,
+                        temp_herm, 0, k,
+                        beta,
+                        c_buffer, c_offset, c_ld);
+
+        // A and B are now reversed, so also reverse the error codes returned from the Xgemm routine
+        switch(status) {
+          case StatusCode::kInvalidMatrixA:      status = StatusCode::kInvalidMatrixB; break;
+          case StatusCode::kInvalidMatrixB:      status = StatusCode::kInvalidMatrixA; break;
+          case StatusCode::kInvalidLeadDimA:     status = StatusCode::kInvalidLeadDimB; break;
+          case StatusCode::kInvalidLeadDimB:     status = StatusCode::kInvalidLeadDimA; break;
+          case StatusCode::kInsufficientMemoryA: status = StatusCode::kInsufficientMemoryB; break;
+          case StatusCode::kInsufficientMemoryB: status = StatusCode::kInsufficientMemoryA; break;
+        }
+      }
+
+      // Return the status of the Xgemm routine
+      return status;
+    } catch (...) { return StatusCode::kInvalidKernel; }
+  } catch (...) { return StatusCode::kTempBufferAllocFailure; }
+}
+
+// =================================================================================================
+
+// Compiles the templated class
+template class Xhemm<float2>;
+template class Xhemm<double2>;
+
+// =================================================================================================
+} // namespace clblast
diff --git a/test/correctness/routines/xhemm.cc b/test/correctness/routines/xhemm.cc
new file mode 100644
index 00000000..e8c82f65
--- /dev/null
+++ b/test/correctness/routines/xhemm.cc
@@ -0,0 +1,98 @@
+
+// =================================================================================================
+// 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 tests for the Xhemm routine.
+//
+// =================================================================================================
+
+#include "correctness/testblas.h"
+#include "routines/xhemm.h"
+
+namespace clblast {
+// =================================================================================================
+
+// The correctness tester
+template <typename T>
+void RunTest(int argc, char *argv[], const bool silent, const std::string &name) {
+
+  // Creates a tester
+  TestBlas<T,T> tester{argc, argv, silent, name, TestXhemm<T>::GetOptions(),
+                       TestXhemm<T>::RunRoutine, TestXhemm<T>::RunReference,
+                       TestXhemm<T>::DownloadResult, TestXhemm<T>::GetResultIndex,
+                       TestXhemm<T>::ResultID1, TestXhemm<T>::ResultID2};
+
+  // This variable holds the arguments relevant for this routine
+  auto args = Arguments<T>{};
+
+  // Loops over the test-cases from a data-layout point of view
+  for (auto &layout: tester.kLayouts) { args.layout = layout;
+    for (auto &side: tester.kSides) { args.side = side;
+      for (auto &triangle: tester.kTriangles) { args.triangle = triangle;
+
+        // Creates the arguments vector for the regular tests
+        auto regular_test_vector = std::vector<Arguments<T>>{};
+        for (auto &m: tester.kMatrixDims) { args.m = m;
+          for (auto &n: tester.kMatrixDims) { args.n = n;
+            for (auto &a_ld: tester.kMatrixDims) { args.a_ld = a_ld;
+              for (auto &a_offset: tester.kOffsets) { args.a_offset = a_offset;
+                for (auto &b_ld: tester.kMatrixDims) { args.b_ld = b_ld;
+                  for (auto &b_offset: tester.kOffsets) { args.b_offset = b_offset;
+                    for (auto &c_ld: tester.kMatrixDims) { args.c_ld = c_ld;
+                      for (auto &c_offset: tester.kOffsets) { args.c_offset = c_offset;
+                        for (auto &alpha: tester.kAlphaValues) { args.alpha = alpha;
+                          for (auto &beta: tester.kBetaValues) { args.beta = beta;
+                            args.a_size = TestXhemm<T>::GetSizeA(args);
+                            args.b_size = TestXhemm<T>::GetSizeB(args);
+                            args.c_size = TestXhemm<T>::GetSizeC(args);
+                            if (args.a_size<1 || args.b_size<1 || args.c_size<1) { continue; }
+                            regular_test_vector.push_back(args);
+                          }
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+
+        // Creates the arguments vector for the invalid-buffer tests
+        auto invalid_test_vector = std::vector<Arguments<T>>{};
+        args.m = args.n = tester.kBufferSize;
+        args.a_ld = args.b_ld = args.c_ld = tester.kBufferSize;
+        args.a_offset = args.b_offset = args.c_offset = 0;
+        for (auto &a_size: tester.kMatSizes) { args.a_size = a_size;
+          for (auto &b_size: tester.kMatSizes) { args.b_size = b_size;
+            for (auto &c_size: tester.kMatSizes) { args.c_size = c_size;
+              invalid_test_vector.push_back(args);
+            }
+          }
+        }
+
+        // Runs the tests
+        const auto case_name = ToString(layout)+" "+ToString(side)+" "+ToString(triangle);
+        tester.TestRegular(regular_test_vector, case_name);
+        tester.TestInvalid(invalid_test_vector, case_name);
+      }
+    }
+  }
+}
+
+// =================================================================================================
+} // namespace clblast
+
+// Main function (not within the clblast namespace)
+int main(int argc, char *argv[]) {
+  clblast::RunTest<clblast::float2>(argc, argv, true, "CHEMM");
+  clblast::RunTest<clblast::double2>(argc, argv, true, "ZHEMM");
+  return 0;
+}
+
+// =================================================================================================
diff --git a/test/performance/routines/xhemm.cc b/test/performance/routines/xhemm.cc
new file mode 100644
index 00000000..34798d8d
--- /dev/null
+++ b/test/performance/routines/xhemm.cc
@@ -0,0 +1,40 @@
+
+// =================================================================================================
+// 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 Xhemm command-line interface performance tester.
+//
+// =================================================================================================
+
+#include "performance/client.h"
+#include "routines/xhemm.h"
+
+// =================================================================================================
+
+// Shortcuts to the clblast namespace
+using float2 = clblast::float2;
+using double2 = clblast::double2;
+
+// Main function (not within the clblast namespace)
+int main(int argc, char *argv[]) {
+  switch(clblast::GetPrecision(argc, argv)) {
+    case clblast::Precision::kHalf:
+      throw std::runtime_error("Unsupported precision mode");
+    case clblast::Precision::kSingle:
+      throw std::runtime_error("Unsupported precision mode");
+    case clblast::Precision::kDouble:
+      throw std::runtime_error("Unsupported precision mode");
+    case clblast::Precision::kComplexSingle:
+      clblast::RunClient<clblast::TestXhemm<float2>, float2, float2>(argc, argv); break;
+    case clblast::Precision::kComplexDouble:
+      clblast::RunClient<clblast::TestXhemm<double2>, double2, double2>(argc, argv); break;
+  }
+  return 0;
+}
+
+// =================================================================================================
diff --git a/test/routines/xhemm.h b/test/routines/xhemm.h
new file mode 100644
index 00000000..75878b06
--- /dev/null
+++ b/test/routines/xhemm.h
@@ -0,0 +1,134 @@
+
+// =================================================================================================
+// 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 a class with static methods to describe the Xhemm routine. Examples of
+// such 'descriptions' are how to calculate the size a of buffer or how to run the routine. These
+// static methods are used by the correctness tester and the performance tester.
+//
+// =================================================================================================
+
+#ifndef CLBLAST_TEST_ROUTINES_XHEMM_H_
+#define CLBLAST_TEST_ROUTINES_XHEMM_H_
+
+#include <vector>
+#include <string>
+
+#include "wrapper_clblas.h"
+
+namespace clblast {
+// =================================================================================================
+
+// See comment at top of file for a description of the class
+template <typename T>
+class TestXhemm {
+ public:
+
+  // The list of arguments relevant for this routine
+  static std::vector<std::string> GetOptions() {
+    return {kArgM, kArgN,
+            kArgLayout, kArgSide, kArgTriangle,
+            kArgALeadDim, kArgBLeadDim, kArgCLeadDim,
+            kArgAOffset, kArgBOffset, kArgCOffset,
+            kArgAlpha, kArgBeta};
+  }
+
+  // Describes how to obtain the sizes of the buffers
+  static size_t GetSizeA(const Arguments<T> &args) {
+    size_t k_value = (args.side == Side::kLeft) ? args.m : args.n;
+    auto a_rotated = (args.layout == Layout::kRowMajor);
+    auto a_two = (a_rotated) ? args.m : k_value;
+    return a_two * args.a_ld + args.a_offset;
+  }
+  static size_t GetSizeB(const Arguments<T> &args) {
+    size_t k_value = (args.side == Side::kLeft) ? args.m : args.n;
+    auto b_rotated = (args.layout == Layout::kRowMajor);
+    auto b_two = (b_rotated) ? k_value : args.n;
+    return b_two * args.b_ld + args.b_offset;
+  }
+  static size_t GetSizeC(const Arguments<T> &args) {
+    auto c_rotated = (args.layout == Layout::kRowMajor);
+    auto c_two = (c_rotated) ? args.m : args.n;
+    return c_two * args.c_ld + args.c_offset;
+  }
+
+  // Describes how to set the sizes of all the buffers
+  static void SetSizes(Arguments<T> &args) {
+    args.a_size = GetSizeA(args);
+    args.b_size = GetSizeB(args);
+    args.c_size = GetSizeC(args);
+  }
+
+  // Describes what the default values of the leading dimensions of the matrices are
+  static size_t DefaultLDA(const Arguments<T> &args) { return args.m; }
+  static size_t DefaultLDB(const Arguments<T> &args) { return args.n; }
+  static size_t DefaultLDC(const Arguments<T> &args) { return args.n; }
+
+  // Describes how to run the CLBlast routine
+  static StatusCode RunRoutine(const Arguments<T> &args, const Buffers &buffers,
+                               CommandQueue &queue) {
+    auto queue_plain = queue();
+    auto event = cl_event{};
+    auto status = Hemm(args.layout, args.side, args.triangle,
+                       args.m, args.n, args.alpha,
+                       buffers.a_mat(), args.a_offset, args.a_ld,
+                       buffers.b_mat(), args.b_offset, args.b_ld, args.beta,
+                       buffers.c_mat(), args.c_offset, args.c_ld,
+                       &queue_plain, &event);
+    clWaitForEvents(1, &event);
+    return status;
+  }
+
+  // Describes how to run the clBLAS routine (for correctness/performance comparison)
+  static StatusCode RunReference(const Arguments<T> &args, const Buffers &buffers,
+                                 CommandQueue &queue) {
+    auto queue_plain = queue();
+    auto event = cl_event{};
+    auto status = clblasXhemm(static_cast<clblasOrder>(args.layout),
+                              static_cast<clblasSide>(args.side),
+                              static_cast<clblasUplo>(args.triangle),
+                              args.m, args.n, args.alpha,
+                              buffers.a_mat(), args.a_offset, args.a_ld,
+                              buffers.b_mat(), args.b_offset, args.b_ld, args.beta,
+                              buffers.c_mat(), args.c_offset, args.c_ld,
+                              1, &queue_plain, 0, nullptr, &event);
+    clWaitForEvents(1, &event);
+    return static_cast<StatusCode>(status);
+  }
+
+  // Describes how to download the results of the computation (more importantly: which buffer)
+  static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers &buffers,
+                                       CommandQueue &queue) {
+    std::vector<T> result(args.c_size, static_cast<T>(0));
+    buffers.c_mat.ReadBuffer(queue, args.c_size*sizeof(T), result);
+    return result;
+  }
+
+  // Describes how to compute the indices of the result buffer
+  static size_t ResultID1(const Arguments<T> &args) { return args.m; }
+  static size_t ResultID2(const Arguments<T> &args) { return args.n; }
+  static size_t GetResultIndex(const Arguments<T> &args, const size_t id1, const size_t id2) {
+    return (args.layout == Layout::kRowMajor) ?
+           id1*args.c_ld + id2 + args.c_offset:
+           id2*args.c_ld + id1 + args.c_offset;
+  }
+
+  // Describes how to compute performance metrics
+  static size_t GetFlops(const Arguments<T> &args) {
+    return 2 * args.m * args.n * args.m;
+  }
+  static size_t GetBytes(const Arguments<T> &args) {
+    return (args.m*args.m + args.m*args.n + 2*args.m*args.n) * sizeof(T);
+  }
+};
+
+// =================================================================================================
+} // namespace clblast
+
+// CLBLAST_TEST_ROUTINES_XHEMM_H_
+#endif
diff --git a/test/wrapper_clblas.h b/test/wrapper_clblas.h
index 03f31a58..4aaf3705 100644
--- a/test/wrapper_clblas.h
+++ b/test/wrapper_clblas.h
@@ -267,6 +267,42 @@ clblasStatus clblasXsymm(
                        num_queues, queues, num_wait_events, wait_events, events);
 }
 
+// This calls {clblasChemm, clblasZhemm} with the arguments forwarded.
+clblasStatus clblasXhemm(
+  clblasOrder layout, clblasSide side, clblasUplo triangle,
+  size_t m, size_t n, float2 alpha,
+  const cl_mem a_mat, size_t a_offset, size_t a_ld,
+  const cl_mem b_mat, size_t b_offset, size_t b_ld, float2 beta,
+  const cl_mem c_mat, size_t c_offset, size_t c_ld,
+  cl_uint num_queues, cl_command_queue *queues,
+  cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) {
+    auto cl_alpha = cl_float2{{alpha.real(), alpha.imag()}};
+    auto cl_beta = cl_float2{{beta.real(), beta.imag()}};
+    return clblasChemm(layout, side, triangle,
+                       m, n, cl_alpha,
+                       a_mat, a_offset, a_ld,
+                       b_mat, b_offset, b_ld, cl_beta,
+                       c_mat, c_offset, c_ld,
+                       num_queues, queues, num_wait_events, wait_events, events);
+}
+clblasStatus clblasXhemm(
+  clblasOrder layout, clblasSide side, clblasUplo triangle,
+  size_t m, size_t n, double2 alpha,
+  const cl_mem a_mat, size_t a_offset, size_t a_ld,
+  const cl_mem b_mat, size_t b_offset, size_t b_ld, double2 beta,
+  const cl_mem c_mat, size_t c_offset, size_t c_ld,
+  cl_uint num_queues, cl_command_queue *queues,
+  cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) {
+    auto cl_alpha = cl_double2{{alpha.real(), alpha.imag()}};
+    auto cl_beta = cl_double2{{beta.real(), beta.imag()}};
+    return clblasZhemm(layout, side, triangle,
+                       m, n, cl_alpha,
+                       a_mat, a_offset, a_ld,
+                       b_mat, b_offset, b_ld, cl_beta,
+                       c_mat, c_offset, c_ld,
+                       num_queues, queues, num_wait_events, wait_events, events);
+}
+
 // This calls {clblasSsyrk, clblasDsyrk, clblasCsyrk, clblasZsyrk} with the arguments forwarded.
 clblasStatus clblasXsyrk(
   clblasOrder layout, clblasUplo triangle, clblasTranspose a_transpose,