macOS Metal中CAMetalLayer的nextDrawable方法性能异常缓慢

问题背景 #

基于macOS Metal编写的渲染代码（在@autoreleasepool块中调用），预期可处理数百万三角形，但实际仅支持数千量级。当三角形数量达30000时，nextDrawable调用耗时0.4秒，性能严重不达预期。相关代码如下：

setupEncoder方法 #

- (void)setupEncoder:(nonnull MTKView *)view  {
    id<MTLCommandBuffer> commandBuffer = nil;
    id<CAMetalDrawable> drawable = nil;
    MTLRenderPassDescriptor *renderPass = nil;
    id<MTLRenderPipelineState> pipeline = nil;
    
    MTLRenderPassDescriptor *renderPassDescriptor = view.currentRenderPassDescriptor;
    
    if (renderPassDescriptor == nil)
    {
        return;
    }
  
    commandBuffer = [_commandQueue commandBuffer];

// VERY SLOW
    drawable = [_metalLayer nextDrawable];
// ---------
    if (drawable == nil) {
        return;
    }
    

    id<MTLTexture> framebufferTexture = drawable.texture;
    
    
    renderPass = [MTLRenderPassDescriptor renderPassDescriptor];
    renderPass.colorAttachments[0].clearColor = self.clearColor;
    renderPass.colorAttachments[0].storeAction = MTLStoreActionStore;
    renderPass.colorAttachments[0].loadAction = MTLLoadActionClear;
    
    renderPass.colorAttachments[0].texture = framebufferTexture;
    renderPass.depthAttachment = _depthAttachmentDescriptor;
    
    
    
    pipeline = [self.device newRenderPipelineStateWithDescriptor:_renderPipelineDescriptor
                                                               error:NULL];

    
    id<MTLRenderCommandEncoder> commandEncoder = [commandBuffer renderCommandEncoderWithDescriptor:renderPass];
    [commandEncoder setRenderPipelineState:pipeline];
    
    
    // Set up the depth buffer
    [commandEncoder setDepthStencilState:_depthStencilState];
    [commandEncoder setFrontFacingWinding:MTLWindingCounterClockwise];
    [commandEncoder setCullMode:MTLCullModeBack];
    
    // Draw root node
    [self drawKSCNNode:commandEncoder node:scene.rootNode uniforms:uniformSet
      startTranslation:SCNVector3Make(0, 0, 0) startRotation:SCNVector4Make(0, 1, 0, 0) startScale:SCNVector3Make(1, 1, 1)];
    
    [commandEncoder endEncoding];
    
    
    // Get the drawable that will be presented at the end of the frame
   // id<MTLDrawable> drawable = view.currentDrawable;

    // Request that the drawable texture be presented by the windowing system once drawing is done
    
    if (commandBuffer != nil && drawable != nil) {

        [commandBuffer presentDrawable:drawable];

        
        [commandBuffer commit];
    }
}

drawKSCNNode方法 #

- (void)drawKSCNNode:(id<MTLRenderCommandEncoder>)commandEncoder
                node:(KSCNNode *)node
            uniforms:(float *)uniforms
    startTranslation:(SCNVector3)startTranslation
       startRotation:(SCNVector4)startRotation
          startScale:(SCNVector3)startScale {
    
    if (node.geometry != nil) {
        if (node.geometry.vertices != nil && node.geometry.colors != nil) {
            //NSLog(@"DRAWING %@, %ld vertices", node.name, node.geometry.vertexCount);
            
            uniformSet[8] = node.position.x;
            uniformSet[9] = node.position.y;
            uniformSet[10] = node.position.z;
            
            // Update uniforms for this node
            memcpy(_uniforms.contents, uniformSet, sizeof(float) * 45);
            
            [commandEncoder setVertexBuffer:[node.geometry.vertices getBuffer] offset:0 atIndex:0];
            [commandEncoder setVertexBuffer:[node.geometry.colors getBuffer] offset:0 atIndex:1];
            [commandEncoder setVertexBuffer:_uniforms offset:0 atIndex:2];
            
            [commandEncoder drawPrimitives:MTLPrimitiveTypeTriangle vertexStart:0 vertexCount:node.geometry.vertexCount instanceCount:node.geometry.vertexCount / 3];
        }

    }
    
    for (int j = 0; j < [node.children count]; j++) {
        KSCNNode * child = [node.children objectAtIndex:j];
        
        [self drawKSCNNode:commandEncoder node:child uniforms:uniforms
          startTranslation:startTranslation startRotation:startRotation startScale:startScale];
    }
}

渲染循环设置 #

- (void)step:(CADisplayLink *)sender {
    [self drawInMTKView:self];
}

- (void)createDisplayLink {
    CADisplayLink *displayLink = [self displayLinkWithTarget:self selector:@selector(step:)];
    [displayLink addToRunLoop:[NSRunLoop currentRunLoop] forMode:NSRunLoopCommonModes];
    [displayLink setPreferredFrameRateRange:CAFrameRateRangeMake(30, 60, 60)];
}

- (void)drawInMTKView:(nonnull MTKView *)view {
    // Wah wah ....

    @autoreleasepool {
        [self setupEncoder:view];
    }
}

核心优化方案 #

1. 预创建并复用渲染管线状态 #

当前每帧调用newRenderPipelineStateWithDescriptor创建管线状态，这是CPU开销极大的操作。管线状态只需初始化一次（如在视图初始化或setup方法中），之后每帧直接复用：

// 初始化阶段执行
- (void)setupRenderPipeline {
    NSError *error = nil;
    _pipelineState = [self.device newRenderPipelineStateWithDescriptor:_renderPipelineDescriptor error:&error];
    if (!_pipelineState) {
        NSLog(@"管线状态创建失败: %@", error);
    }
}

// setupEncoder中直接复用
// 移除原有的pipeline创建代码，替换为：
if (!_pipelineState) return;
[commandEncoder setRenderPipelineState:_pipelineState];

2. 修复Render Pass Descriptor使用逻辑 #

代码中获取了view.currentRenderPassDescriptor却未使用，反而手动创建新的Descriptor，这会导致配置不匹配及额外开销。直接复用MTKView提供的Descriptor即可：

// 移除手动创建renderPass的代码，替换为：
renderPass = view.currentRenderPassDescriptor;
if (renderPass == nil) return;
// 仅修改需要自定义的属性（如clearColor）
renderPass.colorAttachments[0].clearColor = self.clearColor;

3. 优化Drawable获取方式 #

使用view.currentDrawable替代[_metalLayer nextDrawable]，MTKView已内置Drawable生命周期管理，可避免手动调用导致的CPU-GPU同步阻塞：

drawable = view.currentDrawable;

4. 修正实例化渲染参数并减少Draw Call #

当前drawPrimitives的instanceCount设置错误（设为三角形数量而非实例数量），会导致几何重复渲染，GPU负载暴增。同时合并静态几何、使用实例化渲染减少Draw Call：

// 修正instanceCount为1（单个实例渲染）
[commandEncoder drawPrimitives:MTLPrimitiveTypeTriangle vertexStart:0 vertexCount:node.geometry.vertexCount instanceCount:1];

// 进阶优化：将多个静态节点的几何合并为单个Buffer，减少setVertexBuffer和drawPrimitives调用次数

5. 改用MTKView原生渲染循环 #

移除自定义的CADisplayLink，直接实现MTKView的代理方法，Metal框架会自动优化渲染同步逻辑：

// 遵循MTKViewDelegate协议
- (void)mtkView:(MTKView *)view drawableSizeWillChange:(CGSize)size {
    // 处理尺寸变更逻辑
}

- (void)drawInMTKView:(MTKView *)view {
    @autoreleasepool {
        [self setupEncoder:view];
    }
}

内容的提问来源于stack exchange，提问作者Jonathan Claassens