使用LiDAR相机采集深度的Swift示例代码GPU超时崩溃求助

我在苹果开发者官网下载了《使用LiDAR相机采集深度》的Swift示例项目，在iPhone 14 Pro上运行时，应用会间歇性崩溃，抛出错误：

‘Execution of the command buffer was aborted due to an error during execution. Caused GPU Timeout Error (00000002:kIOGPUCommandBufferCallbackErrorTimeout)’

崩溃时间不定，有时几秒内就崩溃，有时需十几分钟甚至更久。

我尝试了多种方法排查，也通过Feedback Assistant向苹果提交了反馈，但始终无法通过Xcode定位崩溃代码行。经过多小时排查，终于发现错误出在MetalTextureViewColor.swift文件的draw方法中，具体是encoder.endEncoding()这一行，报错信息为：

‘MTLCommandBuffer execution failed. The commands associated with an encoder caused an error’

完整的draw函数代码如下：

override func draw(in view: MTKView) {
        guard parent.capturedData.colorY != nil && parent.capturedData.colorCbCr != nil else {
            print("There's no content to display.")
            return
        }
        guard let commandBuffer = metalCommandQueue.makeCommandBuffer() else { return }
        guard let passDescriptor = view.currentRenderPassDescriptor else { return }
        guard let encoder = commandBuffer.makeRenderCommandEncoder(descriptor: passDescriptor) else { return }
        // Vertex and Texture coordinates data (x,y,u,v) * 4 ordered for triangle strip.
        let vertexData: [Float] = [-1, -1, 1, 1,
                                    1, -1, 1, 0,
                                   -1,  1, 0, 1,
                                    1,  1, 0, 0]
        encoder.setVertexBytes(vertexData, length: vertexData.count * MemoryLayout<Float>.stride, index: 0)
        encoder.setFragmentTexture(parent.capturedData.colorY!, index: 0)
        encoder.setFragmentTexture(parent.capturedData.colorCbCr!, index: 1)
        encoder.setDepthStencilState(depthState)
        encoder.setRenderPipelineState(pipelineState)
        encoder.drawPrimitives(type: .triangleStrip, vertexStart: 0, vertexCount: 4)
        
        // error happening on the line below
        encoder.endEncoding()


        commandBuffer.present(view.currentDrawable!)
        commandBuffer.commit()
        //commandBuffer.waitUntilCompleted()
    }

请问是否有办法捕获该错误并忽略，让应用继续运行？

1. 捕获Command Buffer执行错误 #

Metal的MTLCommandBuffer支持添加完成回调，可在回调中捕获执行阶段的错误，避免应用崩溃。同时对encoder.endEncoding()做容错处理：

override func draw(in view: MTKView) {
        guard parent.capturedData.colorY != nil && parent.capturedData.colorCbCr != nil else {
            print("There's no content to display.")
            return
        }
        guard let commandBuffer = metalCommandQueue.makeCommandBuffer() else { return }
        
        // 添加命令缓冲区完成回调，捕获执行错误
        commandBuffer.addCompletedHandler { buffer in
            if let error = buffer.error {
                print("Command Buffer执行错误: \(error.localizedDescription)")
                // 可在此添加错误统计、重试逻辑，或直接忽略
                return
            }
        }
        
        guard let passDescriptor = view.currentRenderPassDescriptor else { return }
        guard let encoder = commandBuffer.makeRenderCommandEncoder(descriptor: passDescriptor) else { return }
        
        let vertexData: [Float] = [-1, -1, 1, 1,
                                    1, -1, 1, 0,
                                   -1,  1, 0, 1,
                                    1,  1, 0, 0]
        encoder.setVertexBytes(vertexData, length: vertexData.count * MemoryLayout<Float>.stride, index: 0)
        encoder.setFragmentTexture(parent.capturedData.colorY!, index: 0)
        encoder.setFragmentTexture(parent.capturedData.colorCbCr!, index: 1)
        encoder.setDepthStencilState(depthState)
        encoder.setRenderPipelineState(pipelineState)
        encoder.drawPrimitives(type: .triangleStrip, vertexStart: 0, vertexCount: 4)
        
        // 尝试安全结束编码
        do {
            try encoder.endEncoding()
        } catch {
            print("Encoder结束编码错误: \(error.localizedDescription)")
            // 即使出错也尝试结束编码，避免资源泄漏
            encoder.endEncoding()
            return
        }

        commandBuffer.present(view.currentDrawable!)
        commandBuffer.commit()
    }

2. 根源优化：减少GPU超时概率 #

虽然捕获错误能避免崩溃，但从根源优化可彻底解决问题：

• 确保colorY和colorCbCr纹理在渲染周期内未被释放或异步修改，可添加纹理引用计数检查
• 将vertexData改为类级静态常量，避免每次draw都重新创建数组
• 关闭Xcode的Metal Validation等调试选项，这些会增加GPU负载
• 确保pipelineState在初始化时预创建，不要在draw方法内重复初始化

3. 全局容错兜底（不推荐） #

如果想做最极端的容错，可将draw逻辑封装到单独方法并添加全局异常捕获，但可能导致未知渲染状态：

override func draw(in view: MTKView) {
    do {
        try performRender(in: view)
    } catch {
        print("渲染过程出错: \(error)")
    }
}

private func performRender(in view: MTKView) throws {
    // 原draw方法的完整逻辑
    guard parent.capturedData.colorY != nil && parent.capturedData.colorCbCr != nil else {
        print("There's no content to display.")
        return
    }
    guard let commandBuffer = metalCommandQueue.makeCommandBuffer() else { return }
    guard let passDescriptor = view.currentRenderPassDescriptor else { return }
    guard let encoder = commandBuffer.makeRenderCommandEncoder(descriptor: passDescriptor) else { return }
    
    let vertexData: [Float] = [-1, -1, 1, 1,
                                1, -1, 1, 0,
                               -1,  1, 0, 1,
                                1,  1, 0, 0]
    encoder.setVertexBytes(vertexData, length: vertexData.count * MemoryLayout<Float>.stride, index: 0)
    encoder.setFragmentTexture(parent.capturedData.colorY!, index: 0)
    encoder.setFragmentTexture(parent.capturedData.colorCbCr!, index: 1)
    encoder.setDepthStencilState(depthState)
    encoder.setRenderPipelineState(pipelineState)
    encoder.drawPrimitives(type: .triangleStrip, vertexStart: 0, vertexCount: 4)
    
    encoder.endEncoding()

    commandBuffer.present(view.currentDrawable!)
    commandBuffer.commit()
}

内容的提问来源于stack exchange，提问作者N D