为什么我不能在Metal中遍历完整数组

Question

我有以下代码，我认为分配2个16kb内存的数组，用7填充一个，用4填充另一个。

然后我运行代码，发现只有前511个项目加在一起。我80％确定它与线程组/线程数的大小有关。任何想法？

func test()
{
    var (device, commandQueue, defaultLibrary, commandBuffer, computeCommandEncoder) = initMetal()

    let alignment:Int = 0x4000
    var xpointer: UnsafeMutableRawPointer? = nil
    var ypointer: UnsafeMutableRawPointer? = nil

    let numberOfFloats = 4096 //56
    let numberOfBytes: Int = 16384

    let retx = posix_memalign(&xpointer, alignment, numberOfBytes)
    if retx != noErr {
        let err = String(validatingUTF8: strerror(retx)) ?? "unknown error"
        fatalError("Unable to allocate aligned memory: \(err).")
    }

    let rety = posix_memalign(&ypointer, alignment, numberOfBytes)
    if rety != noErr {
        let err = String(validatingUTF8: strerror(rety)) ?? "unknown error"
        fatalError("Unable to allocate aligned memory: \(err).")
    }


    let datax = xpointer!.bindMemory(to: Float.self, capacity: numberOfFloats)
    for index in 0..<numberOfFloats {
        datax[index] = 7.0
    }

    let datay = ypointer!.bindMemory(to: Float.self, capacity: numberOfFloats)
    for index in 0..<numberOfFloats {
        datay[index] = 4.0
    }

    kernelFunction = defaultLibrary.makeFunction(name: "sigmoid")
    do
    {
        pipelineState = try device.makeComputePipelineState(function: kernelFunction!)
    }
    catch
    {
        fatalError("Unable to create pipeline state")
    }


    let startTime = CFAbsoluteTimeGetCurrent()

    computeCommandEncoder.setComputePipelineState(pipelineState)


    var xvectorBufferNoCopy = device.makeBuffer(bytesNoCopy: xpointer!, length: numberOfBytes, options: [], deallocator: nil)
    computeCommandEncoder.setBuffer(xvectorBufferNoCopy, offset: 0, at: 0)

    var yvectorBufferNoCopy = device.makeBuffer(bytesNoCopy: ypointer!, length: numberOfBytes, options: [], deallocator: nil)
    computeCommandEncoder.setBuffer(yvectorBufferNoCopy, offset: 0, at: 1)


    var threadgroupCounts = MTLSize(width:32,height:1,depth:1)
    var threadgroups = MTLSize(width:1024, height:1, depth:1)

    computeCommandEncoder.dispatchThreadgroups(threadgroups, threadsPerThreadgroup: threadgroupCounts)

    computeCommandEncoder.endEncoding()
    commandBuffer.commit()
    commandBuffer.waitUntilCompleted()

    let timeElapsed = CFAbsoluteTimeGetCurrent() - startTime
    print("Time elapsed for \(title): \(timeElapsed) s")


    let data = ypointer!.bindMemory(to: Float.self, capacity: numberOfFloats)
    for index in 0..<numberOfFloats {
        print("\(index) - \(data[index])")
    }

}

kernel void addTest(const device float *inVector [[ buffer(0) ]],
                device float *outVector [[ buffer(1) ]],
                uint id [[ thread_position_in_grid ]]) {
    outVector[id] = inVector[id] + outVector[id];
}