在 Swift 中从 MTLBuffer 读取数据

Question

在GPU上处理后，我需要从MTLBuffer读取数据。到目前为止，我已经尝试使用以下代码，但它总是崩溃并显示EXC_BAD_ACCESS错误代码。

struct gauss_model {
  var mean : [Float32] = [0.0, 0.0, 0.0];
  var covMat : [[Float32]] = [[0.0, 0.0, 0.0],[0.0, 0.0, 0.0],[0.0, 0.0, 0.0]];
  var invCovMat : [[Float32]] = [[0.0, 0.0, 0.0],[0.0, 0.0, 0.0],[0.0, 0.0, 0.0]];
  var samples : Int32 = 0;
}
self.gaussModels = [gauss_model](count: Int(10), repeatedValue: gauss_model())
self.modelsBuffer = self.device.newBufferWithBytes(self.gaussModels, length:      self.gaussModels.count * sizeof(gauss_model), options: MTLResourceOptions.OptionCPUCacheModeDefault)
commandEncoder.setBuffer(self.modelsBuffer, offset: 0, atIndex: 0)
    // execute GPU code
var model = unsafeBitCast(self.modelsBuffer.contents(), UnsafeMutablePointer<gauss_model>.self)
NSLog("%@", model.memory.mean) // crashes on this statement

我还尝试了不同的方法来获取值，例如

var model = UnsafeMutablePointer<gauss_model>(self.modelsBuffer.contents())
// iterate over models with model.memory and model.successor()

或

var model = UnsafeMutablePointer<[gauss_model]>(self.modelBuffer.contents())
let models : [gauss_model] = model.memory

但他们都没有奏效。有没有办法做到这一点？

Answer 1

我已经设法解决了这个问题。问题在于我对 Swift 内存管理的错误假设，以及函数newBufferWithBytes只做一个浅拷贝的事实。调用该函数仅将指针复制到mean、covMat和invCovMat数组，self.modelBuffer.contents()包含指向未初始化内存的指针。访问内存导致崩溃。

Answer 2

如果你来自谷歌试图阅读一个简单的整数MTLBuffer，这是我是如何做到的。

我对我的代码不是很有信心，但它似乎有效。（我还向Apple发送了一份错误报告，因为快速文档是用Objective-C编写的）

let bufferSize = 10
let rawbufferSize = bufferSize * MemoryLayout<Int>.stride
let oneArray: [Int] = Array(repeating: 1, count: bufferSize)
let BufferA = device.makeBuffer(bytes: oneArray, length: rawbufferSize, options: MTLResourceOptions.storageModeShared)
let BufferB = device.makeBuffer(bytes: oneArray, length: rawbufferSize, options: MTLResourceOptions.storageModeShared)
let BufferResult = device.makeBuffer(bytes: oneArray, length: rawbufferSize, options: MTLResourceOptions.storageModeShared)
let commandBuffer = queue?.makeCommandBuffer()
let computeEncoder = commandBuffer?.makeComputeCommandEncoder()
computeEncoder?.setComputePipelineState(pipelineState)
computeEncoder?.setBuffer(BufferA, offset: 0, index: 0)
computeEncoder?.setBuffer(BufferB, offset: 0, index: 1)
computeEncoder?.setBuffer(BufferResult, offset: 0, index: 2)
let gridSize = MTLSizeMake(bufferSize * MemoryLayout<Int>.stride, 1, 1)
let threadGroupSize = pipelineState.maxTotalThreadsPerThreadgroup > bufferSize ? bufferSize : pipelineState.maxTotalThreadsPerThreadgroup
computeEncoder?.dispatchThreads(
    MTLSizeMake(rawbufferSize, 1, 1),
    threadsPerThreadgroup: MTLSizeMake(threadGroupSize, 1, 1 ))
computeEncoder?.endEncoding()
commandBuffer?.commit()
commandBuffer?.waitUntilCompleted()
// Ewww
var rawPointer = BufferResult?.contents()
var typedPointer = rawPointer?.bindMemory(to: Int.self, capacity: rawbufferSize)
let bufferedPointer = UnsafeBufferPointer(start: typedPointer, count: rawbufferSize)
for i in 0..<bufferSize {
    print(bufferedPointer[i])
}