Merge branch 'devel'

.gitignore

@@ -1,6 +1,4 @@
 
-common.h
-
 gpu-stream-cuda
 gpu-stream-ocl
 gpu-stream-acc

ACCStream.cpp

@@ -36,13 +36,19 @@ ACCStream<T>::~ACCStream()
 }
 
 template <class T>
-void ACCStream<T>::write_arrays(const std::vector<T>& h_a, const std::vector<T>& h_b, const std::vector<T>& h_c)
+void ACCStream<T>::init_arrays(T initA, T initB, T initC)
 {
-  T *a = this->a;
+  unsigned int array_size = this->array_size;
-  T *b = this->b;
+  T * restrict a = this->a;
-  T *c = this->c;
+  T * restrict b = this->b;
-  #pragma acc update device(a[0:array_size], b[0:array_size], c[0:array_size])
+  T * restrict c = this->c;
-  {}
+  #pragma acc kernels present(a[0:array_size], b[0:array_size], c[0:array_size]) wait
+  for (int i = 0; i < array_size; i++)
+  {
+    a[i] = initA;
+    b[i] = initB;
+    c[i] = initC;
+  }
 }
 
 template <class T>
@@ -112,6 +118,24 @@ void ACCStream<T>::triad()
     a[i] = b[i] + scalar * c[i];
   }
 }
+
+template <class T>
+T ACCStream<T>::dot()
+{
+  T sum = 0.0;
+
+  unsigned int array_size = this->array_size;
+  T * restrict a = this->a;
+  T * restrict b = this->b;
+  #pragma acc kernels present(a[0:array_size], b[0:array_size]) wait
+  for (int i = 0; i < array_size; i++)
+  {
+    sum += a[i] * b[i];
+  }
+
+  return sum;
+}
+
 void listDevices(void)
 {
   // Get number of devices

ACCStream.h

@@ -35,8 +35,9 @@ class ACCStream : public Stream<T>
     virtual void add() override;
     virtual void mul() override;
     virtual void triad() override;
+    virtual T dot() override;
 
-    virtual void write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
+    virtual void init_arrays(T initA, T initB, T initC) override;
     virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
 
 

CMakeLists.txt

@@ -20,7 +20,7 @@ include(CheckIncludeFileCXX)
 include(CheckCXXCompilerFlag)
 
 set(gpu-stream_VERSION_MAJOR 2)
-set(gpu-stream_VERSION_MINOR 1)
+set(gpu-stream_VERSION_MINOR 2)
 
 configure_file(common.h.in common.h)
 include_directories(${CMAKE_BINARY_DIR})

CUDAStream.cu

@@ -8,8 +8,6 @@
 
 #include "CUDAStream.h"
 
-#define TBSIZE 1024
-
 void check_error(void)
 {
   cudaError_t err = cudaGetLastError();
@@ -47,6 +45,9 @@ CUDAStream<T>::CUDAStream(const unsigned int ARRAY_SIZE, const int device_index)
 
   array_size = ARRAY_SIZE;
 
+  // Allocate the host array for partial sums for dot kernels
+  sums = (T*)malloc(sizeof(T) * DOT_NUM_BLOCKS);
+
   // Check buffers fit on the device
   cudaDeviceProp props;
   cudaGetDeviceProperties(&props, 0);
@@ -60,29 +61,42 @@ CUDAStream<T>::CUDAStream(const unsigned int ARRAY_SIZE, const int device_index)
   check_error();
   cudaMalloc(&d_c, ARRAY_SIZE*sizeof(T));
   check_error();
+  cudaMalloc(&d_sum, DOT_NUM_BLOCKS*sizeof(T));
+  check_error();
 }
 
 
 template <class T>
 CUDAStream<T>::~CUDAStream()
 {
+  free(sums);
+
   cudaFree(d_a);
   check_error();
   cudaFree(d_b);
   check_error();
   cudaFree(d_c);
   check_error();
+  cudaFree(d_sum);
+  check_error();
+}
+
+
+template <typename T>
+__global__ void init_kernel(T * a, T * b, T * c, T initA, T initB, T initC)
+{
+  const int i = blockDim.x * blockIdx.x + threadIdx.x;
+  a[i] = initA;
+  b[i] = initB;
+  c[i] = initC;
 }
 
 template <class T>
-void CUDAStream<T>::write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c)
+void CUDAStream<T>::init_arrays(T initA, T initB, T initC)
 {
-  // Copy host memory to device
+  init_kernel<<<array_size/TBSIZE, TBSIZE>>>(d_a, d_b, d_c, initA, initB, initC);
-  cudaMemcpy(d_a, a.data(), a.size()*sizeof(T), cudaMemcpyHostToDevice);
   check_error();
-  cudaMemcpy(d_b, b.data(), b.size()*sizeof(T), cudaMemcpyHostToDevice);
+  cudaDeviceSynchronize();
-  check_error();
-  cudaMemcpy(d_c, c.data(), c.size()*sizeof(T), cudaMemcpyHostToDevice);
   check_error();
 }
 
@@ -165,6 +179,48 @@ void CUDAStream<T>::triad()
   check_error();
 }
 
+template <class T>
+__global__ void dot_kernel(const T * a, const T * b, T * sum, unsigned int array_size)
+{
+
+  extern __shared__ __align__(sizeof(T)) unsigned char smem[];
+  T *tb_sum = reinterpret_cast<T*>(smem);
+
+  int i = blockDim.x * blockIdx.x + threadIdx.x;
+  const size_t local_i = threadIdx.x;
+
+  tb_sum[local_i] = 0.0;
+  for (; i < array_size; i += blockDim.x*gridDim.x)
+    tb_sum[local_i] += a[i] * b[i];
+
+  for (int offset = blockDim.x / 2; offset > 0; offset /= 2)
+  {
+    __syncthreads();
+    if (local_i < offset)
+    {
+      tb_sum[local_i] += tb_sum[local_i+offset];
+    }
+  }
+
+  if (local_i == 0)
+    sum[blockIdx.x] = tb_sum[local_i];
+}
+
+template <class T>
+T CUDAStream<T>::dot()
+{
+  dot_kernel<<<DOT_NUM_BLOCKS, TBSIZE, sizeof(T)*TBSIZE>>>(d_a, d_b, d_sum, array_size);
+  check_error();
+
+  cudaMemcpy(sums, d_sum, DOT_NUM_BLOCKS*sizeof(T), cudaMemcpyDeviceToHost);
+  check_error();
+
+  T sum = 0.0;
+  for (int i = 0; i < DOT_NUM_BLOCKS; i++)
+    sum += sums[i];
+
+  return sum;
+}
 
 void listDevices(void)
 {

CUDAStream.h

@@ -15,16 +15,24 @@
 
 #define IMPLEMENTATION_STRING "CUDA"
 
+#define TBSIZE 1024
+#define DOT_NUM_BLOCKS 256
+
 template <class T>
 class CUDAStream : public Stream<T>
 {
   protected:
     // Size of arrays
     unsigned int array_size;
+
+    // Host array for partial sums for dot kernel
+    T *sums;
+
     // Device side pointers to arrays
     T *d_a;
     T *d_b;
     T *d_c;
+    T *d_sum;
 
 
   public:
@@ -36,8 +44,9 @@ class CUDAStream : public Stream<T>
     virtual void add() override;
     virtual void mul() override;
     virtual void triad() override;
+    virtual T dot() override;
 
-    virtual void write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
+    virtual void init_arrays(T initA, T initB, T initC) override;
     virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
 
 };

HIPStream.cu

@@ -74,15 +74,21 @@ HIPStream<T>::~HIPStream()
   check_error();
 }
 
-template <class T>
+template <typename T>
-void HIPStream<T>::write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c)
+__global__ void init_kernel(hipLaunchParm lp, T * a, T * b, T * c, T initA, T initB, T initC)
 {
-  // Copy host memory to device
+  const int i = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
-  hipMemcpy(d_a, a.data(), a.size()*sizeof(T), hipMemcpyHostToDevice);
+  a[i] = initA;
+  b[i] = initB;
+  c[i] = initC;
+}
+
+template <class T>
+void HIPStream<T>::init_arrays(T initA, T initB, T initC)
+{
+  hipLaunchKernel(HIP_KERNEL_NAME(init_kernel), dim3(array_size/TBSIZE), dim3(TBSIZE), 0, 0, d_a, d_b, d_c, initA, initB, initC);
   check_error();
-  hipMemcpy(d_b, b.data(), b.size()*sizeof(T), hipMemcpyHostToDevice);
+  hipDeviceSynchronize();
-  check_error();
-  hipMemcpy(d_c, c.data(), c.size()*sizeof(T), hipMemcpyHostToDevice);
   check_error();
 }
 

HIPStream.h

@@ -37,7 +37,7 @@ class HIPStream : public Stream<T>
     virtual void mul() override;
     virtual void triad() override;
 
-    virtual void write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
+    virtual void init_arrays(T initA, T initB, T initC) override;
     virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
 
 };

KOKKOSStream.cpp

@@ -34,18 +34,18 @@ KOKKOSStream<T>::~KOKKOSStream()
 }
 
 template <class T>
-void KOKKOSStream<T>::write_arrays(
+void KOKKOSStream<T>::init_arrays(T initA, T initB, T initC)
-        const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c)
 {
-  for(int ii = 0; ii < array_size; ++ii)
+  View<double*, DEVICE> a(*d_a);
+  View<double*, DEVICE> b(*d_b);
+  View<double*, DEVICE> c(*d_c);
+  parallel_for(array_size, KOKKOS_LAMBDA (const int index)
   {
-    (*hm_a)(ii) = a[ii];
+    a[index] = initA;
-    (*hm_b)(ii) = b[ii];
+    b[index] - initB;
-    (*hm_c)(ii) = c[ii];
+    c[index] = initC;
-  }
+  });
-  deep_copy(*d_a, *hm_a);
+  Kokkos::fence();
-  deep_copy(*d_b, *hm_b);
-  deep_copy(*d_c, *hm_c);
 }
 
 template <class T>
@@ -121,6 +121,23 @@ void KOKKOSStream<T>::triad()
   Kokkos::fence();
 }
 
+template <class T>
+T KOKKOSStream<T>::dot()
+{
+  View<double *, DEVICE> a(*d_a);
+  View<double *, DEVICE> b(*d_b);
+
+  T sum = 0.0;
+
+  parallel_reduce(array_size, KOKKOS_LAMBDA (const int index, double &tmp)
+  {
+    tmp += a[index] * b[index];
+  }, sum);
+
+  return sum;
+
+}
+
 void listDevices(void)
 {
   std::cout << "This is not the device you are looking for.";

KOKKOSStream.hpp

@@ -47,9 +47,9 @@ class KOKKOSStream : public Stream<T>
     virtual void add() override;
     virtual void mul() override;
     virtual void triad() override;
+    virtual T dot() override;
 
-    virtual void write_arrays(
+    virtual void init_arrays(T initA, T initB, T initC) override;
-            const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
     virtual void read_arrays(
             std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
 };

OCLStream.cpp

@@ -16,6 +16,18 @@ std::string kernels{R"CLC(
 
   constant TYPE scalar = startScalar;
 
+  kernel void init(
+    global TYPE * restrict a,
+    global TYPE * restrict b,
+    global TYPE * restrict c,
+    TYPE initA, TYPE initB, TYPE initC)
+  {
+    const size_t i = get_global_id(0);
+    a[i] = initA;
+    b[i] = initB;
+    c[i] = initC;
+  }
+
   kernel void copy(
     global const TYPE * restrict a,
     global TYPE * restrict c)
@@ -50,6 +62,32 @@ std::string kernels{R"CLC(
     a[i] = b[i] + scalar * c[i];
   }
 
+  kernel void stream_dot(
+    global const TYPE * restrict a,
+    global const TYPE * restrict b,
+    global TYPE * restrict sum,
+    local TYPE * restrict wg_sum,
+    int array_size)
+  {
+    size_t i = get_global_id(0);
+    const size_t local_i = get_local_id(0);
+    wg_sum[local_i] = 0.0;
+    for (; i < array_size; i += get_global_size(0))
+      wg_sum[local_i] += a[i] * b[i];
+
+    for (int offset = get_local_size(0) / 2; offset > 0; offset /= 2)
+    {
+      barrier(CLK_LOCAL_MEM_FENCE);
+      if (local_i < offset)
+      {
+        wg_sum[local_i] += wg_sum[local_i+offset];
+      }
+    }
+
+    if (local_i == 0)
+      sum[get_group_id(0)] = wg_sum[local_i];
+  }
+
 )CLC"};
 
 
@@ -64,9 +102,22 @@ OCLStream<T>::OCLStream(const unsigned int ARRAY_SIZE, const int device_index)
     throw std::runtime_error("Invalid device index");
   device = devices[device_index];
 
+  // Determine sensible dot kernel NDRange configuration
+  if (device.getInfo<CL_DEVICE_TYPE>() & CL_DEVICE_TYPE_CPU)
+  {
+    dot_num_groups = device.getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>();
+    dot_wgsize     = device.getInfo<CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE>() * 2;
+  }
+  else
+  {
+    dot_num_groups = device.getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>() * 4;
+    dot_wgsize     = device.getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>();
+  }
+
   // Print out device information
   std::cout << "Using OpenCL device " << getDeviceName(device_index) << std::endl;
   std::cout << "Driver: " << getDeviceDriver(device_index) << std::endl;
+  std::cout << "Reduction kernel config: " << dot_num_groups << " groups of size " << dot_wgsize << std::endl;
 
   context = cl::Context(device);
   queue = cl::CommandQueue(context);
@@ -101,10 +152,12 @@ OCLStream<T>::OCLStream(const unsigned int ARRAY_SIZE, const int device_index)
   }
 
   // Create kernels
+  init_kernel = new cl::KernelFunctor<cl::Buffer, cl::Buffer, cl::Buffer, T, T, T>(program, "init");
   copy_kernel = new cl::KernelFunctor<cl::Buffer, cl::Buffer>(program, "copy");
   mul_kernel = new cl::KernelFunctor<cl::Buffer, cl::Buffer>(program, "mul");
   add_kernel = new cl::KernelFunctor<cl::Buffer, cl::Buffer, cl::Buffer>(program, "add");
   triad_kernel = new cl::KernelFunctor<cl::Buffer, cl::Buffer, cl::Buffer>(program, "triad");
+  dot_kernel = new cl::KernelFunctor<cl::Buffer, cl::Buffer, cl::Buffer, cl::LocalSpaceArg, cl_int>(program, "stream_dot");
 
   array_size = ARRAY_SIZE;
 
@@ -120,12 +173,15 @@ OCLStream<T>::OCLStream(const unsigned int ARRAY_SIZE, const int device_index)
   d_a = cl::Buffer(context, CL_MEM_READ_WRITE, sizeof(T) * ARRAY_SIZE);
   d_b = cl::Buffer(context, CL_MEM_READ_WRITE, sizeof(T) * ARRAY_SIZE);
   d_c = cl::Buffer(context, CL_MEM_READ_WRITE, sizeof(T) * ARRAY_SIZE);
+  d_sum = cl::Buffer(context, CL_MEM_WRITE_ONLY, sizeof(T) * dot_num_groups);
 
+  sums = std::vector<T>(dot_num_groups);
 }
 
 template <class T>
 OCLStream<T>::~OCLStream()
 {
+  delete init_kernel;
   delete copy_kernel;
   delete mul_kernel;
   delete add_kernel;
@@ -173,11 +229,29 @@ void OCLStream<T>::triad()
 }
 
 template <class T>
-void OCLStream<T>::write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c)
+T OCLStream<T>::dot()
 {
-  cl::copy(queue, a.begin(), a.end(), d_a);
+  (*dot_kernel)(
-  cl::copy(queue, b.begin(), b.end(), d_b);
+    cl::EnqueueArgs(queue, cl::NDRange(dot_num_groups*dot_wgsize), cl::NDRange(dot_wgsize)),
-  cl::copy(queue, c.begin(), c.end(), d_c);
+    d_a, d_b, d_sum, cl::Local(sizeof(T) * dot_wgsize), array_size
+  );
+  cl::copy(queue, d_sum, sums.begin(), sums.end());
+
+  T sum = 0.0;
+  for (T val : sums)
+    sum += val;
+
+  return sum;
+}
+
+template <class T>
+void OCLStream<T>::init_arrays(T initA, T initB, T initC)
+{
+  (*init_kernel)(
+    cl::EnqueueArgs(queue, cl::NDRange(array_size)),
+    d_a, d_b, d_c, initA, initB, initC
+  );
+  queue.finish();
 }
 
 template <class T>

OCLStream.h

@@ -28,20 +28,30 @@ class OCLStream : public Stream<T>
     // Size of arrays
     unsigned int array_size;
 
+    // Host array for partial sums for dot kernel
+    std::vector<T> sums;
+
     // Device side pointers to arrays
     cl::Buffer d_a;
     cl::Buffer d_b;
     cl::Buffer d_c;
+    cl::Buffer d_sum;
 
     // OpenCL objects
     cl::Device device;
     cl::Context context;
     cl::CommandQueue queue;
 
+    cl::KernelFunctor<cl::Buffer, cl::Buffer, cl::Buffer, T, T, T> *init_kernel;
     cl::KernelFunctor<cl::Buffer, cl::Buffer> *copy_kernel;
     cl::KernelFunctor<cl::Buffer, cl::Buffer> * mul_kernel;
     cl::KernelFunctor<cl::Buffer, cl::Buffer, cl::Buffer> *add_kernel;
     cl::KernelFunctor<cl::Buffer, cl::Buffer, cl::Buffer> *triad_kernel;
+    cl::KernelFunctor<cl::Buffer, cl::Buffer, cl::Buffer, cl::LocalSpaceArg, cl_int> *dot_kernel;
+
+    // NDRange configuration for the dot kernel
+    size_t dot_num_groups;
+    size_t dot_wgsize;
 
   public:
 
@@ -52,8 +62,9 @@ class OCLStream : public Stream<T>
     virtual void add() override;
     virtual void mul() override;
     virtual void triad() override;
+    virtual T dot() override;
 
-    virtual void write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
+    virtual void init_arrays(T initA, T initB, T initC) override;
     virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
 
 };

OMP3Stream.cpp

@@ -1,111 +0,0 @@
-
-// Copyright (c) 2015-16 Tom Deakin, Simon McIntosh-Smith,
-// University of Bristol HPC
-//
-// For full license terms please see the LICENSE file distributed with this
-// source code
-
-#include "OMP3Stream.h"
-
-template <class T>
-OMP3Stream<T>::OMP3Stream(const unsigned int ARRAY_SIZE, T *a, T *b, T *c)
-{
-  array_size = ARRAY_SIZE;
-  this->a = (T*)malloc(sizeof(T)*array_size);
-  this->b = (T*)malloc(sizeof(T)*array_size);
-  this->c = (T*)malloc(sizeof(T)*array_size);
-}
-
-template <class T>
-OMP3Stream<T>::~OMP3Stream()
-{
-  free(a);
-  free(b);
-  free(c);
-}
-
-
-template <class T>
-void OMP3Stream<T>::write_arrays(const std::vector<T>& h_a, const std::vector<T>& h_b, const std::vector<T>& h_c)
-{
-  #pragma omp parallel for
-  for (int i = 0; i < array_size; i++)
-  {
-    a[i] = h_a[i];
-    b[i] = h_b[i];
-    c[i] = h_c[i];
-  }
-}
-
-template <class T>
-void OMP3Stream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::vector<T>& h_c)
-{
-  #pragma omp parallel for
-  for (int i = 0; i < array_size; i++)
-  {
-    h_a[i] = a[i];
-    h_b[i] = b[i];
-    h_c[i] = c[i];
-  }
-}
-
-template <class T>
-void OMP3Stream<T>::copy()
-{
-  #pragma omp parallel for
-  for (int i = 0; i < array_size; i++)
-  {
-    c[i] = a[i];
-  }
-}
-
-template <class T>
-void OMP3Stream<T>::mul()
-{
-  const T scalar = startScalar;
-  #pragma omp parallel for
-  for (int i = 0; i < array_size; i++)
-  {
-    b[i] = scalar * c[i];
-  }
-}
-
-template <class T>
-void OMP3Stream<T>::add()
-{
-  #pragma omp parallel for
-  for (int i = 0; i < array_size; i++)
-  {
-    c[i] = a[i] + b[i];
-  }
-}
-
-template <class T>
-void OMP3Stream<T>::triad()
-{
-  const T scalar = startScalar;
-  #pragma omp parallel for
-  for (int i = 0; i < array_size; i++)
-  {
-    a[i] = b[i] + scalar * c[i];
-  }
-}
-
-void listDevices(void)
-{
-  std::cout << "0: CPU" << std::endl;
-}
-
-std::string getDeviceName(const int)
-{
-  return std::string("Device name unavailable");
-}
-
-std::string getDeviceDriver(const int)
-{
-  return std::string("Device driver unavailable");
-}
-
-
-template class OMP3Stream<float>;
-template class OMP3Stream<double>;

OMP3Stream.h

@@ -1,40 +0,0 @@
-
-// Copyright (c) 2015-16 Tom Deakin, Simon McIntosh-Smith,
-// University of Bristol HPC
-//
-// For full license terms please see the LICENSE file distributed with this
-// source code
-
-#pragma once
-
-#include <iostream>
-#include <stdexcept>
-
-#include "Stream.h"
-
-#define IMPLEMENTATION_STRING "Reference OpenMP"
-
-template <class T>
-class OMP3Stream : public Stream<T>
-{
-  protected:
-    // Size of arrays
-    unsigned int array_size;
-    // Device side pointers
-    T *a;
-    T *b;
-    T *c;
-
-  public:
-    OMP3Stream(const unsigned int, T*, T*, T*);
-    ~OMP3Stream();
-
-    virtual void copy() override;
-    virtual void add() override;
-    virtual void mul() override;
-    virtual void triad() override;
-
-    virtual void write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
-    virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
-
-};

OMPStream.cpp

@@ -5,26 +5,33 @@
 // For full license terms please see the LICENSE file distributed with this
 // source code
 
-#include "OMP45Stream.h"
+#include "OMPStream.h"
 
 template <class T>
-OMP45Stream<T>::OMP45Stream(const unsigned int ARRAY_SIZE, T *a, T *b, T *c, int device)
+OMPStream<T>::OMPStream(const unsigned int ARRAY_SIZE, T *a, T *b, T *c, int device)
 {
-  omp_set_default_device(device);
-
   array_size = ARRAY_SIZE;
 
+#ifdef OMP_TARGET_GPU
+  omp_set_default_device(device);
   // Set up data region on device
   this->a = a;
   this->b = b;
   this->c = c;
-  #pragma omp target enter data map(to: a[0:array_size], b[0:array_size], c[0:array_size])
+  #pragma omp target enter data map(alloc: a[0:array_size], b[0:array_size], c[0:array_size])
   {}
+#else
+  // Allocate on the host
+  this->a = (T*)malloc(sizeof(T)*array_size);
+  this->b = (T*)malloc(sizeof(T)*array_size);
+  this->c = (T*)malloc(sizeof(T)*array_size);
+#endif
 }
 
 template <class T>
-OMP45Stream<T>::~OMP45Stream()
+OMPStream<T>::~OMPStream()
 {
+#ifdef OMP_TARGET_GPU
   // End data region on device
   unsigned int array_size = this->array_size;
   T *a = this->a;
@@ -32,35 +39,64 @@ OMP45Stream<T>::~OMP45Stream()
   T *c = this->c;
   #pragma omp target exit data map(release: a[0:array_size], b[0:array_size], c[0:array_size])
   {}
+#else
+  free(a);
+  free(b);
+  free(c);
+#endif
 }
 
 template <class T>
-void OMP45Stream<T>::write_arrays(const std::vector<T>& h_a, const std::vector<T>& h_b, const std::vector<T>& h_c)
+void OMPStream<T>::init_arrays(T initA, T initB, T initC)
 {
+  unsigned int array_size = this->array_size;
+#ifdef OMP_TARGET_GPU
   T *a = this->a;
   T *b = this->b;
   T *c = this->c;
-  #pragma omp target update to(a[0:array_size], b[0:array_size], c[0:array_size])
+  #pragma omp target teams distribute parallel for simd map(to: a[0:array_size], b[0:array_size], c[0:array_size])
-  {}
+#else
+  #pragma omp parallel for
+#endif
+  for (int i = 0; i < array_size; i++)
+  {
+    a[i] = initA;
+    b[i] = initB;
+    c[i] = initC;
+  }
 }
 
 template <class T>
-void OMP45Stream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::vector<T>& h_c)
+void OMPStream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::vector<T>& h_c)
 {
+#ifdef OMP_TARGET_GPU
   T *a = this->a;
   T *b = this->b;
   T *c = this->c;
   #pragma omp target update from(a[0:array_size], b[0:array_size], c[0:array_size])
   {}
+#else
+  #pragma omp parallel for
+  for (int i = 0; i < array_size; i++)
+  {
+    h_a[i] = a[i];
+    h_b[i] = b[i];
+    h_c[i] = c[i];
+  }
+#endif
 }
 
 template <class T>
-void OMP45Stream<T>::copy()
+void OMPStream<T>::copy()
 {
+#ifdef OMP_TARGET_GPU
   unsigned int array_size = this->array_size;
   T *a = this->a;
   T *c = this->c;
   #pragma omp target teams distribute parallel for simd map(to: a[0:array_size], c[0:array_size])
+#else
+  #pragma omp parallel for
+#endif
   for (int i = 0; i < array_size; i++)
   {
     c[i] = a[i];
@@ -68,14 +104,18 @@ void OMP45Stream<T>::copy()
 }
 
 template <class T>
-void OMP45Stream<T>::mul()
+void OMPStream<T>::mul()
 {
   const T scalar = startScalar;
 
+#ifdef OMP_TARGET_GPU
   unsigned int array_size = this->array_size;
   T *b = this->b;
   T *c = this->c;
   #pragma omp target teams distribute parallel for simd map(to: b[0:array_size], c[0:array_size])
+#else
+  #pragma omp parallel for
+#endif
   for (int i = 0; i < array_size; i++)
   {
     b[i] = scalar * c[i];
@@ -83,13 +123,17 @@ void OMP45Stream<T>::mul()
 }
 
 template <class T>
-void OMP45Stream<T>::add()
+void OMPStream<T>::add()
 {
+#ifdef OMP_TARGET_GPU
   unsigned int array_size = this->array_size;
   T *a = this->a;
   T *b = this->b;
   T *c = this->c;
   #pragma omp target teams distribute parallel for simd map(to: a[0:array_size], b[0:array_size], c[0:array_size])
+#else
+  #pragma omp parallel for
+#endif
   for (int i = 0; i < array_size; i++)
   {
     c[i] = a[i] + b[i];
@@ -97,22 +141,51 @@ void OMP45Stream<T>::add()
 }
 
 template <class T>
-void OMP45Stream<T>::triad()
+void OMPStream<T>::triad()
 {
   const T scalar = startScalar;
 
+#ifdef OMP_TARGET_GPU
   unsigned int array_size = this->array_size;
   T *a = this->a;
   T *b = this->b;
   T *c = this->c;
   #pragma omp target teams distribute parallel for simd map(to: a[0:array_size], b[0:array_size], c[0:array_size])
+#else
+  #pragma omp parallel for
+#endif
   for (int i = 0; i < array_size; i++)
   {
     a[i] = b[i] + scalar * c[i];
   }
 }
+
+template <class T>
+T OMPStream<T>::dot()
+{
+  T sum = 0.0;
+
+#ifdef OMP_TARGET_GPU
+  unsigned int array_size = this->array_size;
+  T *a = this->a;
+  T *b = this->b;
+  #pragma omp target teams distribute parallel for simd reduction(+:sum) map(tofrom: sum)
+#else
+  #pragma omp parallel for reduction(+:sum)
+#endif
+  for (int i = 0; i < array_size; i++)
+  {
+    sum += a[i] * b[i];
+  }
+
+  return sum;
+}
+
+
+
 void listDevices(void)
 {
+#ifdef OMP_TARGET_GPU
   // Get number of devices
   int count = omp_get_num_devices();
 
@@ -125,6 +198,9 @@ void listDevices(void)
   {
     std::cout << "There are " << count << " devices." << std::endl;
   }
+#else
+  std::cout << "0: CPU" << std::endl;
+#endif
 }
 
 std::string getDeviceName(const int)
@@ -136,5 +212,5 @@ std::string getDeviceDriver(const int)
 {
   return std::string("Device driver unavailable");
 }
-template class OMP45Stream<float>;
+template class OMPStream<float>;
-template class OMP45Stream<double>;
+template class OMPStream<double>;

OMPStream.h

@@ -14,10 +14,10 @@
 
 #include <omp.h>
 
-#define IMPLEMENTATION_STRING "OpenMP 4.5"
+#define IMPLEMENTATION_STRING "OpenMP"
 
 template <class T>
-class OMP45Stream : public Stream<T>
+class OMPStream : public Stream<T>
 {
   protected:
     // Size of arrays
@@ -29,15 +29,16 @@ class OMP45Stream : public Stream<T>
     T *c;
 
   public:
-    OMP45Stream(const unsigned int, T*, T*, T*, int);
+    OMPStream(const unsigned int, T*, T*, T*, int);
-    ~OMP45Stream();
+    ~OMPStream();
 
     virtual void copy() override;
     virtual void add() override;
     virtual void mul() override;
     virtual void triad() override;
+    virtual T dot() override;
 
-    virtual void write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
+    virtual void init_arrays(T initA, T initB, T initC) override;
     virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
 
 

RAJAStream.cpp

@@ -21,12 +21,6 @@ RAJAStream<T>::RAJAStream(const unsigned int ARRAY_SIZE, const int device_index)
   d_a = new T[ARRAY_SIZE];
   d_b = new T[ARRAY_SIZE];
   d_c = new T[ARRAY_SIZE];
-  forall<policy>(index_set, [=] RAJA_DEVICE (int index)
-  {
-    d_a[index] = 0.0;
-    d_b[index] = 0.0;
-    d_c[index] = 0.0;
-  });
 #else
   cudaMallocManaged((void**)&d_a, sizeof(T)*ARRAY_SIZE, cudaMemAttachGlobal);
   cudaMallocManaged((void**)&d_b, sizeof(T)*ARRAY_SIZE, cudaMemAttachGlobal);
@@ -50,12 +44,17 @@ RAJAStream<T>::~RAJAStream()
 }
 
 template <class T>
-void RAJAStream<T>::write_arrays(
+void RAJAStream<T>::init_arrays(T initA, T initB, T initC)
-        const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c)
 {
-  std::copy(a.begin(), a.end(), d_a);
+  T* a = d_a;
-  std::copy(b.begin(), b.end(), d_b);
+  T* b = d_b;
-  std::copy(c.begin(), c.end(), d_c);
+  T* c = d_c;
+  forall<policy>(index_set, [=] RAJA_DEVICE (int index)
+  {
+    a[index] = initA;
+    b[index] = initB;
+    c[index] = initC;
+  });
 }
 
 template <class T>
@@ -115,6 +114,23 @@ void RAJAStream<T>::triad()
   });
 }
 
+template <class T>
+T RAJAStream<T>::dot()
+{
+  T* a = d_a;
+  T* b = d_b;
+
+  RAJA::ReduceSum<reduce_policy, T> sum(0.0);
+
+  forall<policy>(index_set, [=] RAJA_DEVICE (int index)
+  {
+    sum += a[index] * b[index];
+  });
+
+  return T(sum);
+}
+
+
 void listDevices(void)
 {
   std::cout << "This is not the device you are looking for.";

RAJAStream.hpp

@@ -18,11 +18,13 @@
 typedef RAJA::IndexSet::ExecPolicy<
         RAJA::seq_segit,
         RAJA::omp_parallel_for_exec> policy;
+typedef RAJA::omp_reduce reduce_policy;
 #else
 const size_t block_size = 128;
 typedef RAJA::IndexSet::ExecPolicy<
         RAJA::seq_segit,
         RAJA::cuda_exec<block_size>> policy;
+typedef RAJA::cuda_reduce<block_size> reduce_policy;
 #endif
 
 template <class T>
@@ -49,9 +51,9 @@ class RAJAStream : public Stream<T>
     virtual void add() override;
     virtual void mul() override;
     virtual void triad() override;
+    virtual T dot() override;
 
-    virtual void write_arrays(
+    virtual void init_arrays(T initA, T initB, T initC) override;
-            const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
     virtual void read_arrays(
             std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
 };

SYCLStream.cpp

@@ -18,14 +18,6 @@ std::vector<device> devices;
 void getDeviceList(void);
 program * p;
 
-/* Forward declaration of SYCL kernels */
-namespace kernels {
-  class copy;
-  class mul;
-  class add;
-  class triad;
-}
-
 template <class T>
 SYCLStream<T>::SYCLStream(const unsigned int ARRAY_SIZE, const int device_index)
 {
@@ -38,24 +30,39 @@ SYCLStream<T>::SYCLStream(const unsigned int ARRAY_SIZE, const int device_index)
     throw std::runtime_error("Invalid device index");
   device dev = devices[device_index];
 
+  // Determine sensible dot kernel NDRange configuration
+  if (dev.is_cpu())
+  {
+    dot_num_groups = dev.get_info<info::device::max_compute_units>();
+    dot_wgsize     = dev.get_info<info::device::native_vector_width_double>() * 2;
+  }
+  else
+  {
+    dot_num_groups = dev.get_info<info::device::max_compute_units>() * 4;
+    dot_wgsize     = dev.get_info<info::device::max_work_group_size>();
+  }
+
   // Print out device information
   std::cout << "Using SYCL device " << getDeviceName(device_index) << std::endl;
   std::cout << "Driver: " << getDeviceDriver(device_index) << std::endl;
+  std::cout << "Reduction kernel config: " << dot_num_groups << " groups of size " << dot_wgsize << std::endl;
 
   queue = new cl::sycl::queue(dev);
 
   /* Pre-build the kernels */
   p = new program(queue->get_context());
-  p->build_from_kernel_name<kernels::copy>();
+  p->build_from_kernel_name<init_kernel>();
-  p->build_from_kernel_name<kernels::mul>();
+  p->build_from_kernel_name<copy_kernel>();
-  p->build_from_kernel_name<kernels::add>();
+  p->build_from_kernel_name<mul_kernel>();
-  p->build_from_kernel_name<kernels::triad>();
+  p->build_from_kernel_name<add_kernel>();
-
+  p->build_from_kernel_name<triad_kernel>();
+  p->build_from_kernel_name<dot_kernel>();
 
   // Create buffers
   d_a = new buffer<T>(array_size);
   d_b = new buffer<T>(array_size);
   d_c = new buffer<T>(array_size);
+  d_sum = new buffer<T>(dot_num_groups);
 }
 
 template <class T>
@@ -64,6 +71,7 @@ SYCLStream<T>::~SYCLStream()
   delete d_a;
   delete d_b;
   delete d_c;
+  delete d_sum;
 
   delete p;
   delete queue;
@@ -76,11 +84,11 @@ void SYCLStream<T>::copy()
   {
     auto ka = d_a->template get_access<access::mode::read>(cgh);
     auto kc = d_c->template get_access<access::mode::write>(cgh);
-    cgh.parallel_for<kernels::copy>(p->get_kernel<kernels::copy>(),
+    cgh.parallel_for<copy_kernel>(p->get_kernel<copy_kernel>(),
           range<1>{array_size}, [=](item<1> item)
     {
-      auto id = item.get();
+      auto id = item.get()[0];
-      kc[id[0]] = ka[id[0]];
+      kc[id] = ka[id];
     });
   });
   queue->wait();
@@ -94,11 +102,11 @@ void SYCLStream<T>::mul()
   {
     auto kb = d_b->template get_access<access::mode::write>(cgh);
     auto kc = d_c->template get_access<access::mode::read>(cgh);
-    cgh.parallel_for<kernels::mul>(p->get_kernel<kernels::mul>(),
+    cgh.parallel_for<mul_kernel>(p->get_kernel<mul_kernel>(),
       range<1>{array_size}, [=](item<1> item)
     {
-      auto id = item.get();
+      auto id = item.get()[0];
-      kb[id[0]] = scalar * kc[id[0]];
+      kb[id] = scalar * kc[id];
     });
   });
   queue->wait();
@@ -112,11 +120,11 @@ void SYCLStream<T>::add()
     auto ka = d_a->template get_access<access::mode::read>(cgh);
     auto kb = d_b->template get_access<access::mode::read>(cgh);
     auto kc = d_c->template get_access<access::mode::write>(cgh);
-    cgh.parallel_for<kernels::add>(p->get_kernel<kernels::add>(),
+    cgh.parallel_for<add_kernel>(p->get_kernel<add_kernel>(),
       range<1>{array_size}, [=](item<1> item)
     {
-      auto id = item.get();
+      auto id = item.get()[0];
-      kc[id[0]] = ka[id[0]] + kb[id[0]];
+      kc[id] = ka[id] + kb[id];
     });
   });
   queue->wait();
@@ -131,28 +139,81 @@ void SYCLStream<T>::triad()
     auto ka = d_a->template get_access<access::mode::write>(cgh);
     auto kb = d_b->template get_access<access::mode::read>(cgh);
     auto kc = d_c->template get_access<access::mode::read>(cgh);
-    cgh.parallel_for<kernels::triad>(p->get_kernel<kernels::triad>(),
+    cgh.parallel_for<triad_kernel>(p->get_kernel<triad_kernel>(),
       range<1>{array_size}, [=](item<1> item)
     {
-      auto id = item.get();
+      auto id = item.get()[0];
-      ka[id] = kb[id[0]] + scalar * kc[id[0]];
+      ka[id] = kb[id] + scalar * kc[id];
     });
   });
   queue->wait();
 }
 
 template <class T>
-void SYCLStream<T>::write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c)
+T SYCLStream<T>::dot()
 {
-  auto _a = d_a->template get_access<access::mode::write, access::target::host_buffer>();
+  queue->submit([&](handler &cgh)
-  auto _b = d_b->template get_access<access::mode::write, access::target::host_buffer>();
-  auto _c = d_c->template get_access<access::mode::write, access::target::host_buffer>();
-  for (int i = 0; i < array_size; i++)
   {
-    _a[i] = a[i];
+    auto ka   = d_a->template get_access<access::mode::read>(cgh);
-    _b[i] = b[i];
+    auto kb   = d_b->template get_access<access::mode::read>(cgh);
-    _c[i] = c[i];
+    auto ksum = d_sum->template get_access<access::mode::write>(cgh);
+
+    auto wg_sum = accessor<T, 1, access::mode::read_write, access::target::local>(range<1>(dot_wgsize), cgh);
+
+    size_t N = array_size;
+
+    cgh.parallel_for<dot_kernel>(p->get_kernel<dot_kernel>(),
+      nd_range<1>(dot_num_groups*dot_wgsize, dot_wgsize), [=](nd_item<1> item)
+    {
+      size_t i = item.get_global(0);
+      size_t li = item.get_local(0);
+      size_t global_size = item.get_global_range()[0];
+
+      wg_sum[li] = 0.0;
+      for (; i < N; i += global_size)
+        wg_sum[li] += ka[i] * kb[i];
+
+      size_t local_size = item.get_local_range()[0];
+      for (int offset = local_size / 2; offset > 0; offset /= 2)
+      {
+        item.barrier(cl::sycl::access::fence_space::local_space);
+        if (li < offset)
+          wg_sum[li] += wg_sum[li + offset];
       }
+
+      if (li == 0)
+        ksum[item.get_group(0)] = wg_sum[0];
+    });
+  });
+
+  T sum = 0.0;
+  auto h_sum = d_sum->template get_access<access::mode::read, access::target::host_buffer>();
+  for (int i = 0; i < dot_num_groups; i++)
+  {
+    sum += h_sum[i];
+  }
+
+  return sum;
+}
+
+template <class T>
+void SYCLStream<T>::init_arrays(T initA, T initB, T initC)
+{
+  queue->submit([&](handler &cgh)
+  {
+    auto ka = d_a->template get_access<access::mode::write>(cgh);
+    auto kb = d_b->template get_access<access::mode::write>(cgh);
+    auto kc = d_c->template get_access<access::mode::write>(cgh);
+    cgh.parallel_for<init_kernel>(p->get_kernel<init_kernel>(),
+      range<1>{array_size}, [=](item<1> item)
+    {
+      auto id = item.get()[0];
+      ka[id] = initA;
+      kb[id] = initB;
+      kc[id] = initC;
+    });
+  });
+  queue->wait();
 }
 
 template <class T>
@@ -244,5 +305,5 @@ std::string getDeviceDriver(const int device)
 
 
 // TODO: Fix kernel names to allow multiple template specializations
-//template class SYCLStream<float>;
+template class SYCLStream<float>;
 template class SYCLStream<double>;

SYCLStream.h

@@ -15,6 +15,16 @@
 
 #define IMPLEMENTATION_STRING "SYCL"
 
+namespace sycl_kernels
+{
+  template <class T> class init;
+  template <class T> class copy;
+  template <class T> class mul;
+  template <class T> class add;
+  template <class T> class triad;
+  template <class T> class dot;
+}
+
 template <class T>
 class SYCLStream : public Stream<T>
 {
@@ -27,6 +37,19 @@ class SYCLStream : public Stream<T>
     cl::sycl::buffer<T> *d_a;
     cl::sycl::buffer<T> *d_b;
     cl::sycl::buffer<T> *d_c;
+    cl::sycl::buffer<T> *d_sum;
+
+    // SYCL kernel names
+    typedef sycl_kernels::init<T> init_kernel;
+    typedef sycl_kernels::copy<T> copy_kernel;
+    typedef sycl_kernels::mul<T> mul_kernel;
+    typedef sycl_kernels::add<T> add_kernel;
+    typedef sycl_kernels::triad<T> triad_kernel;
+    typedef sycl_kernels::dot<T> dot_kernel;
+
+    // NDRange configuration for the dot kernel
+    size_t dot_num_groups;
+    size_t dot_wgsize;
 
   public:
 
@@ -37,8 +60,9 @@ class SYCLStream : public Stream<T>
     virtual void add() override;
     virtual void mul() override;
     virtual void triad() override;
+    virtual T    dot() override;
 
-    virtual void write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
+    virtual void init_arrays(T initA, T initB, T initC) override;
     virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
 
 };

Stream.h

@@ -29,9 +29,10 @@ class Stream
     virtual void mul() = 0;
     virtual void add() = 0;
     virtual void triad() = 0;
+    virtual T dot() = 0;
 
     // Copy memory between host and device
-    virtual void write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) = 0;
+    virtual void init_arrays(T initA, T initB, T initC) = 0;
     virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) = 0;
 
 };

common.h.in

@@ -1,9 +0,0 @@
-
-// Copyright (c) 2015-16 Tom Deakin, Simon McIntosh-Smith,
-// University of Bristol HPC
-//
-// For full license terms please see the LICENSE file distributed with this
-// source code
-
-#define VERSION_STRING "@gpu-stream_VERSION_MAJOR@.@gpu-stream_VERSION_MINOR@"
-

main.cpp

@@ -15,7 +15,8 @@
 #include <iomanip>
 #include <cstring>
 
-#include "common.h"
+#define VERSION_STRING "devel"
+
 #include "Stream.h"
 
 #if defined(CUDA)
@@ -32,10 +33,8 @@
 #include "ACCStream.h"
 #elif defined(SYCL)
 #include "SYCLStream.h"
-#elif defined(OMP3)
+#elif defined(OMP)
-#include "OMP3Stream.h"
+#include "OMPStream.h"
-#elif defined(OMP45)
-#include "OMP45Stream.h"
 #endif
 
 // Default size of 2^25
@@ -45,7 +44,7 @@ unsigned int deviceIndex = 0;
 bool use_float = false;
 
 template <typename T>
-void check_solution(const unsigned int ntimes, std::vector<T>& a, std::vector<T>& b, std::vector<T>& c);
+void check_solution(const unsigned int ntimes, std::vector<T>& a, std::vector<T>& b, std::vector<T>& c, T& sum);
 
 template <typename T>
 void run();
@@ -61,13 +60,11 @@ int main(int argc, char *argv[])
 
   parseArguments(argc, argv);
 
-  // TODO: Fix SYCL to allow multiple template specializations
+  // TODO: Fix Kokkos to allow multiple template specializations
-#ifndef SYCL
 #ifndef KOKKOS
   if (use_float)
     run<float>();
   else
-#endif
 #endif
     run<double>();
 
@@ -84,9 +81,9 @@ void run()
     std::cout << "Precision: double" << std::endl;
 
   // Create host vectors
-  std::vector<T> a(ARRAY_SIZE, startA);
+  std::vector<T> a(ARRAY_SIZE);
-  std::vector<T> b(ARRAY_SIZE, startB);
+  std::vector<T> b(ARRAY_SIZE);
-  std::vector<T> c(ARRAY_SIZE, startC);
+  std::vector<T> c(ARRAY_SIZE);
   std::streamsize ss = std::cout.precision();
   std::cout << std::setprecision(1) << std::fixed
     << "Array size: " << ARRAY_SIZE*sizeof(T)*1.0E-6 << " MB"
@@ -95,6 +92,9 @@ void run()
     << " (=" << 3.0*ARRAY_SIZE*sizeof(T)*1.0E-9 << " GB)" << std::endl;
   std::cout.precision(ss);
 
+  // Result of the Dot kernel
+  T sum;
+
   Stream<T> *stream;
 
 #if defined(CUDA)
@@ -125,20 +125,16 @@ void run()
   // Use the SYCL implementation
   stream = new SYCLStream<T>(ARRAY_SIZE, deviceIndex);
 
-#elif defined(OMP3)
+#elif defined(OMP)
-  // Use the "reference" OpenMP 3 implementation
+  // Use the OpenMP implementation
-  stream = new OMP3Stream<T>(ARRAY_SIZE, a.data(), b.data(), c.data());
+  stream = new OMPStream<T>(ARRAY_SIZE, a.data(), b.data(), c.data(), deviceIndex);
-
-#elif defined(OMP45)
-  // Use the "reference" OpenMP 3 implementation
-  stream = new OMP45Stream<T>(ARRAY_SIZE, a.data(), b.data(), c.data(), deviceIndex);
 
 #endif
 
-  stream->write_arrays(a, b, c);
+  stream->init_arrays(startA, startB, startC);
 
   // List of times
-  std::vector<std::vector<double>> timings(4);
+  std::vector<std::vector<double>> timings(5);
 
   // Declare timers
   std::chrono::high_resolution_clock::time_point t1, t2;
@@ -170,11 +166,17 @@ void run()
     t2 = std::chrono::high_resolution_clock::now();
     timings[3].push_back(std::chrono::duration_cast<std::chrono::duration<double> >(t2 - t1).count());
 
+    // Execute Dot
+    t1 = std::chrono::high_resolution_clock::now();
+    sum = stream->dot();
+    t2 = std::chrono::high_resolution_clock::now();
+    timings[4].push_back(std::chrono::duration_cast<std::chrono::duration<double> >(t2 - t1).count());
+
   }
 
   // Check solutions
   stream->read_arrays(a, b, c);
-  check_solution<T>(num_times, a, b, c);
+  check_solution<T>(num_times, a, b, c, sum);
 
   // Display timing results
   std::cout
@@ -186,15 +188,16 @@ void run()
 
   std::cout << std::fixed;
 
-  std::string labels[4] = {"Copy", "Mul", "Add", "Triad"};
+  std::string labels[5] = {"Copy", "Mul", "Add", "Triad", "Dot"};
-  size_t sizes[4] = {
+  size_t sizes[5] = {
     2 * sizeof(T) * ARRAY_SIZE,
     2 * sizeof(T) * ARRAY_SIZE,
     3 * sizeof(T) * ARRAY_SIZE,
-    3 * sizeof(T) * ARRAY_SIZE
+    3 * sizeof(T) * ARRAY_SIZE,
+    2 * sizeof(T) * ARRAY_SIZE
   };
 
-  for (int i = 0; i < 4; i++)
+  for (int i = 0; i < 5; i++)
   {
     // Get min/max; ignore the first result
     auto minmax = std::minmax_element(timings[i].begin()+1, timings[i].end());
@@ -218,12 +221,13 @@ void run()
 }
 
 template <typename T>
-void check_solution(const unsigned int ntimes, std::vector<T>& a, std::vector<T>& b, std::vector<T>& c)
+void check_solution(const unsigned int ntimes, std::vector<T>& a, std::vector<T>& b, std::vector<T>& c, T& sum)
 {
   // Generate correct solution
   T goldA = startA;
   T goldB = startB;
   T goldC = startC;
+  T goldSum = 0.0;
 
   const T scalar = startScalar;
 
@@ -236,6 +240,9 @@ void check_solution(const unsigned int ntimes, std::vector<T>& a, std::vector<T>
     goldA = goldB + scalar * goldC;
   }
 
+  // Do the reduction
+  goldSum = goldA * goldB * ARRAY_SIZE;
+
   // Calculate the average error
   double errA = std::accumulate(a.begin(), a.end(), 0.0, [&](double sum, const T val){ return sum + fabs(val - goldA); });
   errA /= a.size();
@@ -243,6 +250,7 @@ void check_solution(const unsigned int ntimes, std::vector<T>& a, std::vector<T>
   errB /= b.size();
   double errC = std::accumulate(c.begin(), c.end(), 0.0, [&](double sum, const T val){ return sum + fabs(val - goldC); });
   errC /= c.size();
+  double errSum = fabs(sum - goldSum);
 
   double epsi = std::numeric_limits<T>::epsilon() * 100.0;
 
@@ -258,6 +266,13 @@ void check_solution(const unsigned int ntimes, std::vector<T>& a, std::vector<T>
     std::cerr
       << "Validation failed on c[]. Average error " << errC
       << std::endl;
+  // Check sum to 8 decimal places
+  if (errSum > 1.0E-8)
+    std::cerr
+      << "Validation failed on sum. Error " << errSum
+      << std::endl << std::setprecision(15)
+      << "Sum was " << sum << " but should be " << goldSum
+      << std::endl;
 
 }