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

CMakeLists.txt

@@ -121,7 +121,7 @@ include_directories(${clblast_SOURCE_DIR}/include ${OPENCL_INCLUDE_DIRS})
 set(KERNELS copy pad transpose padtranspose xaxpy xdot xger xgemm xgemv)
 set(SAMPLE_PROGRAMS_CPP sgemm)
 set(SAMPLE_PROGRAMS_C sgemm)
-set(LEVEL1_ROUTINES xswap xscal xcopy xaxpy xdot xdotu xdotc xnrm2)
+set(LEVEL1_ROUTINES xswap xscal xcopy xaxpy xdot xdotu xdotc xnrm2 xasum)
 set(LEVEL2_ROUTINES xgemv xgbmv xhemv xhbmv xhpmv xsymv xsbmv xspmv xtrmv xtbmv xtpmv
                     xger xgeru xgerc xher xhpr xher2 xhpr2 xsyr xspr xsyr2 xspr2)
 set(LEVEL3_ROUTINES xgemm xsymm xhemm xsyrk xherk xsyr2k xher2k xtrmm)

README.md

@@ -185,7 +185,7 @@ CLBlast is in active development but already supports almost all the BLAS routin
 | xDOTU    | - | - | ✔ | ✔ |         |
 | xDOTC    | - | - | ✔ | ✔ |         |
 | xNRM2    | ✔ | ✔ | ✔ | ✔ |         |
-| xASUM    |   |   |   |   |         |
+| xASUM    | ✔ | ✔ | ✔ | ✔ |         |
 | IxAMAX   |   |   |   |   |         |
 
 | Level-2  | S | D | C | Z | Notes   |

include/internal/routines/level1/xasum.h

@@ -0,0 +1,56 @@
+
+// =================================================================================================
+// 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 Xasum routine. The precision is implemented using a template argument.
+//
+// =================================================================================================
+
+#ifndef CLBLAST_ROUTINES_XASUM_H_
+#define CLBLAST_ROUTINES_XASUM_H_
+
+#include "internal/routine.h"
+
+namespace clblast {
+// =================================================================================================
+
+// See comment at top of file for a description of the class
+template <typename T>
+class Xasum: public Routine<T> {
+ public:
+
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::source_string_;
+  using Routine<T>::queue_;
+  using Routine<T>::event_;
+  using Routine<T>::context_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::TestVectorX;
+  using Routine<T>::TestVectorDot;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
+
+  // Constructor
+  Xasum(Queue &queue, EventPointer event, const std::string &name = "ASUM");
+
+  // Templated-precision implementation of the routine
+  StatusCode DoAsum(const size_t n,
+                    const Buffer<T> &asum_buffer, const size_t asum_offset,
+                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc);
+
+ private:
+  // Static variable to get the precision
+  const static Precision precision_;
+};
+
+// =================================================================================================
+} // namespace clblast
+
+// CLBLAST_ROUTINES_XASUM_H_
+#endif

include/internal/utilities.h

@@ -65,6 +65,7 @@ constexpr auto kArgCOffset = "offc";
 constexpr auto kArgAPOffset = "offap";
 constexpr auto kArgDotOffset = "offdot";
 constexpr auto kArgNrm2Offset = "offnrm2";
+constexpr auto kArgAsumOffset = "offasum";
 constexpr auto kArgAlpha = "alpha";
 constexpr auto kArgBeta = "beta";
 
@@ -119,6 +120,7 @@ struct Arguments {
   size_t ap_offset = 0;
   size_t dot_offset = 0;
   size_t nrm2_offset = 0;
+  size_t asum_offset = 0;
   T alpha = T{1.0};
   T beta = T{1.0};
   size_t x_size = 1;

scripts/generator/generator.py

@@ -67,7 +67,7 @@ routines = [
   Routine(True,  "1", "dotu",  T,  [C,Z],     ["n"], [], ["x","y"], ["dot"], [], "n", "Dot product of two complex vectors"),
   Routine(True,  "1", "dotc",  T,  [C,Z],     ["n"], [], ["x","y"], ["dot"], [], "n", "Dot product of two complex vectors, one conjugated"),
   Routine(True,  "1", "nrm2",  T, [S,D,Sc,Dz],["n"], [], ["x"], ["nrm2"], [], "2*n", "Euclidian norm of a vector"),
-  Routine(False, "1", "asum",  T, [S,D,Sc,Dz],["n"], [], ["x"], ["asum"], [], "n", "Absolute sum of values in a vector"),
+  Routine(True,  "1", "asum",  T, [S,D,Sc,Dz],["n"], [], ["x"], ["asum"], [], "n", "Absolute sum of values in a vector"),
 ],
 [ # Level 2: matrix-vector
   Routine(True,  "2a", "gemv",  T,  [S,D,C,Z], ["m","n"], ["layout","a_transpose"], ["a","x"], ["y"], ["alpha","beta"], "", "General matrix-vector multiplication"),
@@ -289,7 +289,7 @@ files = [
   path_clblast+"/test/wrapper_clblas.h",
   path_clblast+"/test/wrapper_cblas.h",
 ]
-header_lines = [84, 65, 93, 22, 22, 38]
+header_lines = [84, 66, 93, 22, 22, 38]
 footer_lines = [6, 3, 9, 2, 6, 6]
 
 # Checks whether the command-line arguments are valid; exists otherwise

src/clblast.cc

@@ -27,6 +27,7 @@
 #include "internal/routines/level1/xdotu.h"
 #include "internal/routines/level1/xdotc.h"
 #include "internal/routines/level1/xnrm2.h"
+#include "internal/routines/level1/xasum.h"
 
 // BLAS level-2 includes
 #include "internal/routines/level2/xgemv.h"
@@ -398,11 +399,17 @@ template StatusCode PUBLIC_API Nrm2<double2>(const size_t,
 
 // Absolute sum of values in a vector: SASUM/DASUM/ScASUM/DzASUM
 template <typename T>
-StatusCode Asum(const size_t,
-                cl_mem, const size_t,
-                const cl_mem, const size_t, const size_t,
-                cl_command_queue*, cl_event*) {
-  return StatusCode::kNotImplemented;
+StatusCode Asum(const size_t n,
+                cl_mem asum_buffer, const size_t asum_offset,
+                const cl_mem x_buffer, const size_t x_offset, const size_t x_inc,
+                cl_command_queue* queue, cl_event* event) {
+  auto queue_cpp = Queue(*queue);
+  auto routine = Xasum<T>(queue_cpp, event);
+  auto status = routine.SetUp();
+  if (status != StatusCode::kSuccess) { return status; }
+  return routine.DoAsum(n,
+                        Buffer<T>(asum_buffer), asum_offset,
+                        Buffer<T>(x_buffer), x_offset, x_inc);
 }
 template StatusCode PUBLIC_API Asum<float>(const size_t,
                                            cl_mem, const size_t,

src/kernels/common.opencl

@@ -109,6 +109,13 @@ R"(
   #define SetToOne(a) a = ONE
 #endif
 
+// The absolute value (component-wise)
+#if PRECISION == 3232 || PRECISION == 6464
+  #define AbsoluteValue(value) value.x = fabs(value.x); value.y = fabs(value.y)
+#else
+  #define AbsoluteValue(value) value = fabs(value)
+#endif
+
 // Adds two complex variables
 #if PRECISION == 3232 || PRECISION == 6464
   #define Add(c, a, b) c.x = a.x + b.x; c.y = a.y + b.y

src/kernels/level1/xasum.opencl

@@ -0,0 +1,108 @@
+
+// =================================================================================================
+// 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 contains the Xasum kernel. It implements a absolute sum computation using reduction
+// kernels. Reduction is split in two parts. In the first (main) kernel the X vector is loaded,
+// followed by a per-thread and a per-workgroup reduction. The second (epilogue) kernel
+// is executed with a single workgroup only, computing the final result.
+//
+// =================================================================================================
+
+// Enables loading of this file using the C++ pre-processor's #include (C++11 standard raw string
+// literal). Comment-out this line for syntax-highlighting when developing.
+R"(
+
+// Parameters set by the tuner or by the database. Here they are given a basic default value in case
+// this kernel file is used outside of the CLBlast library.
+#ifndef WGS1
+  #define WGS1 64     // The local work-group size of the main kernel
+#endif
+#ifndef WGS2
+  #define WGS2 64     // The local work-group size of the epilogue kernel
+#endif
+
+// =================================================================================================
+
+// The main reduction kernel, performing the loading and the majority of the operation
+__attribute__((reqd_work_group_size(WGS1, 1, 1)))
+__kernel void Xasum(const int n,
+                    const __global real* restrict xgm, const int x_offset, const int x_inc,
+                    __global real* output) {
+  __local real lm[WGS1];
+  const int lid = get_local_id(0);
+  const int wgid = get_group_id(0);
+  const int num_groups = get_num_groups(0);
+
+  // Performs loading and the first steps of the reduction
+  real acc;
+  SetToZero(acc);
+  int id = wgid*WGS1 + lid;
+  while (id < n) {
+    real x = xgm[id*x_inc + x_offset];
+    AbsoluteValue(x);
+    Add(acc, acc, x);
+    id += WGS1*num_groups;
+  }
+  lm[lid] = acc;
+  barrier(CLK_LOCAL_MEM_FENCE);
+
+  // Performs reduction in local memory
+  #pragma unroll
+  for (int s=WGS1/2; s>0; s=s>>1) {
+    if (lid < s) {
+      Add(lm[lid], lm[lid], lm[lid + s]);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+  }
+
+  // Stores the per-workgroup result
+  if (lid == 0) {
+    output[wgid] = lm[0];
+  }
+}
+
+// =================================================================================================
+
+// The epilogue reduction kernel, performing the final bit of the operation. This kernel has to
+// be launched with a single workgroup only.
+__attribute__((reqd_work_group_size(WGS2, 1, 1)))
+__kernel void XasumEpilogue(const __global real* restrict input,
+                            __global real* asum, const int asum_offset) {
+  __local real lm[WGS2];
+  const int lid = get_local_id(0);
+
+  // Performs the first step of the reduction while loading the data
+  Add(lm[lid], input[lid], input[lid + WGS2]);
+  barrier(CLK_LOCAL_MEM_FENCE);
+
+  // Performs reduction in local memory
+  #pragma unroll
+  for (int s=WGS2/2; s>0; s=s>>1) {
+    if (lid < s) {
+      Add(lm[lid], lm[lid], lm[lid + s]);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+  }
+
+  // Computes the absolute value and stores the final result
+  if (lid == 0) {
+    #if PRECISION == 3232 || PRECISION == 6464
+      asum[asum_offset].x = lm[0].x + lm[0].y; // the result is a non-complex number
+    #else
+      asum[asum_offset] = lm[0];
+    #endif
+  }
+}
+
+// =================================================================================================
+
+// End of the C++11 raw string literal
+)"
+
+// =================================================================================================

src/kernels/level1/xnrm2.opencl

@@ -7,9 +7,9 @@
 // Author(s):
 //   Cedric Nugteren <www.cedricnugteren.nl>
 //
-// This file contains the Xnrm2 kernel. It implements a dot-product computation using reduction
-// kernels. Reduction is split in two parts. In the first (main) kernel the X and Y vectors are
-// multiplied, followed by a per-thread and a per-workgroup reduction. The second (epilogue) kernel
+// This file contains the Xnrm2 kernel. It implements a squared norm computation using reduction
+// kernels. Reduction is split in two parts. In the first (main) kernel the X vector is squared,
+// followed by a per-thread and a per-workgroup reduction. The second (epilogue) kernel
 // is executed with a single workgroup only, computing the final result.
 //
 // =================================================================================================
@@ -29,7 +29,7 @@ R"(
 
 // =================================================================================================
 
-// The main reduction kernel, performing the multiplication and the majority of the sum operation
+// The main reduction kernel, performing the multiplication and the majority of the operation
 __attribute__((reqd_work_group_size(WGS1, 1, 1)))
 __kernel void Xnrm2(const int n,
                     const __global real* restrict xgm, const int x_offset, const int x_inc,
@@ -70,7 +70,7 @@ __kernel void Xnrm2(const int n,
 
 // =================================================================================================
 
-// The epilogue reduction kernel, performing the final bit of the sum operation. This kernel has to
+// The epilogue reduction kernel, performing the final bit of the operation. This kernel has to
 // be launched with a single workgroup only.
 __attribute__((reqd_work_group_size(WGS2, 1, 1)))
 __kernel void Xnrm2Epilogue(const __global real* restrict input,

src/routines/level1/xasum.cc

@@ -0,0 +1,109 @@
+
+// =================================================================================================
+// 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 Xasum class (see the header for information about the class).
+//
+// =================================================================================================
+
+#include "internal/routines/level1/xasum.h"
+
+#include <string>
+#include <vector>
+
+namespace clblast {
+// =================================================================================================
+
+// Specific implementations to get the memory-type based on a template argument
+template <> const Precision Xasum<float>::precision_ = Precision::kSingle;
+template <> const Precision Xasum<double>::precision_ = Precision::kDouble;
+template <> const Precision Xasum<float2>::precision_ = Precision::kComplexSingle;
+template <> const Precision Xasum<double2>::precision_ = Precision::kComplexDouble;
+
+// =================================================================================================
+
+// Constructor: forwards to base class constructor
+template <typename T>
+Xasum<T>::Xasum(Queue &queue, EventPointer event, const std::string &name):
+    Routine<T>(queue, event, name, {"Xdot"}, precision_) {
+  source_string_ =
+    #include "../../kernels/level1/xasum.opencl"
+  ;
+}
+
+// =================================================================================================
+
+// The main routine
+template <typename T>
+StatusCode Xasum<T>::DoAsum(const size_t n,
+                            const Buffer<T> &asum_buffer, const size_t asum_offset,
+                            const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc) {
+
+  // Makes sure all dimensions are larger than zero
+  if (n == 0) { return StatusCode::kInvalidDimension; }
+
+  // Tests the vectors for validity
+  auto status = TestVectorX(n, x_buffer, x_offset, x_inc, sizeof(T));
+  if (ErrorIn(status)) { return status; }
+  status = TestVectorDot(1, asum_buffer, asum_offset, sizeof(T));
+  if (ErrorIn(status)) { return status; }
+
+  // Retrieves the Xasum kernels from the compiled binary
+  try {
+    auto& program = GetProgramFromCache();
+    auto kernel1 = Kernel(program, "Xasum");
+    auto kernel2 = Kernel(program, "XasumEpilogue");
+
+    // Creates the buffer for intermediate values
+    auto temp_size = 2*db_["WGS2"];
+    auto temp_buffer = Buffer<T>(context_, temp_size);
+
+    // Sets the kernel arguments
+    kernel1.SetArgument(0, static_cast<int>(n));
+    kernel1.SetArgument(1, x_buffer());
+    kernel1.SetArgument(2, static_cast<int>(x_offset));
+    kernel1.SetArgument(3, static_cast<int>(x_inc));
+    kernel1.SetArgument(4, temp_buffer());
+
+    // Event waiting list
+    auto eventWaitList = std::vector<Event>();
+
+    // Launches the main kernel
+    auto global1 = std::vector<size_t>{db_["WGS1"]*temp_size};
+    auto local1 = std::vector<size_t>{db_["WGS1"]};
+    auto kernelEvent = Event();
+    status = RunKernel(kernel1, global1, local1, kernelEvent.pointer());
+    if (ErrorIn(status)) { return status; }
+    eventWaitList.push_back(kernelEvent);
+
+    // Sets the arguments for the epilogue kernel
+    kernel2.SetArgument(0, temp_buffer());
+    kernel2.SetArgument(1, asum_buffer());
+    kernel2.SetArgument(2, static_cast<int>(asum_offset));
+
+    // Launches the epilogue kernel
+    auto global2 = std::vector<size_t>{db_["WGS2"]};
+    auto local2 = std::vector<size_t>{db_["WGS2"]};
+    status = RunKernel(kernel2, global2, local2, event_, eventWaitList);
+    if (ErrorIn(status)) { return status; }
+
+    // Succesfully finished the computation
+    return StatusCode::kSuccess;
+  } catch (...) { return StatusCode::kInvalidKernel; }
+}
+
+// =================================================================================================
+
+// Compiles the templated class
+template class Xasum<float>;
+template class Xasum<double>;
+template class Xasum<float2>;
+template class Xasum<double2>;
+
+// =================================================================================================
+} // namespace clblast

src/routines/level1/xnrm2.cc

@@ -69,6 +69,7 @@ StatusCode Xnrm2<T>::DoNrm2(const size_t n,
     kernel1.SetArgument(2, static_cast<int>(x_offset));
     kernel1.SetArgument(3, static_cast<int>(x_inc));
     kernel1.SetArgument(4, temp_buffer());
+
     // Event waiting list
     auto eventWaitList = std::vector<Event>();
 

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);
+                                                      }
                                                     }
                                                   }
                                                 }

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_) {

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