From 847cdc9b5e2cafcd31689f5a9a3a9becdfdcb07c Mon Sep 17 00:00:00 2001
From: Tom Deakin <tom.deakin@bristol.ac.uk>
Date: Thu, 16 Jul 2015 12:04:07 +0100
Subject: [PATCH] Time the runs and print out results

---
 ocl-stream.cpp | 115 +++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 111 insertions(+), 4 deletions(-)

diff --git a/ocl-stream.cpp b/ocl-stream.cpp
index 91021ca..9cdca91 100644
--- a/ocl-stream.cpp
+++ b/ocl-stream.cpp
@@ -2,12 +2,19 @@
 #include <iostream>
 #include <fstream>
 #include <vector>
+#include <chrono>
+#include <cfloat>
+#include <iomanip>
 
 #define __CL_ENABLE_EXCEPTIONS
 #include "cl.hpp"
 
 #define DATATYPE double
-#define ARRAY_SIZE 1000000
+#define ARRAY_SIZE 50000000
+#define NTIMES 2
+
+#define MIN(a,b) ((a) < (b)) ? (a) : (b)
+#define MAX(a,b) ((a) > (b)) ? (a) : (b)
 
 struct badfile : public std::exception
 {
@@ -17,6 +24,12 @@ struct badfile : public std::exception
   }
 };
 
+size_t sizes[4] = {
+	2 * sizeof(DATATYPE) * ARRAY_SIZE,
+	2 * sizeof(DATATYPE) * ARRAY_SIZE,
+	3 * sizeof(DATATYPE) * ARRAY_SIZE,
+	3 * sizeof(DATATYPE) * ARRAY_SIZE
+};
 
 
 int main(void)
@@ -29,7 +42,7 @@ int main(void)
 		std::string kernels(std::istreambuf_iterator<char>(in), (std::istreambuf_iterator<char>()));
 
 		// Setup OpenCL
-		cl::Context context(CL_DEVICE_TYPE_GPU);
+		cl::Context context(CL_DEVICE_TYPE_CPU);
 		cl::CommandQueue queue(context);
 		cl::Program program(context, kernels);
 
@@ -50,8 +63,8 @@ int main(void)
 
 		cl::make_kernel<cl::Buffer, cl::Buffer> copy(program, "copy");
 		cl::make_kernel<cl::Buffer, cl::Buffer> mul(program, "mul");
-		cl::make_kernel<cl::Buffer, cl::Buffer> add(program, "add");
-		cl::make_kernel<cl::Buffer, cl::Buffer> triad(program, "triad");
+		cl::make_kernel<cl::Buffer, cl::Buffer, cl::Buffer> add(program, "add");
+		cl::make_kernel<cl::Buffer, cl::Buffer, cl::Buffer> triad(program, "triad");
 
 		// Create host vectors
 		std::vector<DATATYPE> h_a(ARRAY_SIZE, 1.0);
@@ -71,6 +84,100 @@ int main(void)
 		// Make sure the copies are finished
 		queue.finish();
 
+		// List of times
+		std::vector< std::vector<double> > timings;
+
+		// Declare timers
+		std::chrono::high_resolution_clock::time_point t1, t2;
+
+		// Main loop
+		for (unsigned int k = 0; k < NTIMES; k++)
+		{
+			std::vector<double> times;
+			t1 = std::chrono::high_resolution_clock::now();
+			copy(
+				cl::EnqueueArgs(
+                queue,
+                cl::NDRange(ARRAY_SIZE)),
+                d_a, d_c);
+			queue.finish();
+			t2 = std::chrono::high_resolution_clock::now();
+			times.push_back(std::chrono::duration_cast<std::chrono::duration<double> >(t2 - t1).count());
+
+
+			t1 = std::chrono::high_resolution_clock::now();
+			mul(
+				cl::EnqueueArgs(
+                queue,
+                cl::NDRange(ARRAY_SIZE)),
+                d_b, d_c);
+			queue.finish();
+			t2 = std::chrono::high_resolution_clock::now();
+			times.push_back(std::chrono::duration_cast<std::chrono::duration<double> >(t2 - t1).count());
+
+
+			t1 = std::chrono::high_resolution_clock::now();
+			add(
+				cl::EnqueueArgs(
+                queue,
+                cl::NDRange(ARRAY_SIZE)),
+                d_a, d_b, d_c);
+			queue.finish();
+			t2 = std::chrono::high_resolution_clock::now();
+			times.push_back(std::chrono::duration_cast<std::chrono::duration<double> >(t2 - t1).count());
+
+
+			t1 = std::chrono::high_resolution_clock::now();
+			triad(
+				cl::EnqueueArgs(
+                queue,
+                cl::NDRange(ARRAY_SIZE)),
+                d_a, d_b, d_c);
+			queue.finish();
+			t2 = std::chrono::high_resolution_clock::now();
+			times.push_back(std::chrono::duration_cast<std::chrono::duration<double> >(t2 - t1).count());
+
+			timings.push_back(times);
+
+		}
+
+		// Crunch results
+		double min[4] = {DBL_MAX, DBL_MAX, DBL_MAX, DBL_MAX};
+		double max[4] = {0.0, 0.0, 0.0, 0.0};
+		double avg[4] = {0.0, 0.0, 0.0, 0.0};
+		// Ignore first result
+		for (unsigned int i = 1; i < NTIMES; i++)
+		{
+			for (int j = 0; j < 4; j++)
+			{
+				avg[j] += timings[i][j];
+				min[j] = MIN(min[j], timings[i][j]);
+				max[j] = MAX(max[j], timings[i][j]);
+			}
+		}
+		for (int j = 0; j < 4; j++)
+			avg[j] /= (double)NTIMES;
+
+		// Display results
+		std::string labels[] = {"Copy", "Mul", "Add", "Triad"};
+		std::cout
+			<< std::left << std::setw(12) << "Function"
+			<< std::left << std::setw(12) << "MBytes/sec"
+			<< std::left << std::setw(12) << "Min (sec)"
+			<< std::left << std::setw(12) << "Max"
+			<< std::left << std::setw(12) << "Average"
+			<< std::endl;
+		for (int j = 0; j < 4; j++)
+		{
+			std::cout
+				<< std::left << std::setw(12) << labels[j]
+				<< std::left << std::setw(12) << 1.0E-06 * sizes[j]/min[j]
+				<< std::left << std::setw(12) << min[j]
+				<< std::left << std::setw(12) << max[j]
+				<< std::left << std::setw(12) << avg[j]
+				<< std::endl;
+		}
+
 	}
 	// Catch OpenCL Errors and display information
 	catch (cl::Error& e)