diff --git a/counter.go b/counter.go
new file mode 100644
index 0000000..3bf2db6
--- /dev/null
+++ b/counter.go
@@ -0,0 +1,58 @@
+package perp
+
+import (
+	"sync/atomic"
+)
+
+// Counter is sharded across processors in Go runtime.
+type Counter struct {
+	vs *Values[*counter]
+}
+
+// NewCounter returns a new Counter initialized to zero.
+func NewCounter() *Counter {
+	vs := NewValues(func() *counter {
+		return &counter{}
+	})
+	return &Counter{vs: vs}
+}
+
+// Add n to the counter.
+func (c *Counter) Add(n int64) {
+	c.vs.Get().Add(n)
+}
+
+// Load the total counter value.
+func (c *Counter) Load() int64 {
+	var sum int64
+	c.vs.Iter(func(v *counter) {
+		sum += v.Load()
+	})
+	return sum
+}
+
+// Reset the counter to zero and return the old value.
+func (c *Counter) Reset() int64 {
+	var sum int64
+	c.vs.Iter(func(v *counter) {
+		sum += v.Reset()
+	})
+	return sum
+}
+
+type counter struct {
+	n int64
+	_ [56]byte // cache line aligment
+}
+
+func (c *counter) Add(n int64) int64 {
+	return atomic.AddInt64(&c.n, n)
+}
+
+func (c *counter) Load() int64 {
+	return atomic.LoadInt64(&c.n)
+}
+
+func (c *counter) Reset() int64 {
+	return atomic.SwapInt64(&c.n, 0)
+}
diff --git a/counter_test.go b/counter_test.go
new file mode 100644
index 0000000..de2e8a2
--- /dev/null
+++ b/counter_test.go
@@ -0,0 +1,178 @@
+package perp
+
+import (
+	"sync"
+	"sync/atomic"
+	"testing"
+)
+
+func TestCounter(t *testing.T) {
+	c := NewCounter()
+
+	const n = 100
+	var wg sync.WaitGroup
+	wg.Add(n)
+
+	for i := 0; i < n; i++ {
+		go func() {
+			defer wg.Done()
+			for i := 0; i < n; i++ {
+				c.Add(1)
+			}
+		}()
+	}
+	wg.Wait()
+
+	mustEqual(t, c.Load(), int64(n*n))
+}
+
+func TestCounterReset(t *testing.T) {
+	c := NewCounter()
+
+	const n = 100
+	var wg sync.WaitGroup
+	wg.Add(n)
+
+	var sum int64
+
+	for i := 0; i < n; i++ {
+		go func() {
+			defer wg.Done()
+			for i := 0; i < n; i++ {
+				c.Add(1)
+				if i%32 == 0 {
+					atomic.AddInt64(&sum, c.Reset())
+				}
+			}
+		}()
+	}
+
+	wg.Wait()
+
+	sum += c.Reset()
+	mustEqual(t, sum, int64(n*n))
+	mustEqual(t, c.Load(), int64(0))
+}
+
+func BenchmarkCounter(b *testing.B) {
+	c := NewCounter()
+
+	for i := 0; i < b.N; i++ {
+		c.Add(1)
+	}
+	mustEqual(b, c.Load(), int64(b.N))
+}
+
+func BenchmarkCounterMutex(b *testing.B) {
+	var c mutexCounter
+
+	for i := 0; i < b.N; i++ {
+		c.Add(1)
+	}
+	mustEqual(b, c.Load(), int64(b.N))
+}
+
+func BenchmarkCounterAtomic(b *testing.B) {
+	var c atomicCounter
+
+	for i := 0; i < b.N; i++ {
+		c.Add(1)
+	}
+	mustEqual(b, c.Load(), int64(b.N))
+}
+
+func BenchmarkCounterPaddedAtomic(b *testing.B) {
+	var c paddedAtomicCounter
+
+	for i := 0; i < b.N; i++ {
+		c.Add(1)
+	}
+	mustEqual(b, c.Load(), int64(b.N))
+}
+
+func BenchmarkParallelCounter(b *testing.B) {
+	c := NewCounter()
+
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			c.Add(1)
+		}
+	})
+	mustEqual(b, c.Load(), int64(b.N))
+}
+
+func BenchmarkParallelCounterMutex(b *testing.B) {
+	var c mutexCounter
+
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			c.Add(1)
+		}
+	})
+	mustEqual(b, c.Load(), int64(b.N))
+}
+
+func BenchmarkParallelCounterAtomic(b *testing.B) {
+	var c atomicCounter
+
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			c.Add(1)
+		}
+	})
+	mustEqual(b, c.Load(), int64(b.N))
+}
+
+func BenchmarkParallelCounterPaddedAtomic(b *testing.B) {
+	var c paddedAtomicCounter
+
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			c.Add(1)
+		}
+	})
+	mustEqual(b, c.Load(), int64(b.N))
+}
+
+type mutexCounter struct {
+	mu sync.Mutex
+	n  int64
+}
+
+func (c *mutexCounter) Add(n int64) {
+	c.mu.Lock()
+	c.n += n
+	c.mu.Unlock()
+}
+
+func (c *mutexCounter) Load() int64 {
+	c.mu.Lock()
+	v := c.n
+	c.mu.Unlock()
+	return v
+}
+
+type atomicCounter struct {
+	n int64
+}
+
+func (c *atomicCounter) Add(n int64) {
+	atomic.AddInt64(&c.n, n)
+}
+
+func (c *atomicCounter) Load() int64 {
+	return atomic.LoadInt64(&c.n)
+}
+
+type paddedAtomicCounter struct {
+	_ [56]byte
+	n int64
+}
+
+func (c *paddedAtomicCounter) Add(n int64) {
+	atomic.AddInt64(&c.n, n)
+}
+
+func (c *paddedAtomicCounter) Load() int64 {
+	return atomic.LoadInt64(&c.n)
+}