// dotproduct_test.go
package dotproduct

import (
	"runtime"
	"sync/atomic"
	"testing"
	"unsafe"

	"golang.org/x/sys/cpu"
)

func TestAVX2hasAVX(t *testing.T) {
	if runtime.GOARCH == "amd64" {
		if cpu.X86.HasAVX2 && !cpu.X86.HasAVX {
			t.Fatal("HasAVX expected true, got false")
		} else if !cpu.X86.HasAVX2 {
			t.Log("HasAVX2=False")
		}
	}
}
func TestAVX512HasAVX2AndAVX(t *testing.T) {
	if runtime.GOARCH == "amd64" {
		if cpu.X86.HasAVX512 && !cpu.X86.HasAVX {
			t.Fatal("HasAVX expected true, got false")
		}
		if cpu.X86.HasAVX512 && !cpu.X86.HasAVX2 {
			t.Fatal("HasAVX2 expected true, got false")
		}
		if !cpu.X86.HasAVX512 {
			t.Log("HasAVX512=False")
		}
	}
}

func DotProduct(a []int32, b []int32, N int32) (sum int32) {
	//N := len(a)
	for i := int32(0); i < N; i++ {
		sum += a[i] * b[i]
	}
	return
}
func TestSumAsm(t *testing.T) {
	if runtime.GOARCH == "amd64" {
		d3 := make([]uint64, 24)
		var sumGo uint64
		for i := 0; i < 24; i++ {
			d3[i] = uint64(i + 1)
			sumGo = sumGo + d3[i]
		}
		sumAsm := Sum(d3)
		t.Log("sumGo ", sumGo, ", sumAsm ", sumAsm)
	} else {
		t.Skip("test case *TestSumAsm* not applicable")
	}
}
func TestDotProductAsm(t *testing.T) {
	if runtime.GOARCH == "amd64" {
		const len32 = 24
		d1 := make([]int32, len32)
		d2 := make([]int32, len32)
		for i := 0; i < len32; i++ {
			d1[i] = int32(i + 1)
			d2[i] = int32(2 * i)
		}

		sumGo := DotProduct(d1, d2, len32)
		sumAsm := DotProductAsm(d1, d2, len32)
		t.Log("DotProductGo ", sumGo, ", DotProductAsm ", sumAsm)

	} else {
		t.Skip("test case *TestDotProductAsm* not applicable")
	}
}

func TestDotProductAsmAvx(t *testing.T) {
	if runtime.GOARCH == "amd64" && cpu.X86.HasAVX2 {
		const len32 = 32
		d1 := make([]int32, len32)
		d2 := make([]int32, len32)
		for i := 0; i < len32; i++ {
			d1[i] = int32(i + 1)
			d2[i] = int32(2 * i)
		}
		sumGo := DotProduct(d1, d2, len32)
		sumAvx := DotProductAsmAvx(d1, d2, len32)
		t.Log("DotProductGo ", sumGo, ", DotProductAvx ", sumAvx)
	} else {
		t.Skip("test case *TestDotProductAsmAvx* not applicable")
	}
}
func BenchmarkSum(b *testing.B) {
	const len32 = 4096
	d1 := make([]uint64, len32)
	for i := 0; i < len32; i++ {
		d1[i] = uint64(i + 1)

	}
	var sum2 uint64 = 0
	b.SetBytes(int64(len32 * unsafe.Sizeof(uint64(0))))
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		sum2 += SumGo(d1) % len32
	}
}

func BenchmarkSumParallel(b *testing.B) {
	const len32 = 4096
	d1 := make([]uint64, len32)
	for i := 0; i < len32; i++ {
		d1[i] = uint64(i + 1)
	}
	b.SetBytes(int64(len32 * unsafe.Sizeof(uint64(0))))
	b.ResetTimer()
	b.RunParallel(func(pb *testing.PB) {
		for pb.Next() {
			var sum2 uint64 = 0
				sum2 += SumGo(d1) % len32		
		}
	})
}

func BenchmarkSumAsm(b *testing.B) {
	if runtime.GOARCH == "amd64" {
		const len32 = 4096
		d1 := make([]uint64, len32)
		for i := 0; i < len32; i++ {
			d1[i] = uint64(i + 1)
		}
		var sum2 uint64 = 0
		b.SetBytes(int64(len32 * unsafe.Sizeof(uint64(0))))
		b.ResetTimer()
		for i := 0; i < b.N; i++ {
			sum2 += Sum(d1) % len32
		}
	} else {
		b.Skip("test case *BenchmarkSumAsm* not applicable")
	}
}

func BenchmarkSumAsmParallel(b *testing.B) {
	if runtime.GOARCH == "amd64" {
		const len32 = 4096
		d1 := make([]uint64, len32)
		for i := 0; i < len32; i++ {
			d1[i] = uint64(i + 1)
		}
		//glob := b.N
		b.SetBytes(int64(len32 * unsafe.Sizeof(uint64(0))))
		b.ResetTimer()
		b.RunParallel(func(pb *testing.PB) {
			for pb.Next() {
				var sum2 uint64 = 0

				//for i := 0; i < 24; i++ {
				sum2 += Sum(d1) % len32
				//}
			}
		})

	} else {
		b.Skip("test case *BenchmarkSumAsm* not applicable")
	}
}

func SumGo(a []uint64) (sum uint64) {
	N := len(a)
	for i := 0; i < N; i++ {
		sum += a[i]
	}
	return
}

func BenchmarkDotProduct(b *testing.B) {
	const len32 = 1024
	d1 := make([]int32, len32)
	d2 := make([]int32, len32)
	for i := 0; i < len32; i++ {
		d1[i] = int32(i + 1)
		d2[i] = int32(2 * i)
	}
	var sum2 int32 = 0
	b.SetBytes(int64(len32 * unsafe.Sizeof(int32(0))))
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		sum2 += DotProduct(d1, d2, len32) % len32
	}
}

func BenchmarkDotProductAsm(b *testing.B) {
	if runtime.GOARCH == "amd64" {
		const len32 = 1024
		d1 := make([]int32, len32)
		d2 := make([]int32, len32)
		for i := 0; i < len32; i++ {
			d1[i] = int32(i + 1)
			d2[i] = int32(2 * i)
		}
		var sum2 int32 = 0
		b.SetBytes(int64(len32 * unsafe.Sizeof(int32(0))))
		b.ResetTimer()
		for i := 0; i < b.N; i++ {
			sum2 += DotProductAsm(d1, d2, len32) % len32
		}
	} else {
		b.Skip("test case *BenchmarkDotProductAsm* not applicable")
	}
}

func BenchmarkDotProductAsmAvx2(b *testing.B) {
	if runtime.GOARCH == "amd64" && cpu.X86.HasAVX2 {
		const len32 = 1024
		d1 := make([]int32, len32)
		d2 := make([]int32, len32)
		for i := 0; i < len32; i++ {
			d1[i] = int32(i + 1)
			d2[i] = int32(2 * i)
		}
		var sum2 int32 = 0
		b.SetBytes(int64(len32 * unsafe.Sizeof(int32(0))))
		b.ResetTimer()
		for i := 0; i < b.N; i++ {
			sum2 += DotProductAsmAvx(d1, d2, len32) % len32
		}
	} else {
		b.Skip("test case *BenchmarkDotProductAsmAvx2* not applicable")
	}
}

func BenchmarkDotProductAsmAvx2Parallel(b *testing.B) {
	if runtime.GOARCH == "amd64" && cpu.X86.HasAVX2 {
		const len32 = 1024
		d1 := make([]int32, len32)
		d2 := make([]int32, len32)
		for i := 0; i < len32; i++ {
			d1[i] = int32(i + 1)
			d2[i] = int32(2 * i)
		}
		b.SetBytes(int64(len32 * unsafe.Sizeof(int32(0))))
		b.ResetTimer()
		b.RunParallel(func(pb *testing.PB) {
			for pb.Next() {
				var sum2 int32 = 0
				//for i := 0; i < 24; i++ {
				sum2 += DotProductAsmAvx(d1, d2, len32) % len32
				//}
			}
		})
	} else {
		b.Skip("test case *BenchmarkDotProductAsmAvx2* not applicable")
	}
}

func BenchmarkHowItWorksParallel(b *testing.B) {
	procs := uint32(0)
	iters := uint64(0)
	//b.SetParallelism(3) //number of goroutine to each b.RunParallel
	b.SetBytes(int64(4 * unsafe.Sizeof(uint64(0))))
	b.ResetTimer()
	b.RunParallel(func(pb *testing.PB) {
		atomic.AddUint32(&procs, 1)
		for pb.Next() {
			atomic.AddUint64(&iters, 1)
		}
	})
	b.Log("goroutine count :", procs, ", retrial count :", iters)
}