Ios 由于向数组值添加数字时执行过程中出错，命令缓冲区的Swift Metal Shader执行被中止

我正在尝试一些非常简单的算法，使用金属GPU加速来计算数组中的一些值。着色器在某些情况下会抛出错误，我将对此进行解释。 错误：由于执行期间出错，命令缓冲区的执行被中止。因导致先前/过多GPU错误IOAF代码4而忽略

着色器仅在向数组索引处的现有值添加值时才会引发此错误。例如：

这不会导致错误：

    kernel void shader (device int *wPointsIntensity [[buffer(0)]],
                const device uint *wPointsXCoord [[buffer(1)]],
                const device uint *wPointsYCoord [[buffer(2)]],
                device float *pixelSignalIntensity [[buffer(3)]],
                device float *pixelDistance [[buffer(4)]],
                const device uint& noOfPoints [[ buffer(5) ]],
                const device uint& width [[ buffer(6) ]],
                const device uint& height [[ buffer(7) ]],
                uint id [[ thread_position_in_grid ]]) {

//this does not throw error
for (uint wpIndex = 0; wpIndex < noOfPoints; wpIndex++) {
    for (uint heightIndex = 0; heightIndex < height; heightIndex++) {
        for (uint widthIndex = 0; widthIndex < width; widthIndex++) {

            uint pixelIndex = heightIndex * width + widthIndex;
            pixelDistance[pixelIndex] = float(pixelIndex);
            pixelSignalIntensity[pixelIndex] = float(pixelIndex);
}}}}

而如果你改变

pixelDistance[pixelIndex]=浮动pixelIndex； 与

pixelDistance[pixelIndex]+=floatpixelIndex

它将抛出一个错误

以下是swift代码：

var wPointsValues = [Int32](repeating:0, count: wPoints.count)
var wPointsXLocations = [Int32](repeating:0, count: wPoints.count)
var wPointsYLocations = [Int32](repeating:0, count: wPoints.count)
        for i in 0..<wPoints.count {
            wPointsValues[i] = Int32(wPoints[i].signalIntensity)
            wPointsXLocations[i] = Int32(wPoints[i].location.x)
            wPointsYLocations[i] = Int32(wPoints[i].location.y)
        }

        var numberOfWPoints:Int32 = Int32(wPoints.count)
        var int32Width = Int32(width)
        var int32Height = Int32(height)

        //output arrays

        let numberOfResults = wPoints.count * Int(width) * Int(height)
        var wPointsSignalIntensity = [Float32](repeating:0.0, count: numberOfResults)
        var wPointsDistance = [Float32](repeating:0.0, count: numberOfResults)


        //local variables
        var signalDensity:[Float32] = [Float32](repeating:0.0, count: numberOfResults)
        var signalDistance:[Float32] = [Float32](repeating:0.0, count: numberOfResults)

            //create input buffers
            let inWPointSignalValues = device.makeBuffer(bytes: wPointsValues, length: (MemoryLayout<Int32>.stride * wPoints.count), options: [])
            let inWPointXCoordBuffer = device.makeBuffer(bytes: wPointsXLocations, length: (MemoryLayout<Int32>.stride * wPoints.count), options: [])
            let inWPointYCoordBuffer = device.makeBuffer(bytes: wPointsYLocations, length: (MemoryLayout<Int32>.stride * wPoints.count), options: [])

            //create putput buffers
            let outPixelSignalIntensityBuffer = device.makeBuffer(bytes: wPointsSignalIntensity, length: (MemoryLayout<Float32>.stride * numberOfResults), options: [])
            let outPixelDistanceBuffer = device.makeBuffer(bytes: wPointsDistance, length: (MemoryLayout<Float32>.stride * numberOfResults), options: [])

            let commandBuffer = (mtlCommmandQueue?.makeCommandBuffer())!
            let computeCommandEncoder = (commandBuffer.makeComputeCommandEncoder())!
            computeCommandEncoder.setComputePipelineState(mtlComputePipelineFilter!)
            //set input buffers
            computeCommandEncoder.setBuffer(inWPointSignalValues, offset: 0, index: 0)
            computeCommandEncoder.setBuffer(inWPointXCoordBuffer, offset: 0, index: 1)
            computeCommandEncoder.setBuffer(inWPointYCoordBuffer, offset: 0, index: 2)

            //set output buffers
            computeCommandEncoder.setBuffer(outPixelSignalIntensityBuffer, offset: 0, index: 3)
            computeCommandEncoder.setBuffer(outPixelDistanceBuffer, offset: 0, index: 4)

            //set constants
            computeCommandEncoder.setBytes(&numberOfWPoints, length: MemoryLayout<Int32>.stride, index: 5)
            computeCommandEncoder.setBytes(&int32Width, length: MemoryLayout<Int32>.stride, index: 6)
            computeCommandEncoder.setBytes(&int32Height, length: MemoryLayout<Int32>.stride, index: 7)


            let threadsPerGroup = MTLSize(width:2,height:2,depth:2)
            let numThreadgroups = MTLSize(width:2, height:2, depth:2)
            computeCommandEncoder.dispatchThreadgroups(numThreadgroups, threadsPerThreadgroup: threadsPerGroup)

            let endBufferAllocation = mach_absolute_time()

            print("time for creating and setting buffert: time: \(Double(endBufferAllocation - start) / Double(NSEC_PER_SEC))")
            computeCommandEncoder.endEncoding()

            commandBuffer.commit()
            commandBuffer.waitUntilCompleted()
            let allComplete = mach_absolute_time()
        self.signalDistance = (outPixelDistanceBuffer?.contents())!
        self.signalDensity = (outPixelSignalIntensityBuffer?.contents())!

---

我有这个问题很久了，程序断断续续地崩溃了。事实证明，我正在访问内核中没有被缓冲区分配的内存。在内核中，我做了一个for循环0..Hmm。您使用的缓冲区超出了需要。您可以将输入和制服合并到结构中。此外，在这种情况下，const device one可以是常量。在任何情况下，都会导致所有计算线程执行相同的工作。他们很可能是步调一致。因此，所有线程都同时执行pixelDistance[pixelIndex]的增量，这至少会产生未定义的结果。GPU加速的主要好处是并行性，使线程独立处理问题。如果您根据线程ID分割工作，您将获得更好的性能，还可能避免错误