🚀 perf: Performance and Memory optimization

internal/render/cache.go

@@ -2,7 +2,9 @@ package render
 
 import (
 	"bytes"
+	"compress/flate"
 	"container/list"
+	"io"
 	"sync"
 	"sync/atomic"
 	"time"
@@ -37,13 +39,48 @@ type Cache struct {
 }
 
 type cacheEntry struct {
-	key   cacheKey
+	key     cacheKey
-	ascii []byte
+	data    []byte
-	cost  int64
+	origLen int
+	cost    int64
 }
 
-func entryCost(_ cacheKey, ascii []byte) int64 {
+func entryCost(_ cacheKey, data []byte) int64 {
-	return int64(cap(ascii)) + 160
+	return int64(cap(data)) + 160
+}
+
+var flateWriters = sync.Pool{New: func() any {
+	w, _ := flate.NewWriter(io.Discard, 1)
+	return w
+}}
+
+type flateReader interface {
+	io.Reader
+	flate.Resetter
+}
+
+var flateReaders = sync.Pool{New: func() any {
+	return flate.NewReader(bytes.NewReader(nil)).(flateReader)
+}}
+
+func compressAscii(src []byte) []byte {
+	var buf bytes.Buffer
+	buf.Grow(len(src)/3 + 64)
+	w := flateWriters.Get().(*flate.Writer)
+	w.Reset(&buf)
+	_, _ = w.Write(src)
+	_ = w.Close()
+	flateWriters.Put(w)
+	return bytes.Clone(buf.Bytes())
+}
+
+func decompressAscii(src []byte, origLen int) []byte {
+	r := flateReaders.Get().(flateReader)
+	_ = r.Reset(bytes.NewReader(src), nil)
+	buf := bytes.NewBuffer(make([]byte, 0, origLen))
+	_, _ = io.Copy(buf, r)
+	flateReaders.Put(r)
+	return buf.Bytes()
 }
 
 func NewCache(maxBytes int64) *Cache {
@@ -82,15 +119,18 @@ func (c *Cache) get(key cacheKey) ([]byte, bool) {
 	}
 	shard := c.shard(key)
 	shard.mu.Lock()
-	defer shard.mu.Unlock()
 	el, ok := shard.entries[key]
 	if !ok {
+		shard.mu.Unlock()
 		c.misses.Add(1)
 		return nil, false
 	}
-	c.hits.Add(1)
 	shard.order.MoveToBack(el)
-	return el.Value.(*cacheEntry).ascii, true
+	entry := el.Value.(*cacheEntry)
+	data, origLen := entry.data, entry.origLen
+	shard.mu.Unlock()
+	c.hits.Add(1)
+	return decompressAscii(data, origLen), true
 }
 
 func (c *Cache) put(key cacheKey, ascii []byte) {
@@ -98,7 +138,8 @@ func (c *Cache) put(key cacheKey, ascii []byte) {
 		return
 	}
 	shard := c.shard(key)
-	cost := entryCost(key, ascii)
+	compressed := compressAscii(ascii)
+	cost := entryCost(key, compressed)
 	if cost > shard.maxBytes {
 		c.rejections.Add(1)
 		return
@@ -108,7 +149,8 @@ func (c *Cache) put(key cacheKey, ascii []byte) {
 	if _, ok := shard.entries[key]; ok {
 		return
 	}
-	shard.entries[key] = shard.order.PushBack(&cacheEntry{key: key, ascii: ascii, cost: cost})
+	entry := &cacheEntry{key: key, data: compressed, origLen: len(ascii), cost: cost}
+	shard.entries[key] = shard.order.PushBack(entry)
 	shard.size += cost
 	c.size.Add(cost)
 	for shard.size > shard.maxBytes {
@@ -219,9 +261,6 @@ func (r *Renderer) Render(index, width, height int, keepAspectRatio bool, tier C
 	}
 	putPixBuf(pix)
 
-	if r.cache != nil && cap(ascii) > len(ascii)+len(ascii)/4 {
-		ascii = bytes.Clone(ascii)
-	}
 	r.cache.put(key, ascii)
 	if r.cache != nil {
 		r.cache.renders.Add(1)

internal/render/render_test.go

@@ -88,12 +88,25 @@ func TestRenderConcurrentSameKey(t *testing.T) {
 	}
 }
 
+func incompressible(n int) []byte {
+	b := make([]byte, n)
+	x := uint32(0x9e3779b9)
+	for i := range b {
+		x ^= x << 13
+		x ^= x >> 17
+		x ^= x << 5
+		b[i] = byte(x)
+	}
+	return b
+}
+
 func TestRenderCacheEvictsByBytes(t *testing.T) {
 	key := func(index int) cacheKey { return cacheKey{index: index} }
-	budget := 3 * entryCost(key(0), bytes.Repeat([]byte("x"), 1000))
+	payload := incompressible(1000)
+	budget := 3 * entryCost(key(0), compressAscii(payload))
 	c := NewCache(budget)
 	for i := range 5 {
-		c.put(key(i), bytes.Repeat([]byte("x"), 1000))
+		c.put(key(i), payload)
 	}
 	if c.size.Load() > budget {
 		t.Errorf("size %d exceeds budget %d", c.size.Load(), budget)
@@ -106,6 +119,19 @@ func TestRenderCacheEvictsByBytes(t *testing.T) {
 	}
 }
 
+func TestRenderCacheRoundTrips(t *testing.T) {
+	c := NewCache(1 << 20)
+	want := incompressible(4096)
+	c.put(cacheKey{}, want)
+	got, ok := c.get(cacheKey{})
+	if !ok {
+		t.Fatal("entry should be cached")
+	}
+	if !bytes.Equal(got, want) {
+		t.Fatal("decompressed entry does not match original")
+	}
+}
+
 func TestRenderCacheDisabled(t *testing.T) {
 	c := NewCache(0)
 	if c != nil {
@@ -119,7 +145,7 @@ func TestRenderCacheDisabled(t *testing.T) {
 
 func TestRenderCacheRejectsOversizedEntry(t *testing.T) {
 	c := NewCache(256)
-	c.put(cacheKey{}, bytes.Repeat([]byte("x"), 10_000))
+	c.put(cacheKey{}, incompressible(10_000))
 	if _, ok := c.get(cacheKey{}); ok {
 		t.Error("entry larger than budget should not be cached")
 	}
@@ -128,15 +154,15 @@ func TestRenderCacheRejectsOversizedEntry(t *testing.T) {
 	}
 }
 
-func TestRenderCacheAccountsRetainedCapacity(t *testing.T) {
+func TestRenderCacheAccountsCompressedSize(t *testing.T) {
 	cache := NewCache(512)
 	value := make([]byte, 1, 4096)
 	cache.put(cacheKey{}, value)
-	if _, ok := cache.get(cacheKey{}); ok {
+	if _, ok := cache.get(cacheKey{}); !ok {
-		t.Fatal("cache accepted an entry whose backing allocation exceeds its budget")
+		t.Fatal("small value should be cached regardless of its backing capacity")
 	}
-	if cache.Stats().Rejections != 1 {
+	if got := cache.size.Load(); got > 512 {
-		t.Fatalf("rejections = %d, want 1", cache.Stats().Rejections)
+		t.Fatalf("size = %d, want <= 512", got)
 	}
 }
 

internal/tsf/format.go

@@ -112,6 +112,9 @@ func Load(filename string) (*FramesContainer, error) {
 	if off != len(raw) {
 		return nil, invalid()
 	}
+
+	file.dropResident()
+
 	data := &FramesContainer{ColorFrames: colorFrames, FPS: fps}
 	frameFileOwners.Store(data, file)
 	owned = true

internal/tsf/mmap_other.go

@@ -8,3 +8,5 @@ func readFrameFile(filename string) (*frameFile, error) {
 	data, err := os.ReadFile(filename)
 	return &frameFile{data: data}, err
 }
+
+func (f *frameFile) dropResident() {}

internal/tsf/mmap_unix.go

@@ -28,6 +28,8 @@ func readFrameFile(filename string) (*frameFile, error) {
 		data, err := os.ReadFile(filename)
 		return &frameFile{data: data}, err
 	}
+
+	_ = syscall.Madvise(data, syscall.MADV_RANDOM)
 	return &frameFile{
 		data: data,
 		cleanup: func() error {
@@ -35,3 +37,10 @@ func readFrameFile(filename string) (*frameFile, error) {
 		},
 	}, nil
 }
+
+func (f *frameFile) dropResident() {
+	if f == nil || f.cleanup == nil || len(f.data) == 0 {
+		return
+	}
+	_ = syscall.Madvise(f.data, syscall.MADV_DONTNEED)
+}