Fix C++ benchmark.

Author: calder

.gitignore

@@ -1,3 +1,2 @@
-*.bin
 *.out
 perf.data*

Makefile

@@ -1,14 +1,20 @@
+.PHONY: gbench-cc
+gbench-cc: gbench.cc
+	clang++ gbench.cc -O2 -march=haswell -lbenchmark
+
 .PHONY: naive-cc
 naive-cc: naive.cc
-	clang++ naive.cc -O3 -march=haswell -lbenchmark -o naive-cc.bin
-	./naive-cc.bin
+	clang++ naive.cc -O2 -march=haswell
 
 .PHONY: naive-rs
 naive-rs: naive.rs
-	rustc -C opt-level=3 -C target-feature=+avx2,+fma -o naive-rs.bin naive.rs
-	./naive-rs.bin
+	rustc -C opt-level=3 -C target-feature=+avx2,+fma -o a.out naive.rs
 
 .PHONY: functional-rs
 functional-rs: functional.rs
-	rustc -C opt-level=3 -C target-feature=+avx2,+fma -o functional-rs.bin functional.rs
-	./functional-rs.bin
+	rustc -C opt-level=3 -C target-feature=+avx2,+fma -o a.out functional.rs
+
+.PHONE: profile
+profile:
+	perf record ./a.out
+	perf annotate

README.md

@@ -1,14 +1,14 @@
 # Dot Product Benchmarks
 
-## Prerequisites
-
-1. Install https://github.com/google/benchmark
-
-## Usage
+Usage:
+```sh
+make naive-cc && ./a.out
+make naive-rs && ./a.out
+make functional-rs && ./a.out
+```
 
-Run benchmarks:
+Inspect assembly:
 ```sh
-make naive-cc
-make naive-rs
-make functional-rs
+make naive-cc && make profile
+...
 ```

functional.rs

@@ -9,7 +9,7 @@ fn dot(a: &[f32], b: &[f32]) -> f32 {
 
 fn main() {
     const SAMPLES: usize = 10;
-    const ITERS: usize = 1000;
+    const ITERS: usize = 10000;
     const LEN: usize = 100000;
 
     let a = [0.0; LEN];
@@ -21,6 +21,6 @@ fn main() {
             black_box(dot(black_box(&a), black_box(&b)));
         }
         let time_us = 1e6 * start.elapsed().unwrap().as_secs_f32() / ITERS as f32;
-        println!("{:4.2} us", time_us);
+        println!("{:8.2} us", time_us);
     }
 }

gbench.cc

@@ -0,0 +1,23 @@
+#include <benchmark/benchmark.h>
+
+float dot(float *a, float *b, size_t len) {
+    #pragma clang fp reassociate(on)
+    float sum = 0.0;
+    for (size_t i = 0; i < len; ++i) {
+        sum += a[i] * b[i];
+    }
+    return sum;
+}
+
+void BM_dot(benchmark::State& state) {
+    constexpr size_t LEN = 100000;
+    float a[LEN] = {};
+    float b[LEN] = {};
+
+    for (auto _ : state) {
+        benchmark::DoNotOptimize(dot(a, b, LEN));
+    }
+}
+
+BENCHMARK(BM_dot);
+BENCHMARK_MAIN();

naive.cc

@@ -1,4 +1,17 @@
-#include <benchmark/benchmark.h>
+#include <chrono>
+#include <cstdio>
+#include <type_traits>
+
+using namespace std::literals::chrono_literals;
+
+template<typename T> inline __attribute__((always_inline))
+void black_box(T &&value) noexcept {
+    if constexpr(std::is_pointer_v<T>) {
+        asm volatile("":"+m"(value)::"memory");
+    } else {
+        asm volatile("":"+r"(value)::);
+    }
+}
 
 float dot(float *a, float *b, size_t len) {
     #pragma clang fp reassociate(on)
@@ -9,13 +22,20 @@ float dot(float *a, float *b, size_t len) {
     return sum;
 }
 
-void BM_dot(benchmark::State& state) {
-    constexpr size_t len = 100000;
-    float a[len], b[len];
-    for (auto _ : state) {
-        benchmark::DoNotOptimize(dot(a, b, len));
+int main() {
+    constexpr size_t SAMPLES = 10;
+    constexpr size_t ITERS = 10000;
+    constexpr size_t LEN = 100000;
+
+    float a[LEN] = {};
+    float b[LEN] = {};
+
+    for (size_t s = 0; s < SAMPLES; ++s) {
+        auto start = std::chrono::system_clock::now();
+        for (size_t i = 0; i < ITERS; ++i) {
+            black_box(dot(a, b, LEN));
+        }
+        float time_us = (std::chrono::system_clock::now() - start) / 1ns / 1e3 / ITERS;
+        printf("%8.2f us\n", time_us);
     }
 }
-
-BENCHMARK(BM_dot);
-BENCHMARK_MAIN();

naive.rs

@@ -14,7 +14,7 @@ fn dot(a: &[f32], b: &[f32]) -> f32 {
 
 fn main() {
     const SAMPLES: usize = 10;
-    const ITERS: usize = 1000;
+    const ITERS: usize = 10000;
     const LEN: usize = 100000;
 
     let a = [0.0; LEN];
@@ -26,6 +26,6 @@ fn main() {
             black_box(dot(black_box(&a), black_box(&b)));
         }
         let time_us = 1e6 * start.elapsed().unwrap().as_secs_f32() / ITERS as f32;
-        println!("{:4.2} us", time_us);
+        println!("{:8.2} us", time_us);
     }
 }