From 1e94a41f3c9f9abbb5c7666cfe6381c21eb37e5e Mon Sep 17 00:00:00 2001
From: Tom Deakin <tom.deakin@bristol.ac.uk>
Date: Mon, 1 Feb 2021 17:41:30 +0000
Subject: [PATCH] Add nstream kernel from PRK

PRK has a nstream kernel, which is Triad with a += update.
This means there are 3 reads and a write, which is a higher
read/write ratio. In addition, non-temporal stores for the
write on CPUs will not be beneficial, and so compilers should
take care to emit these for the other kernels, but not these.
---
 Stream.h |  1 +
 main.cpp | 21 ++++++++++++++++-----
 2 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/Stream.h b/Stream.h
index ff00a54..eb4ffd4 100644
--- a/Stream.h
+++ b/Stream.h
@@ -29,6 +29,7 @@ class Stream
     virtual void mul() = 0;
     virtual void add() = 0;
     virtual void triad() = 0;
+    virtual void nstream() = 0;
     virtual T dot() = 0;
 
     // Copy memory between host and device
diff --git a/main.cpp b/main.cpp
index fd64546..5b931f7 100644
--- a/main.cpp
+++ b/main.cpp
@@ -186,7 +186,7 @@ void run()
   T sum;
 
   // List of times
-  std::vector<std::vector<double>> timings(5);
+  std::vector<std::vector<double>> timings(6);
 
   // Declare timers
   std::chrono::high_resolution_clock::time_point t1, t2;
@@ -218,11 +218,17 @@ 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 nstream
+    t1 = std::chrono::high_resolution_clock::now();
+    stream->nstream();
+    t2 = std::chrono::high_resolution_clock::now();
+    timings[4].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());
+    timings[5].push_back(std::chrono::duration_cast<std::chrono::duration<double> >(t2 - t1).count());
 
   }
 
@@ -262,16 +268,17 @@ void run()
 
 
 
-  std::string labels[5] = {"Copy", "Mul", "Add", "Triad", "Dot"};
-  size_t sizes[5] = {
+  std::string labels[6] = {"Copy", "Mul", "Add", "Triad", "nstream", "Dot"};
+  size_t sizes[6] = {
     2 * sizeof(T) * ARRAY_SIZE,
     2 * sizeof(T) * ARRAY_SIZE,
     3 * sizeof(T) * ARRAY_SIZE,
     3 * sizeof(T) * ARRAY_SIZE,
+    4 * sizeof(T) * ARRAY_SIZE,
     2 * sizeof(T) * ARRAY_SIZE
   };
 
-  for (int i = 0; i < 5; i++)
+  for (int i = 0; i < 6; i++)
   {
     // Get min/max; ignore the first result
     auto minmax = std::minmax_element(timings[i].begin()+1, timings[i].end());
@@ -473,6 +480,10 @@ void check_solution(const unsigned int ntimes, std::vector<T>& a, std::vector<T>
       goldC = goldA + goldB;
     }
     goldA = goldB + scalar * goldC;
+    if (!triad_only)
+    {
+      goldA += goldB + scalar * goldC;
+    }
   }
 
   // Do the reduction