From 10aac0e25382b962476dcba65dbd420864b9134f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Wojciech=20Ma=C5=82ota-W=C3=B3jcik?=
 <59281144+outofforest@users.noreply.github.com>
Date: Thu, 21 Nov 2024 21:35:54 +0100
Subject: [PATCH] Benchmark double copy (#220)

---
 wal/asm.s             |  14 ++++++
 wal/asm_stub.go       |   6 +++
 wal/avo/avo.go        |  31 +++++++++++++
 wal/benchmark_test.go | 103 ++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 154 insertions(+)
 create mode 100644 wal/asm.s
 create mode 100644 wal/asm_stub.go
 create mode 100644 wal/avo/avo.go
 create mode 100644 wal/benchmark_test.go

diff --git a/wal/asm.s b/wal/asm.s
new file mode 100644
index 0000000..a7aa0a8
--- /dev/null
+++ b/wal/asm.s
@@ -0,0 +1,14 @@
+// Code generated by command: go run avo.go -out ../asm.s -stubs ../asm_stub.go -pkg wal. DO NOT EDIT.
+
+#include "textflag.h"
+
+// func Copy(x *byte, y *byte, z *byte)
+// Requires: AVX512F
+TEXT ·Copy(SB), NOSPLIT, $0-24
+	MOVQ      z+16(FP), AX
+	VMOVDQU64 (AX), Z0
+	MOVQ      x+0(FP), AX
+	VMOVDQU64 Z0, (AX)
+	MOVQ      y+8(FP), AX
+	VMOVDQU64 Z0, (AX)
+	RET
diff --git a/wal/asm_stub.go b/wal/asm_stub.go
new file mode 100644
index 0000000..16b4b39
--- /dev/null
+++ b/wal/asm_stub.go
@@ -0,0 +1,6 @@
+// Code generated by command: go run avo.go -out ../asm.s -stubs ../asm_stub.go -pkg wal. DO NOT EDIT.
+
+package wal
+
+// Copy copies data.
+func Copy(x *byte, y *byte, z *byte)
diff --git a/wal/avo/avo.go b/wal/avo/avo.go
new file mode 100644
index 0000000..c16c7ee
--- /dev/null
+++ b/wal/avo/avo.go
@@ -0,0 +1,31 @@
+package main
+
+//go:generate go run . -out ../asm.s -stubs ../asm_stub.go -pkg wal
+
+import (
+	. "github.com/mmcloughlin/avo/build"
+	. "github.com/mmcloughlin/avo/operand"
+)
+
+// Copy copies data.
+func Copy() {
+	TEXT("Copy", NOSPLIT, "func(x *byte, y *byte, z *byte)")
+	Doc("Copy copies data.")
+
+	r := ZMM()
+	memZ := Mem{Base: Load(Param("z"), GP64())}
+	VMOVDQU64(memZ, r)
+
+	memX := Mem{Base: Load(Param("x"), GP64())}
+	VMOVDQU64(r, memX)
+	memY := Mem{Base: Load(Param("y"), GP64())}
+	VMOVDQU64(r, memY)
+
+	RET()
+}
+
+func main() {
+	Copy()
+
+	Generate()
+}
diff --git a/wal/benchmark_test.go b/wal/benchmark_test.go
new file mode 100644
index 0000000..20614f3
--- /dev/null
+++ b/wal/benchmark_test.go
@@ -0,0 +1,103 @@
+package wal_test
+
+import (
+	"crypto/rand"
+	"fmt"
+	"io"
+	"testing"
+	"unsafe"
+
+	"github.com/outofforest/quantum/wal"
+)
+
+type item struct {
+	Field00 uint32
+	Field01 uint32
+	Field02 uint64
+	Field03 uint64
+	Field04 uint16
+	Field05 uint16
+	Field06 uint16
+	Field07 uint16
+	Field08 uint64
+	Field09 uint64
+	Field10 uint64
+	Field11 uint64
+}
+
+var (
+	data = func() item {
+		var i item
+		_, _ = rand.Read(unsafe.Slice((*byte)(unsafe.Pointer(&i)), unsafe.Sizeof(i)))
+		return i
+	}()
+)
+
+func BenchmarkGoCopy(b *testing.B) {
+	b.StopTimer()
+	b.ResetTimer()
+
+	var data0, data1, data2, data3, data4, data5, data6, data7, data8, data9 item
+
+	dataB := unsafe.Slice((*byte)(unsafe.Pointer(&data)), unsafe.Sizeof(data))
+	data0B := unsafe.Slice((*byte)(unsafe.Pointer(&data0)), unsafe.Sizeof(data0))
+	data1B := unsafe.Slice((*byte)(unsafe.Pointer(&data1)), unsafe.Sizeof(data1))
+	data2B := unsafe.Slice((*byte)(unsafe.Pointer(&data2)), unsafe.Sizeof(data2))
+	data3B := unsafe.Slice((*byte)(unsafe.Pointer(&data3)), unsafe.Sizeof(data3))
+	data4B := unsafe.Slice((*byte)(unsafe.Pointer(&data4)), unsafe.Sizeof(data4))
+	data5B := unsafe.Slice((*byte)(unsafe.Pointer(&data5)), unsafe.Sizeof(data5))
+	data6B := unsafe.Slice((*byte)(unsafe.Pointer(&data6)), unsafe.Sizeof(data6))
+	data7B := unsafe.Slice((*byte)(unsafe.Pointer(&data7)), unsafe.Sizeof(data7))
+	data8B := unsafe.Slice((*byte)(unsafe.Pointer(&data8)), unsafe.Sizeof(data8))
+	data9B := unsafe.Slice((*byte)(unsafe.Pointer(&data9)), unsafe.Sizeof(data9))
+
+	b.StartTimer()
+	for range b.N {
+		copy(data0B, dataB)
+		copy(data1B, dataB)
+		copy(data2B, dataB)
+		copy(data3B, dataB)
+		copy(data4B, dataB)
+		copy(data5B, dataB)
+		copy(data6B, dataB)
+		copy(data7B, dataB)
+		copy(data8B, dataB)
+		copy(data9B, dataB)
+	}
+	b.StopTimer()
+
+	_, _ = fmt.Fprint(io.Discard, data0B)
+	_, _ = fmt.Fprint(io.Discard, data1B)
+	_, _ = fmt.Fprint(io.Discard, data2B)
+	_, _ = fmt.Fprint(io.Discard, data3)
+	_, _ = fmt.Fprint(io.Discard, data4B)
+	_, _ = fmt.Fprint(io.Discard, data5B)
+	_, _ = fmt.Fprint(io.Discard, data6B)
+	_, _ = fmt.Fprint(io.Discard, data7B)
+	_, _ = fmt.Fprint(io.Discard, data8B)
+	_, _ = fmt.Fprint(io.Discard, data9B)
+}
+
+func BenchmarkAVX(b *testing.B) {
+	b.StopTimer()
+	b.ResetTimer()
+
+	var data0, data1 item
+
+	dataB := (*byte)(unsafe.Pointer(&data))
+	data0B := (*byte)(unsafe.Pointer(&data0))
+	data1B := (*byte)(unsafe.Pointer(&data1))
+
+	b.StartTimer()
+	for range b.N {
+		wal.Copy(data0B, data1B, dataB)
+		wal.Copy(data0B, data1B, dataB)
+		wal.Copy(data0B, data1B, dataB)
+		wal.Copy(data0B, data1B, dataB)
+		wal.Copy(data0B, data1B, dataB)
+	}
+	b.StopTimer()
+
+	_, _ = fmt.Fprint(io.Discard, data0)
+	_, _ = fmt.Fprint(io.Discard, data1)
+}