internal/graphicsdriver/metal: Reuse MTLBuffer objects

In Metal, MTLBuffer objects are not 'transient' and are expensive
to create. Reuse them whenever possible.

See also: https://developer.apple.com/library/archive/documentation/Miscellaneous/Conceptual/MetalProgrammingGuide/Cmd-Submiss/Cmd-Submiss.html

Updates #1196

internal/graphicsdriver/metal/graphics.go

@@ -20,6 +20,7 @@ package metal
 import (
 	"fmt"
 	"math"
+	"sort"
 	"strings"
 	"unsafe"
 
@@ -316,6 +317,9 @@ type Graphics struct {
 
 	screenDrawable ca.MetalDrawable
 
+	buffers       map[mtl.CommandBuffer][]mtl.Buffer
+	unusedBuffers map[mtl.Buffer]struct{}
+
 	lastDstTexture  mtl.Texture
 	lastStencilMode stencilMode
 
@@ -376,15 +380,90 @@ func (g *Graphics) SetUIView(uiview uintptr) {
 	g.view.setUIView(uiview)
 }
 
+func pow2(x uintptr) uintptr {
+	var p2 uintptr = 1
+	for p2 < x {
+		p2 *= 2
+	}
+	return p2
+}
+
+func (g *Graphics) availableBuffer(length uintptr) mtl.Buffer {
+	if g.cb == (mtl.CommandBuffer{}) {
+		g.cb = g.cq.MakeCommandBuffer()
+	}
+
+	var newBuf mtl.Buffer
+
+	for cb, bs := range g.buffers {
+		// If the command buffer still lives, the buffer must not be updated.
+		// TODO: Handle an error?
+		if cb.Status() != mtl.CommandBufferStatusCompleted {
+			continue
+		}
+
+		for _, b := range bs {
+			if newBuf == (mtl.Buffer{}) && b.Length() >= length {
+				newBuf = b
+				continue
+			}
+			if g.unusedBuffers == nil {
+				g.unusedBuffers = map[mtl.Buffer]struct{}{}
+			}
+			g.unusedBuffers[b] = struct{}{}
+		}
+		delete(g.buffers, cb)
+		cb.Release()
+	}
+
+	for b := range g.unusedBuffers {
+		if b.Length() >= length {
+			newBuf = b
+			delete(g.unusedBuffers, b)
+			break
+		}
+	}
+
+	// GC unused buffers.
+	const maxUnusedBuffers = 20
+	if len(g.unusedBuffers) > maxUnusedBuffers {
+		bufs := make([]mtl.Buffer, 0, len(g.unusedBuffers))
+		for b := range g.unusedBuffers {
+			bufs = append(bufs, b)
+		}
+		sort.Slice(bufs, func(a, b int) bool {
+			return bufs[a].Length() < bufs[b].Length()
+		})
+		for _, b := range bufs[maxUnusedBuffers:] {
+			println("delete!", b.Length())
+			delete(g.unusedBuffers, b)
+			b.Release()
+		}
+	}
+
+	if newBuf == (mtl.Buffer{}) {
+		newBuf = g.view.getMTLDevice().MakeBufferWithLength(pow2(length), resourceStorageMode)
+	}
+
+	if g.buffers == nil {
+		g.buffers = map[mtl.CommandBuffer][]mtl.Buffer{}
+	}
+	if _, ok := g.buffers[g.cb]; !ok {
+		g.cb.Retain()
+	}
+	g.buffers[g.cb] = append(g.buffers[g.cb], newBuf)
+	return newBuf
+}
+
 func (g *Graphics) SetVertices(vertices []float32, indices []uint16) {
-	if g.vb != (mtl.Buffer{}) {
-		g.vb.Release()
-	}
-	if g.ib != (mtl.Buffer{}) {
-		g.ib.Release()
-	}
-	g.vb = g.view.getMTLDevice().MakeBufferWithBytes(unsafe.Pointer(&vertices[0]), unsafe.Sizeof(vertices[0])*uintptr(len(vertices)), resourceStorageMode)
-	g.ib = g.view.getMTLDevice().MakeBufferWithBytes(unsafe.Pointer(&indices[0]), unsafe.Sizeof(indices[0])*uintptr(len(indices)), resourceStorageMode)
+	vbSize := unsafe.Sizeof(vertices[0]) * uintptr(len(vertices))
+	ibSize := unsafe.Sizeof(indices[0]) * uintptr(len(indices))
+
+	g.vb = g.availableBuffer(vbSize)
+	g.vb.CopyToContents(unsafe.Pointer(&vertices[0]), vbSize)
+
+	g.ib = g.availableBuffer(ibSize)
+	g.ib.CopyToContents(unsafe.Pointer(&indices[0]), ibSize)
 }
 
 func (g *Graphics) flushIfNeeded(present bool) {

internal/graphicsdriver/metal/mtl/mtl.go

@@ -330,6 +330,17 @@ const (
 	CompareFunctionAlways       CompareFunction = 7
 )
 
+type CommandBufferStatus uint8
+
+const (
+	CommandBufferStatusNotEnqueued CommandBufferStatus = 0 //The command buffer is not enqueued yet.
+	CommandBufferStatusEnqueued    CommandBufferStatus = 1 // The command buffer is enqueued.
+	CommandBufferStatusCommitted   CommandBufferStatus = 2 // The command buffer is committed for execution.
+	CommandBufferStatusScheduled   CommandBufferStatus = 3 // The command buffer is scheduled.
+	CommandBufferStatusCompleted   CommandBufferStatus = 4 // The command buffer completed execution successfully.
+	CommandBufferStatusError       CommandBufferStatus = 5 // Execution of the command buffer was aborted due to an error during execution.
+)
+
 // Resource represents a memory allocation for storing specialized data
 // that is accessible to the GPU.
 //
@@ -623,6 +634,21 @@ type CommandBuffer struct {
 	commandBuffer unsafe.Pointer
 }
 
+func (cb CommandBuffer) Retain() {
+	C.CommandBuffer_Retain(cb.commandBuffer)
+}
+
+func (cb CommandBuffer) Release() {
+	C.CommandBuffer_Release(cb.commandBuffer)
+}
+
+// Status returns the current stage in the lifetime of the command buffer.
+//
+// Reference: https://developer.apple.com/documentation/metal/mtlcommandbuffer/1443048-status
+func (cb CommandBuffer) Status() CommandBufferStatus {
+	return CommandBufferStatus(C.CommandBuffer_Status(cb.commandBuffer))
+}
+
 // PresentDrawable registers a drawable presentation to occur as soon as possible.
 //
 // Reference: https://developer.apple.com/documentation/metal/mtlcommandbuffer/1443029-presentdrawable.
@@ -873,6 +899,12 @@ type Buffer struct {
 	buffer unsafe.Pointer
 }
 
+func (b Buffer) resource() unsafe.Pointer { return b.buffer }
+
+func (b Buffer) Length() uintptr {
+	return uintptr(C.Buffer_Length(b.buffer))
+}
+
 func (b Buffer) CopyToContents(data unsafe.Pointer, lengthInBytes uintptr) {
 	C.Buffer_CopyToContents(b.buffer, data, C.size_t(lengthInBytes))
 }

internal/graphicsdriver/metal/mtl/mtl.h

@@ -147,7 +147,9 @@ void *Device_MakeDepthStencilState(void *device,
 void CommandQueue_Release(void *commandQueue);
 void *CommandQueue_MakeCommandBuffer(void *commandQueue);
 
+void CommandBuffer_Retain(void *commandBuffer);
 void CommandBuffer_Release(void *commandBuffer);
+uint8_t CommandBuffer_Status(void *commandBuffer);
 void CommandBuffer_PresentDrawable(void *commandBuffer, void *drawable);
 void CommandBuffer_Commit(void *commandBuffer);
 void CommandBuffer_WaitUntilCompleted(void *commandBuffer);
@@ -208,6 +210,7 @@ void Texture_ReplaceRegion(void *texture, struct Region region, uint_t level,
 int Texture_Width(void *texture);
 int Texture_Height(void *texture);
 
+size_t Buffer_Length(void *buffer);
 void Buffer_CopyToContents(void *buffer, void *data, size_t lengthInBytes);
 void Buffer_Retain(void *buffer);
 void Buffer_Release(void *buffer);

internal/graphicsdriver/metal/mtl/mtl.m

@@ -163,6 +163,18 @@ void *CommandQueue_MakeCommandBuffer(void *commandQueue) {
   return [(id<MTLCommandQueue>)commandQueue commandBuffer];
 }
 
+void CommandBuffer_Retain(void *commandBuffer) {
+  [(id<MTLCommandBuffer>)commandBuffer retain];
+}
+
+void CommandBuffer_Release(void *commandBuffer) {
+  [(id<MTLCommandBuffer>)commandBuffer release];
+}
+
+uint8_t CommandBuffer_Status(void *commandBuffer) {
+  return [(id<MTLCommandBuffer>)commandBuffer status];
+}
+
 void CommandBuffer_PresentDrawable(void *commandBuffer, void *drawable) {
   [(id<MTLCommandBuffer>)commandBuffer
       presentDrawable:(id<MTLDrawable>)drawable];
@@ -401,8 +413,13 @@ int Texture_Width(void *texture) { return [(id<MTLTexture>)texture width]; }
 
 int Texture_Height(void *texture) { return [(id<MTLTexture>)texture height]; }
 
+size_t Buffer_Length(void *buffer) { return [(id<MTLBuffer>)buffer length]; }
+
 void Buffer_CopyToContents(void *buffer, void *data, size_t lengthInBytes) {
   memcpy(((id<MTLBuffer>)buffer).contents, data, lengthInBytes);
+#if !TARGET_OS_IPHONE
+  [(id<MTLBuffer>)buffer didModifyRange:NSMakeRange(0, lengthInBytes)];
+#endif
 }
 
 void Buffer_Retain(void *buffer) { [(id<MTLBuffer>)buffer retain]; }