Question

我无法在 macOS Designed for iPad 目标上运行在 iOS 上运行良好的应用程序，但在 MTKView 中运行 M1 处理器的 Mac 上运行。

尽管出现以下消息，但它似乎仍在工作，但它非常不稳定且缓慢：

<块引用>

[14917:391638] [GPUDebug] 执行内核的设备加载无效函数“计算”编码器：“0”，调度：0，偏移量 896024

缓冲区：“” [14917:391037] [生命周期] 场景状态更改错误：场景不再连接。 (NSMenuBarScene_D600133B-F270-427C-A7E8-1DCF225C886F) [14917:391635] [GPUDebug] 执行内核函数“计算”编码器的设备加载无效：“0”，调度：0，偏移量 896024

然后他们重复：

<块引用>

缓冲区：“” ::: delegate_identifier：GPUToolsDiagnostics [14917:391640] [Metal Diagnostics] 消息：执行内核函数“计算”编码器的设备加载无效：“0”，调度：0，在偏移量 896024

还有许多其他 Metal 函数正在使用（包括内核类型），但它们没有报告任何错误。

内核定义为：

kernel void compute ( device Point *points [[buffer(0)]],
    device Particle *out [[buffer(1)]],
    device atomic_int *counter [[buffer(2)]],
    device Sample *samples [[buffer(3)]],
    device Settings *settings[[buffer(4)]],
    device Record *records[[buffer(5)]],
    constant ComputeParameters *params [[buffer(6)]],
    texture2d<half> colorMap [[ texture(0) ]],
    texture2d<half> colorMap2 [[ texture(1) ]],
    uint id [[thread_position_in_grid]]) {

管道配置为：

do {
    guard let library = device.makeDefaultLibrary(),
    let function = library.makeFunction(name: "compute") else { return }

computeState = try device.makeComputePipelineState(function: function)
       
} catch let error {
    print(error.localizedDescription)
}

它被调用：

guard let computeBuffer = commandQueue.makeCommandBuffer() else { return }
guard let computeEncoder = computeBuffer.makeComputeCommandEncoder() else { return }
    
computeEncoder.setComputePipelineState(computeState)
    
let width = computePipelineState.threadExecutionWidth
let threadsPerGroup = MTLSizeMake(width, 1, 1)
    
func dispatchThreads(particleCount: Int) {
    let threadsPerGrid = MTLSizeMake(particleCount, 1, 1)
    computeArtPointsEncoder.dispatchThreads(threadsPerGrid,
                                       threadsPerThreadgroup: threadsPerGroup)
    }

computeEncoder.setBuffer(pointBuffer, offset: MemoryLayout<Point>.stride, index: 0)     
computeEncoder.setBuffer(outBuffer, offset: 0, index: 1)                                        
computeEncoder.setBuffer(counterBuffer, offset: 0, index: 2)
computeEncoder.setBuffer(sampleBuffer, offset: 0, index: 3)
computeEncoder.setBuffer(settingsBuffer, offset: 0, index: 4)
computeEncoder.setBuffer(recordBuffer, offset: 0, index: 5)
    
computeEncoder.setBytes(&params, length: MemoryLayout<ComputeParameters>.size, index: 6)
    
computeEncoder.setTexture(colorTexture, index: 0)
computeEncoder.setTexture(colorTexture2, index: 1)
            
if device.supportsFeatureSet(.iOS_GPUFamily4_v1) {
    dispatchThreads(particleCount: Int(params.length))
} else {
    let threads = min(computePipelineState.threadExecutionWidth, Int(params.length))
    let threadsPerThreadgroup = MTLSize(width: threads,
                                            height: 1,
                                            depth: 1)
    let groups = Int(params.length) / threads + 1
    let groupsPerGrid = MTLSize(width: groups,
                                    height: 1,
                                    depth: 1)
    computeEncoder.dispatchThreadgroups(groupsPerGrid, threadsPerThreadgroup: threadsPerThreadgroup)
}

computeEncoder.endEncoding()
computeBuffer.commit()
computeBuffer.waitUntilCompleted()

在 M1 Mac 上使用“Designed for iPad”目标执行内核的 Metal 设备无效

0 个答案: