all: speed optimization

internal/graphics/vertex.go

@@ -36,6 +36,15 @@ const (
 	TextureSourceRegionOriginUniformVariableIndex      = 5
 	TextureSourceRegionSizeUniformVariableIndex        = 6
 	ProjectionMatrixUniformVariableIndex               = 7
+
+	PreservedUniformUint32Count = 2 + // the destination texture size
+		2*ShaderImageCount + // the texture sizes array
+		2 + // the texture destination region's origin
+		2 + // the texture destination region's size
+		2*(ShaderImageCount-1) + // the offsets array of the second and the following images
+		2 + // the texture source region's origin
+		2 + // the texture source region's size
+		16 // the projection matrix
 )
 
 const (

internal/graphicscommand/command.go

@@ -580,14 +580,9 @@ func roundUpPower2(x int) int {
 }
 
 func (q *commandQueue) prependPreservedUniforms(uniforms []uint32, shader *Shader, dst *Image, srcs [graphics.ShaderImageCount]*Image, offsets [graphics.ShaderImageCount - 1][2]float32, dstRegion, srcRegion graphicsdriver.Region) []uint32 {
-	var n int
-	for _, typ := range shader.ir.Uniforms[:graphics.PreservedUniformVariablesCount] {
-		n += typ.Uint32Count()
-	}
-
 	origUniforms := uniforms
-	uniforms = q.uint32sBuffer.alloc(len(origUniforms) + n)
-	copy(uniforms[n:], origUniforms)
+	uniforms = q.uint32sBuffer.alloc(len(origUniforms) + graphics.PreservedUniformUint32Count)
+	copy(uniforms[graphics.PreservedUniformUint32Count:], origUniforms)
 
 	var idx int
 

internal/shaderir/program.go

@@ -32,7 +32,8 @@ type Program struct {
 	VertexFunc   VertexFunc
 	FragmentFunc FragmentFunc
 
-	reachableUniforms []bool
+	reachableUniforms   []bool
+	uniformUint32Counts []int
 }
 
 type Func struct {
@@ -473,9 +474,15 @@ func (p *Program) FilterUniformVariables(uniforms []uint32) {
 		}
 	}
 
+	if p.uniformUint32Counts == nil {
+		p.uniformUint32Counts = make([]int, len(p.Uniforms))
+		for i, typ := range p.Uniforms {
+			p.uniformUint32Counts[i] = typ.Uint32Count()
+		}
+	}
+
 	var idx int
-	for i, typ := range p.Uniforms {
-		n := typ.Uint32Count()
+	for i, n := range p.uniformUint32Counts {
 		if !p.reachableUniforms[i] {
 			for j := 0; j < n; j++ {
 				uniforms[idx+j] = 0