Merge pull request #563 from markus-wa/s2-perf-improvements

Source 2 performance improvements

Author: markus-wa

go.mod

@@ -6,7 +6,7 @@ require (
 	github.com/llgcode/draw2d v0.0.0-20230723155556-e595d7c7e75e
 	github.com/markus-wa/go-heatmap/v2 v2.0.0
 	github.com/markus-wa/go-unassert v0.1.3
-	github.com/markus-wa/gobitread v0.2.3
+	github.com/markus-wa/gobitread v0.2.4
 	github.com/markus-wa/godispatch v1.4.1
 	github.com/markus-wa/ice-cipher-go v0.0.0-20230901094113-348096939ba7
 	github.com/markus-wa/quickhull-go/v2 v2.2.0

go.sum

@@ -21,6 +21,8 @@ github.com/markus-wa/go-unassert v0.1.3 h1:4N2fPLUS3929Rmkv94jbWskjsLiyNT2yQpCul
 github.com/markus-wa/go-unassert v0.1.3/go.mod h1:/pqt7a0LRmdsRNYQ2nU3SGrXfw3bLXrvIkakY/6jpPY=
 github.com/markus-wa/gobitread v0.2.3 h1:COx7dtYQ7Q+77hgUmD+O4MvOcqG7y17RP3Z7BbjRvPs=
 github.com/markus-wa/gobitread v0.2.3/go.mod h1:PcWXMH4gx7o2CKslbkFkLyJB/aHW7JVRG3MRZe3PINg=
+github.com/markus-wa/gobitread v0.2.4 h1:BDr3dZnsqntDD4D8E7DzhkQlASIkQdfxCXLhWcI2K5A=
+github.com/markus-wa/gobitread v0.2.4/go.mod h1:PcWXMH4gx7o2CKslbkFkLyJB/aHW7JVRG3MRZe3PINg=
 github.com/markus-wa/godispatch v1.4.1 h1:Cdff5x33ShuX3sDmUbYWejk7tOuoHErFYMhUc2h7sLc=
 github.com/markus-wa/godispatch v1.4.1/go.mod h1:tk8L0yzLO4oAcFwM2sABMge0HRDJMdE8E7xm4gK/+xM=
 github.com/markus-wa/ice-cipher-go v0.0.0-20230901094113-348096939ba7 h1:aR9pvnlnBxifXBmzidpAiq2prLSGlkhE904qnk2sCz4=

pkg/demoinfocs/common/player.go

@@ -16,7 +16,7 @@ type Player struct {
 	demoInfoProvider demoInfoProvider // provider for demo info such as tick-rate or current tick
 
 	SteamID64             uint64             // 64-bit representation of the user's Steam ID. See https://developer.valvesoftware.com/wiki/SteamID
-	LastAlivePosition     r3.Vector          // The location where the player was last alive. Should be equal to Position if the player is still alive.
+	LastAlivePosition     r3.Vector          // Deprecated: will be removed in v5 due to performance concerns, track this yourself.
 	UserID                int                // Mostly used in game-events to address this player
 	Name                  string             // Steam / in-game user name
 	Inventory             map[int]*Equipment // All weapons / equipment the player is currently carrying. See also Weapons().
@@ -33,7 +33,7 @@ type Player struct {
 	IsPlanting            bool
 	IsReloading           bool
 	IsUnknown             bool      // Used to identify unknown/broken players. see https://github.com/markus-wa/demoinfocs-golang/issues/162
-	PreviousFramePosition r3.Vector // CS2 only, used to compute velocity as it's not networked in CS2 demos
+	PreviousFramePosition r3.Vector // Deprecated: may be removed in v5 due to performance concerns, track this yourself.
 }
 
 func (p *Player) PlayerPawnEntity() st.Entity {
@@ -85,9 +85,11 @@ func (p *Player) IsAlive() bool {
 	}
 
 	if p.demoInfoProvider.IsSource2() {
-		if pawnEntity := p.PlayerPawnEntity(); pawnEntity != nil {
+		pawnEntity := p.PlayerPawnEntity()
+		if pawnEntity != nil {
 			return pawnEntity.PropertyValueMust("m_lifeState").S2UInt64() == 0
 		}
+
 		return getBool(p.Entity, "m_bPawnIsAlive")
 	}
 
@@ -535,6 +537,7 @@ func (p *Player) PositionEyes() r3.Vector {
 }
 
 // Velocity returns the player's velocity.
+// Deprecated: will be removed due to performance concerns, you will need to track this yourself.
 func (p *Player) Velocity() r3.Vector {
 	if p.demoInfoProvider.IsSource2() {
 		t := 64.0

pkg/demoinfocs/datatables.go

@@ -3,6 +3,7 @@ package demoinfocs
 import (
 	"fmt"
 	"math"
+	"os"
 	"strings"
 
 	"github.com/golang/geo/r3"
@@ -637,6 +638,7 @@ func (p *parser) bindNewPlayerPawnS2(pawnEntity st.Entity) {
 		if pl == nil {
 			return
 		}
+
 		if pl.IsAlive() {
 			pl.LastAlivePosition = pos
 		}
@@ -919,7 +921,7 @@ func (p *parser) bindGrenadeProjectiles(entity st.Entity) {
 				if exists {
 					wep = weaponType
 				} else {
-					fmt.Printf("unknown grenade model %d\n", model)
+					fmt.Fprintf(os.Stderr, "unknown grenade model %d\n", model)
 				}
 			}
 		}

pkg/demoinfocs/demoinfocs_test.go

@@ -493,7 +493,7 @@ func TestConcurrent(t *testing.T) {
 func parseDefaultDemo(tb testing.TB) {
 	tb.Helper()
 
-	f := openFile(tb, defaultDemPath)
+	f := openFile(tb, s2DemPath)
 	defer mustClose(tb, f)
 
 	p := demoinfocs.NewParser(f)
@@ -599,15 +599,15 @@ func BenchmarkDemoInfoCs(b *testing.B) {
 }
 
 func BenchmarkInMemory(b *testing.B) {
-	f := openFile(b, defaultDemPath)
+	f := openFile(b, s2DemPath)
 	defer mustClose(b, f)
 
 	inf, err := f.Stat()
-	assert.NoError(b, err, "failed to stat file %q", defaultDemPath)
+	assert.NoError(b, err, "failed to stat file %q", s2DemPath)
 
 	d := make([]byte, inf.Size())
 	n, err := f.Read(d)
-	assert.NoError(b, err, "failed to read file %q", defaultDemPath)
+	assert.NoError(b, err, "failed to read file %q", s2DemPath)
 	assert.Equal(b, int64(n), inf.Size(), "byte count not as expected")
 
 	b.ResetTimer()

pkg/demoinfocs/parser.go

@@ -104,6 +104,7 @@ type parser struct {
 	equipmentTypePerModel map[uint64]common.EquipmentType                 // Used to retrieve the EquipmentType of grenade projectiles based on models value. Source 2 only.
 	stringTables          []createStringTable                             // Contains all created sendtables, needed when updating them
 	delayedEventHandlers  []func()                                        // Contains event handlers that need to be executed at the end of a tick (e.g. flash events because FlashDuration isn't updated before that)
+	pendingMessagesCache  []pendingMessage                                // Cache for pending messages that need to be dispatched after the current tick
 }
 
 // NetMessageCreator creates additional net-messages to be dispatched to net-message handlers.

pkg/demoinfocs/s2_commands.go

@@ -310,21 +310,21 @@ func (p *parser) handleDemoPacket(pack *msgs2.CDemoPacket) {
 
 	r := bitread.NewSmallBitReader(bytes.NewReader(b))
 
-	ms := make([]pendingMessage, 0)
+	p.pendingMessagesCache = p.pendingMessagesCache[:0]
 
 	for len(b)*8-r.ActualPosition() > 7 {
 		t := int32(r.ReadUBitInt())
 		size := r.ReadVarInt32()
 		buf := r.ReadBytes(int(size))
 
-		ms = append(ms, pendingMessage{t, buf})
+		p.pendingMessagesCache = append(p.pendingMessagesCache, pendingMessage{t, buf})
 	}
 
-	sort.SliceStable(ms, func(i, j int) bool {
-		return ms[i].priority() < ms[j].priority() // TODO: taken from dotabuff/manta. do we really need this?
+	sort.SliceStable(p.pendingMessagesCache, func(i, j int) bool {
+		return p.pendingMessagesCache[i].priority() < p.pendingMessagesCache[j].priority()
 	})
 
-	for _, m := range ms {
+	for _, m := range p.pendingMessagesCache {
 		var msgCreator NetMessageCreator
 
 		if m.t < int32(msgs2.SVC_Messages_svc_ServerInfo) {

pkg/demoinfocs/sendtables2/entity.go

@@ -424,9 +424,9 @@ func (p *Parser) FilterEntity(fb func(*Entity) bool) []*Entity {
 }
 
 func (e *Entity) readFields(r *reader, paths *[]*fieldPath) {
-	readFieldPaths(r, paths)
+	n := readFieldPaths(r, paths)
 
-	for _, fp := range *paths {
+	for _, fp := range (*paths)[:n] {
 		f := e.class.serializer.getFieldForFieldPath(fp, 0)
 		name := e.class.getNameForFieldPath(fp)
 		decoder, base := e.class.serializer.getDecoderForFieldPath2(fp, 0)
@@ -462,8 +462,6 @@ func (e *Entity) readFields(r *reader, paths *[]*fieldPath) {
 				S2:        true,
 			})
 		}
-
-		fp.release()
 	}
 }
 
@@ -486,15 +484,7 @@ func (p *Parser) OnPacketEntities(m *msgs2.CSVCMsg_PacketEntities) error {
 		p.entityFullPackets++
 	}
 
-	type tuple struct {
-		ent *Entity
-		op  st.EntityOp
-	}
-
-	var (
-		tuples []tuple
-		paths  = make([]*fieldPath, 0)
-	)
+	p.tuplesCache = p.tuplesCache[:0]
 
 	for ; updates > 0; updates-- {
 		var (
@@ -530,12 +520,10 @@ func (p *Parser) OnPacketEntities(m *msgs2.CSVCMsg_PacketEntities) error {
 
 				if baseline != nil {
 					// POV demos are missing some baselines?
-					e.readFields(newReader(baseline), &paths)
-					paths = paths[:0]
+					e.readFields(newReader(baseline), &p.pathCache)
 				}
 
-				e.readFields(r, &paths)
-				paths = paths[:0]
+				e.readFields(r, &p.pathCache)
 
 				// Fire created-handlers so update-handlers can be registered
 				for _, h := range class.createdHandlers {
@@ -559,8 +547,7 @@ func (p *Parser) OnPacketEntities(m *msgs2.CSVCMsg_PacketEntities) error {
 					op |= st.EntityOpEntered
 				}
 
-				e.readFields(r, &paths)
-				paths = paths[:0]
+				e.readFields(r, &p.pathCache)
 			}
 		} else {
 			e = p.entities[index]
@@ -583,10 +570,10 @@ func (p *Parser) OnPacketEntities(m *msgs2.CSVCMsg_PacketEntities) error {
 			}
 		}
 
-		tuples = append(tuples, tuple{e, op})
+		p.tuplesCache = append(p.tuplesCache, tuple{e, op})
 	}
 
-	for _, t := range tuples {
+	for _, t := range p.tuplesCache {
 		e := t.ent
 
 		for _, h := range p.entityHandlers {

pkg/demoinfocs/sendtables2/field_decoder.go

@@ -260,6 +260,7 @@ func quantizedFactory(f *field) fieldDecoder {
 	}
 
 	qfd := newQuantizedFloatDecoder(f.bitCount, f.encodeFlags, f.lowValue, f.highValue)
+
 	return func(r *reader) interface{} {
 		return qfd.decode(r)
 	}
@@ -274,9 +275,11 @@ func vectorFactory(n int) fieldFactory {
 		d := floatFactory(f)
 		return func(r *reader) interface{} {
 			x := make([]float32, n)
+
 			for i := 0; i < n; i++ {
 				x[i] = d(r).(float32)
 			}
+
 			return x
 		}
 	}
@@ -319,7 +322,7 @@ func ammoDecoder(r *reader) interface{} {
 }
 
 func noscaleDecoder(r *reader) interface{} {
-	return math.Float32frombits(r.readBits(32))
+	return math.Float32frombits(r.readLeUint32())
 }
 
 func runeTimeDecoder(r *reader) interface{} {

pkg/demoinfocs/sendtables2/field_path.go

@@ -306,10 +306,10 @@ func (fp *fieldPath) release() {
 }
 
 // readFieldPaths reads a new slice of fieldPath values from the given reader
-func readFieldPaths(r *reader, paths *[]*fieldPath) {
+func readFieldPaths(r *reader, paths *[]*fieldPath) int {
 	fp := newFieldPath()
-
 	node := huffTree
+	i := 0
 
 	for !fp.done {
 		var next huffmanTree
@@ -326,14 +326,26 @@ func readFieldPaths(r *reader, paths *[]*fieldPath) {
 			fieldPathTable[next.Value()].fn(r, fp)
 
 			if !fp.done {
-				*paths = append(*paths, fp.copy())
+				if len(*paths) <= i {
+					*paths = append(*paths, fp.copy())
+				} else {
+					x := (*paths)[i]
+					x.last = fp.last
+					x.done = fp.done
+
+					copy(x.path, fp.path)
+				}
+
+				i++
 			}
 		} else {
 			node = next
 		}
 	}
 
 	fp.release()
+
+	return i
 }
 
 // newHuffmanTree creates a new huffmanTree from the field path table

pkg/demoinfocs/sendtables2/field_state.go

@@ -32,11 +32,13 @@ func (s *fieldState) get(fp *fieldPath) interface{} {
 func (s *fieldState) set(fp *fieldPath, v interface{}) {
 	x := s
 	z := 0
+
 	for i := 0; i <= fp.last; i++ {
 		z = fp.path[i]
+
 		if y := len(x.state); y <= z {
 			newCap := max(z+2, y*2)
-			if newCap > cap(x.state) {
+			if z+2 > cap(x.state) {
 				newSlice := make([]interface{}, z+1, newCap)
 				copy(newSlice, x.state)
 				x.state = newSlice
@@ -45,15 +47,18 @@ func (s *fieldState) set(fp *fieldPath, v interface{}) {
 				x.state = x.state[:z+1]
 			}
 		}
+
 		if i == fp.last {
 			if _, ok := x.state[z].(*fieldState); !ok {
 				x.state[z] = v
 			}
 			return
 		}
+
 		if _, ok := x.state[z].(*fieldState); !ok {
 			x.state[z] = newFieldState()
 		}
+
 		x = x.state[z].(*fieldState)
 	}
 }

pkg/demoinfocs/sendtables2/parser.go

@@ -51,6 +51,11 @@ var itemCounts = map[string]int{
 	"MAX_ABILITY_DRAFT_ABILITIES": 48,
 }
 
+type tuple struct {
+	ent *Entity
+	op  st.EntityOp
+}
+
 type Parser struct {
 	serializers       map[string]*serializer
 	classIdSize       uint32
@@ -60,6 +65,8 @@ type Parser struct {
 	entityFullPackets int
 	entities          map[int32]*Entity
 	entityHandlers    []st.EntityHandler
+	pathCache         []*fieldPath
+	tuplesCache       []tuple
 }
 
 func (p *Parser) ReadEnterPVS(r *bit.BitReader, index int, entities map[int]st.Entity, slot int) st.Entity {

pkg/demoinfocs/sendtables2/reader.go

@@ -39,13 +39,15 @@ func (r *reader) remBytes() uint32 {
 
 // nextByte reads the next byte from the buffer
 func (r *reader) nextByte() byte {
-	r.pos++
-
-	if r.pos > r.size {
+	if r.pos >= r.size {
 		_panicf("nextByte: insufficient buffer (%d of %d)", r.pos, r.size)
 	}
 
-	return r.buf[r.pos-1]
+	x := r.buf[r.pos]
+
+	r.pos++
+
+	return x
 }
 
 // readBits returns the uint32 value for the given number of sequential bits
@@ -77,22 +79,31 @@ func (r *reader) readBytes(n uint32) []byte {
 	// Fast path if we're byte aligned
 	if r.bitCount == 0 {
 		r.pos += n
+
 		if r.pos > r.size {
 			_panicf("readBytes: insufficient buffer (%d of %d)", r.pos, r.size)
 		}
+
 		return r.buf[r.pos-n : r.pos]
 	}
 
 	buf := make([]byte, n)
+
 	for i := uint32(0); i < n; i++ {
 		buf[i] = byte(r.readBits(8))
 	}
+
 	return buf
 }
 
 // readLeUint32 reads an little-endian uint32
 func (r *reader) readLeUint32() uint32 {
-	return binary.LittleEndian.Uint32(r.readBytes(4))
+	// Fast path if we're byte aligned
+	if r.bitCount == 0 {
+		return binary.LittleEndian.Uint32(r.readBytes(4))
+	}
+
+	return r.readBits(32)
 }
 
 // readLeUint64 reads a little-endian uint64