mirror of
https://github.com/hajimehoshi/ebiten.git
synced 2024-11-10 04:57:26 +01:00
internal/ui: move dotsBuffer from internal/ui.Image to internal/buffered.Image
This is a refactoring as 'caching' things should be in internal/buffered, but also this is a performance optimization to make pixels and dotsBuffer work well together.
This commit is contained in:
Hajime Hoshi
parent
a0cb0d0568
commit
fb426cd0dd
@ -23,30 +23,52 @@ import (
|
||||
"github.com/hajimehoshi/ebiten/v2/internal/graphicsdriver"
|
||||
)
|
||||
|
||||
var whiteImage *Image
|
||||
|
||||
func init() {
|
||||
whiteImage = NewImage(3, 3, atlas.ImageTypeRegular)
|
||||
pix := make([]byte, 4*3*3)
|
||||
for i := range pix {
|
||||
pix[i] = 0xff
|
||||
}
|
||||
whiteImage.WritePixels(pix, image.Rect(0, 0, 3, 3))
|
||||
}
|
||||
|
||||
type Image struct {
|
||||
img *atlas.Image
|
||||
width int
|
||||
height int
|
||||
|
||||
// pixels is cached pixels for ReadPixels.
|
||||
pixels []byte
|
||||
|
||||
dotsBuffer map[image.Point][4]byte
|
||||
}
|
||||
|
||||
func NewImage(width, height int, imageType atlas.ImageType) *Image {
|
||||
return &Image{
|
||||
img: atlas.NewImage(width, height, imageType),
|
||||
width: width,
|
||||
height: height,
|
||||
img: atlas.NewImage(width, height, imageType),
|
||||
}
|
||||
}
|
||||
|
||||
func (i *Image) invalidatePixels() {
|
||||
i.pixels = nil
|
||||
}
|
||||
|
||||
func (i *Image) Deallocate() {
|
||||
i.img.Deallocate()
|
||||
i.dotsBuffer = nil
|
||||
}
|
||||
|
||||
func (i *Image) ReadPixels(graphicsDriver graphicsdriver.Graphics, pixels []byte, region image.Rectangle) error {
|
||||
// Do not call flushDotsBufferIfNeeded here. This would slow (image/draw).Draw.
|
||||
// See ebiten.TestImageDrawOver.
|
||||
|
||||
if region.Dx() == 1 && region.Dy() == 1 {
|
||||
if c, ok := i.dotsBuffer[region.Min]; ok {
|
||||
copy(pixels, c[:])
|
||||
return nil
|
||||
}
|
||||
}
|
||||
|
||||
if i.pixels == nil {
|
||||
pix := make([]byte, 4*i.width*i.height)
|
||||
if err := i.img.ReadPixels(graphicsDriver, pix, image.Rect(0, 0, i.width, i.height)); err != nil {
|
||||
@ -61,10 +83,25 @@ func (i *Image) ReadPixels(graphicsDriver graphicsdriver.Graphics, pixels []byte
|
||||
srcX := 4 * ((region.Min.Y+j)*i.width + region.Min.X)
|
||||
copy(pixels[dstX:dstX+lineWidth], i.pixels[srcX:srcX+lineWidth])
|
||||
}
|
||||
|
||||
// ReadPixels doesn't modify the actual pixels.
|
||||
// Just 'apply' the dots buffer to the pixels for reading.
|
||||
for pos, clr := range i.dotsBuffer {
|
||||
if !pos.In(region) {
|
||||
continue
|
||||
}
|
||||
idx := 4 * ((pos.Y-region.Min.Y)*region.Dx() + (pos.X - region.Min.X))
|
||||
pixels[idx] = clr[0]
|
||||
pixels[idx+1] = clr[1]
|
||||
pixels[idx+2] = clr[2]
|
||||
pixels[idx+3] = clr[3]
|
||||
}
|
||||
|
||||
return nil
|
||||
}
|
||||
|
||||
func (i *Image) DumpScreenshot(graphicsDriver graphicsdriver.Graphics, name string, blackbg bool) (string, error) {
|
||||
i.flushDotsBufferIfNeeded()
|
||||
return i.img.DumpScreenshot(graphicsDriver, name, blackbg)
|
||||
}
|
||||
|
||||
@ -73,7 +110,33 @@ func (i *Image) WritePixels(pix []byte, region image.Rectangle) {
|
||||
if l := 4 * region.Dx() * region.Dy(); len(pix) != l {
|
||||
panic(fmt.Sprintf("buffered: len(pix) was %d but must be %d", len(pix), l))
|
||||
}
|
||||
i.invalidatePixels()
|
||||
|
||||
// It is in theroy possible to copy the argument pixels, but this tends to consume a lot of memory.
|
||||
// Avoid this unless ReadPixels is called.
|
||||
i.pixels = nil
|
||||
|
||||
if region.Dx() == 1 && region.Dy() == 1 {
|
||||
if i.dotsBuffer == nil {
|
||||
i.dotsBuffer = map[image.Point][4]byte{}
|
||||
}
|
||||
|
||||
var clr [4]byte
|
||||
copy(clr[:], pix)
|
||||
i.dotsBuffer[region.Min] = clr
|
||||
|
||||
if len(i.dotsBuffer) >= 10000 {
|
||||
i.flushDotsBufferIfNeeded()
|
||||
}
|
||||
return
|
||||
}
|
||||
|
||||
// Remove entries in the dots buffer that are overwritten by this WritePixels call.
|
||||
for pos := range i.dotsBuffer {
|
||||
if !pos.In(region) {
|
||||
continue
|
||||
}
|
||||
delete(i.dotsBuffer, pos)
|
||||
}
|
||||
|
||||
i.img.WritePixels(pix, region)
|
||||
}
|
||||
@ -86,8 +149,15 @@ func (i *Image) DrawTriangles(srcs [graphics.ShaderImageCount]*Image, vertices [
|
||||
if i == src {
|
||||
panic("buffered: Image.DrawTriangles: source images must be different from the receiver")
|
||||
}
|
||||
if src != nil {
|
||||
// src's dots buffer should be applied to sync data in CPU and GPU,
|
||||
// but src's pixels doesn't have to be cleared since src is not modified.
|
||||
src.flushDotsBufferIfNeeded()
|
||||
}
|
||||
i.invalidatePixels()
|
||||
}
|
||||
|
||||
i.pixels = nil
|
||||
i.flushDotsBufferIfNeeded()
|
||||
|
||||
var imgs [graphics.ShaderImageCount]*atlas.Image
|
||||
for i, img := range srcs {
|
||||
@ -99,3 +169,86 @@ func (i *Image) DrawTriangles(srcs [graphics.ShaderImageCount]*Image, vertices [
|
||||
|
||||
i.img.DrawTriangles(imgs, vertices, indices, blend, dstRegion, srcRegions, shader, uniforms, fillRule)
|
||||
}
|
||||
|
||||
func (i *Image) flushDotsBufferIfNeeded() {
|
||||
if len(i.dotsBuffer) == 0 {
|
||||
return
|
||||
}
|
||||
|
||||
l := len(i.dotsBuffer)
|
||||
vs := make([]float32, l*4*graphics.VertexFloatCount)
|
||||
is := make([]uint32, l*6)
|
||||
sx, sy := float32(1), float32(1)
|
||||
var idx int
|
||||
for p, c := range i.dotsBuffer {
|
||||
dx := float32(p.X)
|
||||
dy := float32(p.Y)
|
||||
crf := float32(c[0]) / 0xff
|
||||
cgf := float32(c[1]) / 0xff
|
||||
cbf := float32(c[2]) / 0xff
|
||||
caf := float32(c[3]) / 0xff
|
||||
|
||||
vs[graphics.VertexFloatCount*4*idx] = dx
|
||||
vs[graphics.VertexFloatCount*4*idx+1] = dy
|
||||
vs[graphics.VertexFloatCount*4*idx+2] = sx
|
||||
vs[graphics.VertexFloatCount*4*idx+3] = sy
|
||||
vs[graphics.VertexFloatCount*4*idx+4] = crf
|
||||
vs[graphics.VertexFloatCount*4*idx+5] = cgf
|
||||
vs[graphics.VertexFloatCount*4*idx+6] = cbf
|
||||
vs[graphics.VertexFloatCount*4*idx+7] = caf
|
||||
vs[graphics.VertexFloatCount*4*idx+8] = dx + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+9] = dy
|
||||
vs[graphics.VertexFloatCount*4*idx+10] = sx + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+11] = sy
|
||||
vs[graphics.VertexFloatCount*4*idx+12] = crf
|
||||
vs[graphics.VertexFloatCount*4*idx+13] = cgf
|
||||
vs[graphics.VertexFloatCount*4*idx+14] = cbf
|
||||
vs[graphics.VertexFloatCount*4*idx+15] = caf
|
||||
vs[graphics.VertexFloatCount*4*idx+16] = dx
|
||||
vs[graphics.VertexFloatCount*4*idx+17] = dy + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+18] = sx
|
||||
vs[graphics.VertexFloatCount*4*idx+19] = sy + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+20] = crf
|
||||
vs[graphics.VertexFloatCount*4*idx+21] = cgf
|
||||
vs[graphics.VertexFloatCount*4*idx+22] = cbf
|
||||
vs[graphics.VertexFloatCount*4*idx+23] = caf
|
||||
vs[graphics.VertexFloatCount*4*idx+24] = dx + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+25] = dy + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+26] = sx + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+27] = sy + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+28] = crf
|
||||
vs[graphics.VertexFloatCount*4*idx+29] = cgf
|
||||
vs[graphics.VertexFloatCount*4*idx+30] = cbf
|
||||
vs[graphics.VertexFloatCount*4*idx+31] = caf
|
||||
|
||||
is[6*idx] = uint32(4 * idx)
|
||||
is[6*idx+1] = uint32(4*idx + 1)
|
||||
is[6*idx+2] = uint32(4*idx + 2)
|
||||
is[6*idx+3] = uint32(4*idx + 1)
|
||||
is[6*idx+4] = uint32(4*idx + 2)
|
||||
is[6*idx+5] = uint32(4*idx + 3)
|
||||
|
||||
idx++
|
||||
}
|
||||
|
||||
if i.pixels != nil {
|
||||
for pos, clr := range i.dotsBuffer {
|
||||
idx := 4 * (pos.Y*i.width + pos.X)
|
||||
i.pixels[idx] = clr[0]
|
||||
i.pixels[idx+1] = clr[1]
|
||||
i.pixels[idx+2] = clr[2]
|
||||
i.pixels[idx+3] = clr[3]
|
||||
delete(i.dotsBuffer, pos)
|
||||
}
|
||||
} else {
|
||||
// TODO: Use clear if Go 1.21 is available.
|
||||
for pos := range i.dotsBuffer {
|
||||
delete(i.dotsBuffer, pos)
|
||||
}
|
||||
}
|
||||
|
||||
srcs := [graphics.ShaderImageCount]*atlas.Image{whiteImage.img}
|
||||
dr := image.Rect(0, 0, i.width, i.height)
|
||||
blend := graphicsdriver.BlendCopy
|
||||
i.img.DrawTriangles(srcs, vs, is, blend, dr, [graphics.ShaderImageCount]image.Rectangle{}, atlas.NearestFilterShader, nil, graphicsdriver.FillAll)
|
||||
}
|
||||
|
||||
@ -43,8 +43,6 @@ type Image struct {
|
||||
height int
|
||||
imageType atlas.ImageType
|
||||
|
||||
dotsBuffer map[image.Point][4]byte
|
||||
|
||||
// lastBlend is the lastly-used blend for mipmap.Image.
|
||||
lastBlend graphicsdriver.Blend
|
||||
|
||||
@ -77,7 +75,6 @@ func (i *Image) Deallocate() {
|
||||
i.bigOffscreenBuffer.deallocate()
|
||||
}
|
||||
i.mipmap.Deallocate()
|
||||
i.dotsBuffer = nil
|
||||
}
|
||||
|
||||
func (i *Image) DrawTriangles(srcs [graphics.ShaderImageCount]*Image, vertices []float32, indices []uint32, blend graphicsdriver.Blend, dstRegion image.Rectangle, srcRegions [graphics.ShaderImageCount]image.Rectangle, shader *Shader, uniforms []uint32, fillRule graphicsdriver.FillRule, canSkipMipmap bool, antialias bool) {
|
||||
@ -88,9 +85,6 @@ func (i *Image) DrawTriangles(srcs [graphics.ShaderImageCount]*Image, vertices [
|
||||
i.lastBlend = blend
|
||||
|
||||
if antialias {
|
||||
// Flush the other buffer to make the buffers exclusive.
|
||||
i.flushDotsBufferIfNeeded()
|
||||
|
||||
if i.bigOffscreenBuffer == nil {
|
||||
var imageType atlas.ImageType
|
||||
switch i.imageType {
|
||||
@ -126,25 +120,6 @@ func (i *Image) WritePixels(pix []byte, region image.Rectangle) {
|
||||
if i.modifyCallback != nil {
|
||||
i.modifyCallback()
|
||||
}
|
||||
|
||||
if region.Dx() == 1 && region.Dy() == 1 {
|
||||
// Flush the other buffer to make the buffers exclusive.
|
||||
i.flushBigOffscreenBufferIfNeeded()
|
||||
|
||||
if i.dotsBuffer == nil {
|
||||
i.dotsBuffer = map[image.Point][4]byte{}
|
||||
}
|
||||
|
||||
var clr [4]byte
|
||||
copy(clr[:], pix)
|
||||
i.dotsBuffer[region.Min] = clr
|
||||
|
||||
if len(i.dotsBuffer) >= 10000 {
|
||||
i.flushDotsBufferIfNeeded()
|
||||
}
|
||||
return
|
||||
}
|
||||
|
||||
i.flushBufferIfNeeded()
|
||||
i.mipmap.WritePixels(pix, region)
|
||||
}
|
||||
@ -157,17 +132,6 @@ func (i *Image) ReadPixels(pixels []byte, region image.Rectangle) {
|
||||
|
||||
i.flushBigOffscreenBufferIfNeeded()
|
||||
|
||||
if region.Dx() == 1 && region.Dy() == 1 {
|
||||
if c, ok := i.dotsBuffer[region.Min]; ok {
|
||||
copy(pixels, c[:])
|
||||
return
|
||||
}
|
||||
// Do not call flushDotsBufferIfNeeded here. This would slow (image/draw).Draw.
|
||||
// See ebiten.TestImageDrawOver.
|
||||
} else {
|
||||
i.flushDotsBufferIfNeeded()
|
||||
}
|
||||
|
||||
if err := i.ui.readPixels(i.mipmap, pixels, region); err != nil {
|
||||
if panicOnErrorOnReadingPixels {
|
||||
panic(err)
|
||||
@ -182,80 +146,9 @@ func (i *Image) DumpScreenshot(name string, blackbg bool) (string, error) {
|
||||
}
|
||||
|
||||
func (i *Image) flushBufferIfNeeded() {
|
||||
// The buffers are exclusive and the order should not matter.
|
||||
i.flushDotsBufferIfNeeded()
|
||||
i.flushBigOffscreenBufferIfNeeded()
|
||||
}
|
||||
|
||||
func (i *Image) flushDotsBufferIfNeeded() {
|
||||
if len(i.dotsBuffer) == 0 {
|
||||
return
|
||||
}
|
||||
|
||||
l := len(i.dotsBuffer)
|
||||
vs := make([]float32, l*4*graphics.VertexFloatCount)
|
||||
is := make([]uint32, l*6)
|
||||
sx, sy := float32(1), float32(1)
|
||||
var idx int
|
||||
for p, c := range i.dotsBuffer {
|
||||
dx := float32(p.X)
|
||||
dy := float32(p.Y)
|
||||
crf := float32(c[0]) / 0xff
|
||||
cgf := float32(c[1]) / 0xff
|
||||
cbf := float32(c[2]) / 0xff
|
||||
caf := float32(c[3]) / 0xff
|
||||
|
||||
vs[graphics.VertexFloatCount*4*idx] = dx
|
||||
vs[graphics.VertexFloatCount*4*idx+1] = dy
|
||||
vs[graphics.VertexFloatCount*4*idx+2] = sx
|
||||
vs[graphics.VertexFloatCount*4*idx+3] = sy
|
||||
vs[graphics.VertexFloatCount*4*idx+4] = crf
|
||||
vs[graphics.VertexFloatCount*4*idx+5] = cgf
|
||||
vs[graphics.VertexFloatCount*4*idx+6] = cbf
|
||||
vs[graphics.VertexFloatCount*4*idx+7] = caf
|
||||
vs[graphics.VertexFloatCount*4*idx+8] = dx + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+9] = dy
|
||||
vs[graphics.VertexFloatCount*4*idx+10] = sx + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+11] = sy
|
||||
vs[graphics.VertexFloatCount*4*idx+12] = crf
|
||||
vs[graphics.VertexFloatCount*4*idx+13] = cgf
|
||||
vs[graphics.VertexFloatCount*4*idx+14] = cbf
|
||||
vs[graphics.VertexFloatCount*4*idx+15] = caf
|
||||
vs[graphics.VertexFloatCount*4*idx+16] = dx
|
||||
vs[graphics.VertexFloatCount*4*idx+17] = dy + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+18] = sx
|
||||
vs[graphics.VertexFloatCount*4*idx+19] = sy + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+20] = crf
|
||||
vs[graphics.VertexFloatCount*4*idx+21] = cgf
|
||||
vs[graphics.VertexFloatCount*4*idx+22] = cbf
|
||||
vs[graphics.VertexFloatCount*4*idx+23] = caf
|
||||
vs[graphics.VertexFloatCount*4*idx+24] = dx + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+25] = dy + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+26] = sx + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+27] = sy + 1
|
||||
vs[graphics.VertexFloatCount*4*idx+28] = crf
|
||||
vs[graphics.VertexFloatCount*4*idx+29] = cgf
|
||||
vs[graphics.VertexFloatCount*4*idx+30] = cbf
|
||||
vs[graphics.VertexFloatCount*4*idx+31] = caf
|
||||
|
||||
is[6*idx] = uint32(4 * idx)
|
||||
is[6*idx+1] = uint32(4*idx + 1)
|
||||
is[6*idx+2] = uint32(4*idx + 2)
|
||||
is[6*idx+3] = uint32(4*idx + 1)
|
||||
is[6*idx+4] = uint32(4*idx + 2)
|
||||
is[6*idx+5] = uint32(4*idx + 3)
|
||||
|
||||
idx++
|
||||
}
|
||||
i.dotsBuffer = nil
|
||||
|
||||
srcs := [graphics.ShaderImageCount]*mipmap.Mipmap{i.ui.whiteImage.mipmap}
|
||||
dr := image.Rect(0, 0, i.width, i.height)
|
||||
blend := graphicsdriver.BlendCopy
|
||||
i.lastBlend = blend
|
||||
i.mipmap.DrawTriangles(srcs, vs, is, blend, dr, [graphics.ShaderImageCount]image.Rectangle{}, NearestFilterShader.shader, nil, graphicsdriver.FillAll, true)
|
||||
}
|
||||
|
||||
func (i *Image) flushBigOffscreenBufferIfNeeded() {
|
||||
if i.bigOffscreenBuffer != nil {
|
||||
i.bigOffscreenBuffer.flush()
|
||||
|
||||
Loading…
Reference in New Issue
Block a user