add unroll test

04/8_benchmark/01/CMakeLists.txt

@@ -8,7 +8,11 @@ project(testbench LANGUAGES CXX)
 add_executable(testbench
     main.cpp
     aos.cpp
+    aos_aligned.cpp
     soa.cpp
+    soa_size_t.cpp
+    soa_unroll.cpp
+    aosoa.cpp
     )
 
 find_package(OpenMP REQUIRED)

04/8_benchmark/01/aos.cpp

@@ -1,4 +1,3 @@
-#include <cstdint>
 #include "common.h"
 
 namespace aos {
@@ -12,7 +11,7 @@ struct Point {
 Point ps[N];
 
 void compute() {
-    for (std::size_t i = 0; i < N; i++) {
+    for (int i = 0; i < N; i++) {
         ps[i].x = ps[i].x + ps[i].y + ps[i].z;
     }
 }

04/8_benchmark/01/aos_aligned.cpp

@@ -0,0 +1,20 @@
+#include "common.h"
+
+namespace aos_aligned {
+
+struct Point {
+    float x;
+    float y;
+    float z;
+    char padding[4];
+};
+
+Point ps[N];
+
+void compute() {
+    for (int i = 0; i < N; i++) {
+        ps[i].x = ps[i].x + ps[i].y + ps[i].z;
+    }
+}
+
+}

04/8_benchmark/01/aosoa.cpp

@@ -0,0 +1,21 @@
+#include "common.h"
+
+namespace aosoa {
+
+struct Point {
+    float x[M];
+    float y[M];
+    float z[M];
+};
+
+Point ps[N / M];
+
+void compute() {
+    for (int i = 0; i < N / M; i++) {
+        for (int j = 0; j < M; j++) {
+            ps[i].x[j] = ps[i].x[j] + ps[i].y[j] + ps[i].z[j];
+        }
+    }
+}
+
+}

04/8_benchmark/01/common.h

@@ -1,3 +1,4 @@
 #pragma once
 
-#define N (1<<20)
+#define N (1<<16)
+#define M (1<<4)

04/8_benchmark/01/main.cpp

@@ -4,15 +4,43 @@ namespace aos {
 void compute();
 }
 
+namespace aos_aligned {
+void compute();
+}
+
 namespace soa {
 void compute();
 }
 
+namespace soa_size_t {
+void compute();
+}
+
+namespace soa_unroll {
+void compute();
+}
+
+namespace aosoa {
+void compute();
+}
+
 int main() {
     profile(1000, "aos", [&] {
         aos::compute();
     });
+    profile(1000, "aos_aligned", [&] {
+        aos_aligned::compute();
+    });
     profile(1000, "soa", [&] {
         soa::compute();
     });
+    profile(1000, "soa_size_t", [&] {
+        soa_size_t::compute();
+    });
+    profile(1000, "soa_unroll", [&] {
+        soa_unroll::compute();
+    });
+    profile(1000, "aosoa", [&] {
+        aosoa::compute();
+    });
 }

04/8_benchmark/01/profile.h

@@ -10,6 +10,6 @@ static inline void profile(int times, Name const &name, Func const &func) {
         func();
     }
     auto t1 = std::chrono::steady_clock::now();
-    long dt = std::chrono::duration_cast<std::chrono::microseconds>(t1 - t0).count() / times;
-    std::cout << name << ": " << dt << " us" << std::endl;
+    long dt = std::chrono::duration_cast<std::chrono::nanoseconds>(t1 - t0).count() / times;
+    std::cout << name << ": " << dt << " ns" << std::endl;
 }

04/8_benchmark/01/soa.cpp

@@ -1,4 +1,3 @@
-#include <cstdint>
 #include "common.h"
 
 namespace soa {
@@ -12,7 +11,7 @@ struct Point {
 Point ps;
 
 void compute() {
-    for (std::size_t i = 0; i < N; i++) {
+    for (int i = 0; i < N; i++) {
         ps.x[i] = ps.x[i] + ps.y[i] + ps.z[i];
     }
 }

04/8_benchmark/01/soa_size_t.cpp

@@ -0,0 +1,20 @@
+#include <cstdint>
+#include "common.h"
+
+namespace soa_size_t {
+
+struct Point {
+    float x[N];
+    float y[N];
+    float z[N];
+};
+
+Point ps;
+
+void compute() {
+    for (std::size_t i = 0; i < N; i++) {
+        ps.x[i] = ps.x[i] + ps.y[i] + ps.z[i];
+    }
+}
+
+}

04/8_benchmark/01/soa_unroll.cpp

@@ -0,0 +1,24 @@
+#include "common.h"
+
+namespace soa_unroll {
+
+struct Point {
+    float x[N];
+    float y[N];
+    float z[N];
+};
+
+Point ps;
+
+void compute() {
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC unroll 32
+#elif defined(_MSC_VER)
+#pragma unroll 32
+#endif
+    for (int i = 0; i < N; i++) {
+        ps.x[i] = ps.x[i] + ps.y[i] + ps.z[i];
+    }
+}
+
+}