From 9ff46ec2be7e9adb5be82f214a348b1cafc7c812 Mon Sep 17 00:00:00 2001
From: Gonzalo Brito Gadeschi <gonzalob@nvidia.com>
Date: Wed, 5 Jun 2024 01:51:51 -0700
Subject: [PATCH] Update Serial

---
 src/StreamModels.h          |  2 +-
 src/main.cpp                | 15 ++++++++-------
 src/serial/SerialStream.cpp | 20 +++++++++-----------
 src/serial/SerialStream.h   | 23 +++++++++++------------
 4 files changed, 29 insertions(+), 31 deletions(-)

diff --git a/src/StreamModels.h b/src/StreamModels.h
index b803b0f1..b13d5b34 100644
--- a/src/StreamModels.h
+++ b/src/StreamModels.h
@@ -97,7 +97,7 @@ std::unique_ptr<Stream<T>> make_stream(Args... args) {
 
 #elif defined(SERIAL)
   // Use the Serial implementation
-  return std::make_unique<SerialStream<T>>(array_size, deviceIndex);
+  return std::make_unique<SerialStream<T>>(args...);
 
 #elif defined(FUTHARK)
   // Use the Futhark implementation
diff --git a/src/main.cpp b/src/main.cpp
index 55fd6f43..604e14a2 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -328,10 +328,11 @@ void check_solution(const size_t num_times, T const* a, T const* b, T const* c,
     abort();
   }
 
-  // Error relative tolerance check
+  // Error relative tolerance check - a higher tolerance is used for reductions.
   size_t failed = 0;
-  T max_rel = std::numeric_limits<T>::epsilon() * T(1000000.0);
-  auto check = [&](const char* name, T is, T should, size_t i = size_t(-1)) {
+  T max_rel = std::numeric_limits<T>::epsilon() * T(100.0);
+  T max_rel_dot = std::numeric_limits<T>::epsilon() * T(10000000.0);
+  auto check = [&](const char* name, T is, T should, T max_rel, size_t i = size_t(-1)) {
     // Relative difference:
     T diff = std::abs(is - should);
     T abs_is = std::abs(is);
@@ -353,15 +354,15 @@ void check_solution(const size_t num_times, T const* a, T const* b, T const* c,
   for (size_t i = 0; i < num_benchmarks; ++i) {
     if (bench[i].id != BenchId::Dot) continue;
     if (run_benchmark(bench[i]))
-      check("sum", sum,  goldS);
+      check("sum", sum, goldS, max_rel_dot);
     break;
   }
 
   // Calculate the L^infty-norm relative error
   for (size_t i = 0; i < array_size; ++i) {
-    check("a", a[i], goldA, i);
-    check("b", b[i], goldB, i);
-    check("c", c[i], goldC, i);
+    check("a", a[i], goldA, i, max_rel);
+    check("b", b[i], goldB, i, max_rel);
+    check("c", c[i], goldC, i, max_rel);
   }
 
   if (failed > 0 && !silence_errors)
diff --git a/src/serial/SerialStream.cpp b/src/serial/SerialStream.cpp
index 0fc0fdb1..394a657c 100644
--- a/src/serial/SerialStream.cpp
+++ b/src/serial/SerialStream.cpp
@@ -13,14 +13,16 @@
 #endif
 
 template <class T>
-SerialStream<T>::SerialStream(const intptr_t ARRAY_SIZE, int device)
+SerialStream<T>::SerialStream(BenchId bs, const intptr_t array_size, const int device_id,
+			      T initA, T initB, T initC)
+  : array_size{array_size}
 {
-  array_size = ARRAY_SIZE;
-
   // Allocate on the host
   this->a = (T*)aligned_alloc(ALIGNMENT, sizeof(T)*array_size);
   this->b = (T*)aligned_alloc(ALIGNMENT, sizeof(T)*array_size);
   this->c = (T*)aligned_alloc(ALIGNMENT, sizeof(T)*array_size);
+
+  init_arrays(initA, initB, initC);
 }
 
 template <class T>
@@ -44,15 +46,11 @@ void SerialStream<T>::init_arrays(T initA, T initB, T initC)
 }
 
 template <class T>
-void SerialStream<T>::read_arrays(std::vector<T>& h_a, std::vector<T>& h_b, std::vector<T>& h_c)
+void SerialStream<T>::get_arrays(T const*& h_a, T const*& h_b, T const*& h_c)
 {
-  for (intptr_t i = 0; i < array_size; i++)
-  {
-    h_a[i] = a[i];
-    h_b[i] = b[i];
-    h_c[i] = c[i];
-  }
-
+  h_a = a;
+  h_b = b;
+  h_c = c;
 }
 
 template <class T>
diff --git a/src/serial/SerialStream.h b/src/serial/SerialStream.h
index 3c62d296..9f5653d4 100644
--- a/src/serial/SerialStream.h
+++ b/src/serial/SerialStream.h
@@ -23,21 +23,20 @@ class SerialStream : public Stream<T>
     intptr_t array_size;
 
     // Device side pointers
-    T *a;
-    T *b;
-    T *c;
+    T *a, *b, *c;
 
   public:
-    SerialStream(const intptr_t, int);
+    SerialStream(BenchId bs, const intptr_t array_size, const int device_id,
+		 T initA, T initB, T initC);
     ~SerialStream();
 
-    virtual void copy() override;
-    virtual void add() override;
-    virtual void mul() override;
-    virtual void triad() override;
-    virtual void nstream() override;
-    virtual T dot() override;
+    void copy() override;
+    void add() override;
+    void mul() override;
+    void triad() override;
+    void nstream() override;
+    T dot() 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;
+    void get_arrays(T const*& a, T const*& b, T const*& c) override;
+    void init_arrays(T initA, T initB, T initC);
 };