Added support for the SASUM/DASUM/ScASUM/DzASUM routines

3 files changed, 151 insertions, 5 deletions

diff --git a/test/correctness/testblas.h b/test/correctness/testblas.h
index 8181aaf6..aa61c2ec 100644
--- a/test/correctness/testblas.h
+++ b/test/correctness/testblas.h
@@ -153,6 +153,7 @@ void RunTests(int argc, char *argv[], const bool silent, const std::string &name
   auto ap_offsets = std::vector<size_t>{args.ap_offset};
   auto dot_offsets = std::vector<size_t>{args.dot_offset};
   auto nrm2_offsets = std::vector<size_t>{args.nrm2_offset};
+  auto asum_offsets = std::vector<size_t>{args.asum_offset};
   auto alphas = std::vector<U>{args.alpha};
   auto betas = std::vector<U>{args.beta};
   auto x_sizes = std::vector<size_t>{args.x_size};
@@ -193,6 +194,7 @@ void RunTests(int argc, char *argv[], const bool silent, const std::string &name
     if (option == kArgAPOffset) { ap_offsets = tester.kOffsets; }
     if (option == kArgDotOffset) { dot_offsets = tester.kOffsets; }
     if (option == kArgNrm2Offset) { nrm2_offsets = tester.kOffsets; }
+    if (option == kArgAsumOffset) { asum_offsets = tester.kOffsets; }
     if (option == kArgAlpha) { alphas = tester.kAlphaValues; }
     if (option == kArgBeta) { betas = tester.kBetaValues; }
 
@@ -233,10 +235,12 @@ void RunTests(int argc, char *argv[], const bool silent, const std::string &name
                                             for (auto &ap_offset: ap_offsets) { r_args.ap_offset = ap_offset;
                                               for (auto &dot_offset: dot_offsets) { r_args.dot_offset = dot_offset;
                                                 for (auto &nrm2_offset: nrm2_offsets) { r_args.nrm2_offset = nrm2_offset;
-                                                  for (auto &alpha: alphas) { r_args.alpha = alpha;
-                                                    for (auto &beta: betas) { r_args.beta = beta;
-                                                      C::SetSizes(r_args);
-                                                      regular_test_vector.push_back(r_args);
+                                                  for (auto &asum_offset: asum_offsets) { r_args.asum_offset = asum_offset;
+                                                    for (auto &alpha: alphas) { r_args.alpha = alpha;
+                                                      for (auto &beta: betas) { r_args.beta = beta;
+                                                        C::SetSizes(r_args);
+                                                        regular_test_vector.push_back(r_args);
+                                                      }
                                                     }
                                                   }
                                                 }
diff --git a/test/performance/client.cc b/test/performance/client.cc
index 56ab8c8d..f22c9666 100644
--- a/test/performance/client.cc
+++ b/test/performance/client.cc
@@ -80,8 +80,10 @@ Arguments<U> Client<T,U>::ParseArguments(int argc, char *argv[], const GetMetric
     if (o == kArgCOffset)  { args.c_offset = GetArgument(argc, argv, help, kArgCOffset, size_t{0}); }
     if (o == kArgAPOffset) { args.ap_offset= GetArgument(argc, argv, help, kArgAPOffset, size_t{0}); }
 
-    // Dot arguments
+    // Scalar result arguments
     if (o == kArgDotOffset)  { args.dot_offset = GetArgument(argc, argv, help, kArgDotOffset, size_t{0}); }
+    if (o == kArgNrm2Offset)  { args.nrm2_offset = GetArgument(argc, argv, help, kArgNrm2Offset, size_t{0}); }
+    if (o == kArgAsumOffset)  { args.asum_offset = GetArgument(argc, argv, help, kArgAsumOffset, size_t{0}); }
 
     // Scalar values 
     if (o == kArgAlpha) { args.alpha = GetArgument(argc, argv, help, kArgAlpha, GetScalar<U>()); }
@@ -292,6 +294,7 @@ void Client<T,U>::PrintTableRow(const Arguments<U>& args,
     else if (o == kArgAPOffset) { integers.push_back(args.ap_offset); }
     else if (o == kArgDotOffset) {integers.push_back(args.dot_offset); }
     else if (o == kArgNrm2Offset){integers.push_back(args.nrm2_offset); }
+    else if (o == kArgAsumOffset){integers.push_back(args.asum_offset); }
   }
   auto strings = std::vector<std::string>{};
   for (auto &o: options_) {
diff --git a/test/routines/level1/xasum.h b/test/routines/level1/xasum.h
new file mode 100644
index 00000000..6eae3c83
--- /dev/null
+++ b/test/routines/level1/xasum.h
@@ -0,0 +1,139 @@
+
+// =================================================================================================
+// 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 Xasum 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_XASUM_H_
+#define CLBLAST_TEST_ROUTINES_XASUM_H_
+
+#include <vector>
+#include <string>
+
+#ifdef CLBLAST_REF_CLBLAS
+  #include "wrapper_clblas.h"
+#endif
+#ifdef CLBLAST_REF_CBLAS
+  #include "wrapper_cblas.h"
+#endif
+
+namespace clblast {
+// =================================================================================================
+
+// See comment at top of file for a description of the class
+template <typename T>
+class TestXasum {
+ public:
+
+  // The BLAS level: 1, 2, or 3
+  static size_t BLASLevel() { return 1; }
+
+  // The list of arguments relevant for this routine
+  static std::vector<std::string> GetOptions() {
+    return {kArgN,
+            kArgXInc,
+            kArgXOffset, kArgAsumOffset};
+  }
+
+  // Describes how to obtain the sizes of the buffers
+  static size_t GetSizeX(const Arguments<T> &args) {
+    return args.n * args.x_inc + args.x_offset;
+  }
+  static size_t GetSizeAsum(const Arguments<T> &args) {
+    return 1 + args.asum_offset;
+  }
+
+  // Describes how to set the sizes of all the buffers
+  static void SetSizes(Arguments<T> &args) {
+    args.x_size = GetSizeX(args);
+    args.scalar_size = GetSizeAsum(args);
+  }
+
+  // Describes what the default values of the leading dimensions of the matrices are
+  static size_t DefaultLDA(const Arguments<T> &) { return 1; } // N/A for this routine
+  static size_t DefaultLDB(const Arguments<T> &) { return 1; } // N/A for this routine
+  static size_t DefaultLDC(const Arguments<T> &) { return 1; } // N/A for this routine
+
+  // Describes which transpose options are relevant for this routine
+  using Transposes = std::vector<Transpose>;
+  static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine
+  static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine
+
+  // Describes how to run the CLBlast routine
+  static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
+    auto queue_plain = queue();
+    auto event = cl_event{};
+    auto status = Asum<T>(args.n,
+                          buffers.scalar(), args.asum_offset,
+                          buffers.x_vec(), args.x_offset, args.x_inc,
+                          &queue_plain, &event);
+    clWaitForEvents(1, &event);
+    return status;
+  }
+
+  // Describes how to run the clBLAS routine (for correctness/performance comparison)
+  #ifdef CLBLAST_REF_CLBLAS
+    static StatusCode RunReference1(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
+      auto queue_plain = queue();
+      auto event = cl_event{};
+      auto status = clblasXasum<T>(args.n,
+                                   buffers.scalar(), args.asum_offset,
+                                   buffers.x_vec(), args.x_offset, args.x_inc,
+                                   1, &queue_plain, 0, nullptr, &event);
+      clWaitForEvents(1, &event);
+      return static_cast<StatusCode>(status);
+    }
+  #endif
+
+  // Describes how to run the CPU BLAS routine (for correctness/performance comparison)
+  #ifdef CLBLAST_REF_CBLAS
+    static StatusCode RunReference2(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
+      std::vector<T> scalar_cpu(args.scalar_size, static_cast<T>(0));
+      std::vector<T> x_vec_cpu(args.x_size, static_cast<T>(0));
+      buffers.scalar.Read(queue, args.scalar_size, scalar_cpu);
+      buffers.x_vec.Read(queue, args.x_size, x_vec_cpu);
+      cblasXasum(args.n,
+                 scalar_cpu, args.asum_offset,
+                 x_vec_cpu, args.x_offset, args.x_inc);
+      buffers.scalar.Write(queue, args.scalar_size, scalar_cpu);
+      return StatusCode::kSuccess;
+    }
+  #endif
+
+  // Describes how to download the results of the computation (more importantly: which buffer)
+  static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
+    std::vector<T> result(args.scalar_size, static_cast<T>(0));
+    buffers.scalar.Read(queue, args.scalar_size, result);
+    return result;
+  }
+
+  // Describes how to compute the indices of the result buffer
+  static size_t ResultID1(const Arguments<T> &) { return 1; } // N/A for this routine
+  static size_t ResultID2(const Arguments<T> &) { return 1; } // N/A for this routine
+  static size_t GetResultIndex(const Arguments<T> &args, const size_t, const size_t) {
+    return args.asum_offset;
+  }
+
+  // Describes how to compute performance metrics
+  static size_t GetFlops(const Arguments<T> &args) {
+    return args.n;
+  }
+  static size_t GetBytes(const Arguments<T> &args) {
+    return ((args.n) + 1) * sizeof(T);
+  }
+};
+
+// =================================================================================================
+} // namespace clblast
+
+// CLBLAST_TEST_ROUTINES_XASUM_H_
+#endif