graphicsdriver/metal: Avoid synching texture data at ReplacePixels

Instead, use a temporary texture, transfer texture data to it, and
copy the data on GPU side.

internal/graphicsdriver/metal/graphics.go

@@ -326,7 +326,7 @@ type Graphics struct {
 
 	transparent  bool
 	maxImageSize int
-	drawCalled   bool
+	tmpTexture   mtl.Texture
 
 	t *thread.Thread
 
@@ -353,7 +353,7 @@ func (g *Graphics) Begin() {
 }
 
 func (g *Graphics) End() {
-	g.flush(false, true)
+	g.flushIfNeeded(false, true)
 	g.t.Call(func() error {
 		g.screenDrawable = ca.MetalDrawable{}
 		C.releaseAutoreleasePool(g.pool)
@@ -390,7 +390,7 @@ func (g *Graphics) SetVertices(vertices []float32, indices []uint16) {
 	})
 }
 
-func (g *Graphics) flush(wait bool, present bool) {
+func (g *Graphics) flushIfNeeded(wait bool, present bool) {
 	g.t.Call(func() error {
 		if g.cb == (mtl.CommandBuffer{}) {
 			return nil
@@ -639,8 +639,6 @@ func (g *Graphics) Draw(dstID, srcID driver.ImageID, indexLen int, indexOffset i
 	dst := g.images[dstID]
 	src := g.images[srcID]
 
-	g.drawCalled = true
-
 	if err := g.t.Call(func() error {
 		g.view.update()
 
@@ -841,7 +839,7 @@ func (i *Image) syncTexture() {
 
 	i.graphics.t.Call(func() error {
 		if i.graphics.cb != (mtl.CommandBuffer{}) {
-			panic("metal: command buffer must be empty at syncTexture: flush is not called yet?")
+			panic("metal: command buffer must be empty at syncTexture: flushIfNeeded is not called yet?")
 		}
 
 		cb := i.graphics.cq.MakeCommandBuffer()
@@ -855,7 +853,7 @@ func (i *Image) syncTexture() {
 }
 
 func (i *Image) Pixels() ([]byte, error) {
-	i.graphics.flush(true, false)
+	i.graphics.flushIfNeeded(true, false)
 	i.syncTexture()
 
 	b := make([]byte, 4*i.width*i.height)
@@ -868,33 +866,12 @@ func (i *Image) Pixels() ([]byte, error) {
 	return b, nil
 }
 
-// needsToSyncBeforeReplaceRegion reports whether syncTexture is needed before calling ReplaceRegion.
-//
-// Before a part region is updated, the texture data needs to be loaded to CPU. Otherwise, the remaining region will
-// be uninitialized (#1213).
-func (i *Image) needsToSyncBeforeReplaceRegion(args []*driver.ReplacePixelsArgs) bool {
-	if len(args) == 0 {
-		return false
-	}
-	if len(args) > 1 {
-		return true
-	}
-	if a := args[0]; a.X == 0 && a.Y == 0 && a.Width == i.width && a.Height == i.height {
-		return false
-	}
-	return true
-}
-
 func (i *Image) ReplacePixels(args []*driver.ReplacePixelsArgs) {
 	g := i.graphics
-	if g.drawCalled {
+	g.flushIfNeeded(true, false)
-		g.flush(true, false)
-		g.drawCalled = false
-		if i.needsToSyncBeforeReplaceRegion(args) {
-			i.syncTexture()
-		}
-	}
 
+	// If the memory is shared (e.g., iOS), texture data doen't have to be synced. Send the data directly.
+	if storageMode == mtl.StorageModeShared {
 		g.t.Call(func() error {
 			for _, a := range args {
 				i.texture.ReplaceRegion(mtl.Region{
@@ -904,4 +881,47 @@ func (i *Image) ReplacePixels(args []*driver.ReplacePixelsArgs) {
 			}
 			return nil
 		})
+		return
+	}
+
+	// If the memory is managed (e.g., macOS), texture data cannot be sent to the destination directly because
+	// this requires synchronizing data between CPU and GPU. As synchronizing is inefficient, let's send the
+	// data to a temporary texture once, and then copy it in GPU.
+	g.t.Call(func() error {
+		w, h := i.texture.Width(), i.texture.Height()
+		if g.tmpTexture == (mtl.Texture{}) || w > g.tmpTexture.Width() || h > g.tmpTexture.Height() {
+			if g.tmpTexture != (mtl.Texture{}) {
+				g.tmpTexture.Release()
+			}
+			td := mtl.TextureDescriptor{
+				TextureType: mtl.TextureType2D,
+				PixelFormat: mtl.PixelFormatRGBA8UNorm,
+				Width:       w,
+				Height:      h,
+				StorageMode: storageMode,
+				Usage:       mtl.TextureUsageShaderRead | mtl.TextureUsageRenderTarget,
+			}
+			g.tmpTexture = g.view.getMTLDevice().MakeTexture(td)
+		}
+
+		for _, a := range args {
+			g.tmpTexture.ReplaceRegion(mtl.Region{
+				Origin: mtl.Origin{X: a.X, Y: a.Y, Z: 0},
+				Size:   mtl.Size{Width: a.Width, Height: a.Height, Depth: 1},
+			}, 0, unsafe.Pointer(&a.Pixels[0]), 4*a.Width)
+		}
+
+		if g.cb == (mtl.CommandBuffer{}) {
+			g.cb = i.graphics.cq.MakeCommandBuffer()
+		}
+		bce := g.cb.MakeBlitCommandEncoder()
+		for _, a := range args {
+			o := mtl.Origin{X: a.X, Y: a.Y, Z: 0}
+			s := mtl.Size{Width: a.Width, Height: a.Height, Depth: 1}
+			bce.CopyFromTexture(g.tmpTexture, 0, 0, o, s, i.texture, 0, 0, o)
+		}
+		bce.EndEncoding()
+
+		return nil
+	})
 }

internal/graphicsdriver/metal/mtl/example_test.go

@@ -182,9 +182,9 @@ fragment float4 FragmentShader(Vertex in [[stage_in]]) {
 	cb.WaitUntilCompleted()
 
 	// Read pixels from output texture into an image.
-	img := image.NewNRGBA(image.Rect(0, 0, texture.Width, texture.Height))
+	img := image.NewNRGBA(image.Rect(0, 0, texture.Width(), texture.Height()))
-	bytesPerRow := 4 * texture.Width
+	bytesPerRow := 4 * texture.Width()
-	region := mtl.RegionMake2D(0, 0, texture.Width, texture.Height)
+	region := mtl.RegionMake2D(0, 0, texture.Width(), texture.Height())
 	texture.GetBytes(&img.Pix[0], uintptr(bytesPerRow), region, 0)
 
 	// Output image to stdout as grayscale ASCII art.

internal/graphicsdriver/metal/mtl/mtl.go

@@ -501,8 +501,6 @@ func (d Device) MakeTexture(td TextureDescriptor) Texture {
 	}
 	return Texture{
 		texture: C.Device_MakeTexture(d.device, descriptor),
-		Width:   td.Width,  // TODO: Fetch dimensions of actually created texture.
-		Height:  td.Height, // TODO: Fetch dimensions of actually created texture.
 	}
 }
 
@@ -700,6 +698,10 @@ func (bce BlitCommandEncoder) SynchronizeTexture(texture Texture, slice int, lev
 	C.BlitCommandEncoder_SynchronizeTexture(bce.commandEncoder, texture.texture, C.uint_t(slice), C.uint_t(level))
 }
 
+func (bce BlitCommandEncoder) CopyFromTexture(sourceTexture Texture, sourceSlice int, sourceLevel int, sourceOrigin Origin, sourceSize Size, destinationTexture Texture, destinationSlice int, destinationLevel int, destinationOrigin Origin) {
+	C.BlitCommandEncoder_CopyFromTexture(bce.commandEncoder, sourceTexture.texture, C.uint_t(sourceSlice), C.uint_t(sourceLevel), sourceOrigin.c(), sourceSize.c(), destinationTexture.texture, C.uint_t(destinationSlice), C.uint_t(destinationLevel), destinationOrigin.c())
+}
+
 // Library is a collection of compiled graphics or compute functions.
 //
 // Reference: https://developer.apple.com/documentation/metal/mtllibrary.
@@ -725,14 +727,6 @@ func (l Library) MakeFunction(name string) (Function, error) {
 // Reference: https://developer.apple.com/documentation/metal/mtltexture.
 type Texture struct {
 	texture unsafe.Pointer
-
-	// TODO: Change these fields into methods.
-
-	// Width is the width of the texture image for the base level mipmap, in pixels.
-	Width int
-
-	// Height is the height of the texture image for the base level mipmap, in pixels.
-	Height int
 }
 
 // NewTexture returns a Texture that wraps an existing id<MTLTexture> pointer.
@@ -764,6 +758,20 @@ func (t Texture) ReplaceRegion(region Region, level int, pixelBytes unsafe.Point
 	C.Texture_ReplaceRegion(t.texture, r, C.uint_t(level), pixelBytes, C.uint_t(bytesPerRow))
 }
 
+// Width is the width of the texture image for the base level mipmap, in pixels.
+//
+// Reference: https://developer.apple.com/documentation/metal/mtltexture/1515339-width
+func (t Texture) Width() int {
+	return int(C.Texture_Width(t.texture))
+}
+
+// Height is the height of the texture image for the base level mipmap, in pixels.
+//
+// Reference: https://developer.apple.com/documentation/metal/mtltexture/1515938-height
+func (t Texture) Height() int {
+	return int(C.Texture_Height(t.texture))
+}
+
 // Buffer is a memory allocation for storing unformatted data
 // that is accessible to the GPU.
 //
@@ -814,16 +822,8 @@ type Region struct {
 
 func (r *Region) c() C.struct_Region {
 	return C.struct_Region{
-		Origin: C.struct_Origin{
+		Origin: r.Origin.c(),
-			X: C.uint_t(r.Origin.X),
+		Size:   r.Size.c(),
-			Y: C.uint_t(r.Origin.Y),
-			Z: C.uint_t(r.Origin.Z),
-		},
-		Size: C.struct_Size{
-			Width:  C.uint_t(r.Size.Width),
-			Height: C.uint_t(r.Size.Height),
-			Depth:  C.uint_t(r.Size.Depth),
-		},
 	}
 }
 
@@ -833,12 +833,28 @@ func (r *Region) c() C.struct_Region {
 // Reference: https://developer.apple.com/documentation/metal/mtlorigin.
 type Origin struct{ X, Y, Z int }
 
+func (o *Origin) c() C.struct_Origin {
+	return C.struct_Origin{
+		X: C.uint_t(o.X),
+		Y: C.uint_t(o.Y),
+		Z: C.uint_t(o.Z),
+	}
+}
+
 // Size represents the set of dimensions that declare the size of an object,
 // such as an image, texture, threadgroup, or grid.
 //
 // Reference: https://developer.apple.com/documentation/metal/mtlsize.
 type Size struct{ Width, Height, Depth int }
 
+func (s *Size) c() C.struct_Size {
+	return C.struct_Size{
+		Width:  C.uint_t(s.Width),
+		Height: C.uint_t(s.Height),
+		Depth:  C.uint_t(s.Depth),
+	}
+}
+
 // RegionMake2D returns a 2D, rectangular region for image or texture data.
 //
 // Reference: https://developer.apple.com/documentation/metal/1515675-mtlregionmake2d.

internal/graphicsdriver/metal/mtl/mtl.h

@@ -163,6 +163,11 @@ void BlitCommandEncoder_Synchronize(void *blitCommandEncoder, void *resource);
 void BlitCommandEncoder_SynchronizeTexture(void *blitCommandEncoder,
                                            void *texture, uint_t slice,
                                            uint_t level);
+void BlitCommandEncoder_CopyFromTexture(
+    void *blitCommandEncoder, void *sourceTexture, uint_t sourceSlice,
+    uint_t sourceLevel, struct Origin sourceOrigin, struct Size sourceSize,
+    void *destinationTexture, uint_t destinationSlice, uint_t destinationLevel,
+    struct Origin destinationOrigin);
 
 void *Library_MakeFunction(void *library, const char *name);
 
@@ -171,6 +176,8 @@ void Texture_GetBytes(void *texture, void *pixelBytes, size_t bytesPerRow,
                       struct Region region, uint_t level);
 void Texture_ReplaceRegion(void *texture, struct Region region, uint_t level,
                            void *pixelBytes, uint_t bytesPerRow);
+int Texture_Width(void *texture);
+int Texture_Height(void *texture);
 
 void Buffer_CopyToContents(void *buffer, void *data, size_t lengthInBytes);
 void Buffer_Retain(void *buffer);

internal/graphicsdriver/metal/mtl/mtl.m

@@ -280,6 +280,26 @@ void BlitCommandEncoder_SynchronizeTexture(void *blitCommandEncoder,
 #endif
 }
 
+void BlitCommandEncoder_CopyFromTexture(
+    void *blitCommandEncoder, void *sourceTexture, uint_t sourceSlice,
+    uint_t sourceLevel, struct Origin sourceOrigin, struct Size sourceSize,
+    void *destinationTexture, uint_t destinationSlice, uint_t destinationLevel,
+    struct Origin destinationOrigin) {
+  [(id<MTLBlitCommandEncoder>)blitCommandEncoder
+        copyFromTexture:(id<MTLTexture>)sourceTexture
+            sourceSlice:(NSUInteger)sourceSlice
+            sourceLevel:(NSUInteger)sourceLevel
+           sourceOrigin:(MTLOrigin){sourceOrigin.X, sourceOrigin.Y,
+                                    sourceOrigin.Z}
+             sourceSize:(MTLSize){sourceSize.Width, sourceSize.Height,
+                                  sourceSize.Depth}
+              toTexture:(id<MTLTexture>)destinationTexture
+       destinationSlice:(NSUInteger)destinationSlice
+       destinationLevel:(NSUInteger)destinationLevel
+      destinationOrigin:(MTLOrigin){destinationOrigin.X, destinationOrigin.Y,
+                                    destinationOrigin.Z}];
+}
+
 void *Library_MakeFunction(void *library, const char *name) {
   return [(id<MTLLibrary>)library
       newFunctionWithName:[NSString stringWithUTF8String:name]];
@@ -312,6 +332,10 @@ void Texture_ReplaceRegion(void *texture, struct Region region, uint_t level,
                              bytesPerRow:(NSUInteger)bytesPerRow];
 }
 
+int Texture_Width(void *texture) { return [(id<MTLTexture>)texture width]; }
+
+int Texture_Height(void *texture) { return [(id<MTLTexture>)texture height]; }
+
 void Buffer_CopyToContents(void *buffer, void *data, size_t lengthInBytes) {
   memcpy(((id<MTLBuffer>)buffer).contents, data, lengthInBytes);
 }

internal/graphicsdriver/metal/mtl/mtl_test.go

@@ -125,9 +125,9 @@ fragment float4 FragmentShader(Vertex in [[stage_in]]) {
 	cb.WaitUntilCompleted()
 
 	// Read pixels from output texture into an image.
-	got := image.NewNRGBA(image.Rect(0, 0, texture.Width, texture.Height))
+	got := image.NewNRGBA(image.Rect(0, 0, texture.Width(), texture.Height()))
-	bytesPerRow := 4 * texture.Width
+	bytesPerRow := 4 * texture.Width()
-	region := mtl.RegionMake2D(0, 0, texture.Width, texture.Height)
+	region := mtl.RegionMake2D(0, 0, texture.Width(), texture.Height())
 	texture.GetBytes(&got.Pix[0], uintptr(bytesPerRow), region, 0)
 
 	// TODO: Embed this file?