Added the HEMM routine, tester, and client

4 files changed, 308 insertions, 0 deletions

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,