/// <summary>
/// Encode a MPSCnnKernel into a command Buffer. The operation shall proceed out-of-place.
/// We calculate the appropriate offset as per how TensorFlow calculates its padding using input image size and stride here.
/// This [Link](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/nn.py) has an explanation in header comments how tensorFlow pads its convolution input images.
/// </summary>
/// <param name="commandBuffer">A valid MTLCommandBuffer to receive the encoded filter</param>
/// <param name="sourceImage">A valid MPSImage object containing the source image.</param>
/// <param name="destinationImage">A valid MPSImage to be overwritten by result image. destinationImage may not alias sourceImage</param>
public override void EncodeToCommandBuffer(IMTLCommandBuffer commandBuffer, MPSImage sourceImage, MPSImage destinationImage)
{
// select offset according to padding being used or not
if (padding)
{
var pad_along_height = ((destinationImage.Height - 1) * StrideInPixelsY + KernelHeight - sourceImage.Height);
var pad_along_width = ((destinationImage.Width - 1) * StrideInPixelsX + KernelWidth - sourceImage.Width);
var pad_top = pad_along_height / 2;
var pad_left = pad_along_width / 2;
Offset = new MPSOffset {
X = (nint)(KernelWidth / 2 - pad_left),
Y = (nint)(KernelHeight / 2 - pad_top),
Z = 0
};
}
else
{
Offset = new MPSOffset {
X = (nint)(KernelWidth / 2),
Y = (nint)(KernelHeight / 2),
Z = 0
};
}
base.EncodeToCommandBuffer(commandBuffer, sourceImage, destinationImage);
}
/// <summary>
/// This function encodes all the layers of the network into given commandBuffer, it calls subroutines for each piece of the network
/// Returns: Guess of the network as to what the digit is as UInt
/// </summary>
/// <param name="inputImage">Image coming in on which the network will run</param>
/// <param name="imageNum">If the test set is being used we will get a value between 0 and 9999 for which of the 10,000 images is being evaluated</param>
/// <param name="correctLabel">The correct label for the inputImage while testing</param>
public virtual uint Forward (MPSImage inputImage = null, int imageNum = 9999, int correctLabel = 10)
{
uint label = 99;
// Get command buffer to use in MetalPerformanceShaders.
using (var commandBuffer = commandQueue.CommandBuffer ()) {
// output will be stored in this image
var finalLayer = new MPSImage (commandBuffer.Device, DID);
// encode layers to metal commandBuffer
if (inputImage == null)
layer.EncodeToCommandBuffer (commandBuffer, SrcImage, dstImage);
else
layer.EncodeToCommandBuffer (commandBuffer, inputImage, dstImage);
softmax.EncodeToCommandBuffer (commandBuffer, dstImage, finalLayer);
// add a completion handler to get the correct label the moment GPU is done and compare it to the correct output or return it
commandBuffer.AddCompletedHandler (buffer => {
label = GetLabel (finalLayer);
if (correctLabel == label)
Atomics.Increment ();
});
// commit commandbuffer to run on GPU and wait for completion
commandBuffer.Commit ();
if (imageNum == 9999 || inputImage == null)
commandBuffer.WaitUntilCompleted ();
}
return label;
}
/// <summary>
/// This function reads the output probabilities from finalLayer to CPU, sorts them and gets the label with heighest probability
/// </summary>
/// <param name="finalLayer">output image of the network this has probabilities of each digit</param>
/// <returns>Guess of the network as to what the digit is as uint</returns>
public uint GetLabel(MPSImage finalLayer)
{
// even though we have 10 labels outputed the MTLTexture format used is RGBAFloat16 thus 3 slices will have 3*4 = 12 outputs
var resultHalfArray = Enumerable.Repeat((ushort)6, 12).ToArray();
var resultHalfArrayHandle = GCHandle.Alloc(resultHalfArray, GCHandleType.Pinned);
var resultHalfArrayPtr = resultHalfArrayHandle.AddrOfPinnedObject();
var resultFloatArray = Enumerable.Repeat(0.3f, 10).ToArray();
var resultFloatArrayHandle = GCHandle.Alloc(resultFloatArray, GCHandleType.Pinned);
var resultFloatArrayPtr = resultFloatArrayHandle.AddrOfPinnedObject();
for (uint i = 0; i <= 2; i++)
{
finalLayer.Texture.GetBytes(resultHalfArrayPtr + 4 * (int)i * sizeof(ushort),
sizeof(ushort) * 1 * 4, sizeof(ushort) * 1 * 1 * 4,
new MTLRegion(new MTLOrigin(0, 0, 0), new MTLSize(1, 1, 1)),
0, i);
}
// we use vImage to convert our data to float16, Metal GPUs use float16 and swift float is 32-bit
var fullResultVImagebuf = new vImageBuffer {
Data = resultFloatArrayPtr,
Height = 1,
Width = 10,
BytesPerRow = 10 * 4
};
var halfResultVImagebuf = new vImageBuffer {
Data = resultHalfArrayPtr,
Height = 1,
Width = 10,
BytesPerRow = 10 * 2
};
if (Planar16FtoPlanarF(ref halfResultVImagebuf, ref fullResultVImagebuf, 0) != vImageError.NoError)
{
Console.WriteLine("Error in vImage");
}
// poll all labels for probability and choose the one with max probability to return
float max = 0f;
uint mostProbableDigit = 10;
for (uint i = 0; i <= 9; i++)
{
if (max < resultFloatArray [i])
{
max = resultFloatArray [i];
mostProbableDigit = i;
}
}
resultHalfArrayHandle.Free();
resultFloatArrayHandle.Free();
return(mostProbableDigit);
}
public (NSArray <MPSImage> Inputs, NSArray <MPSState> Losses) GetRandomBatch(IMTLDevice device, int batchSize)
{
var trainImageDesc = MPSImageDescriptor.GetImageDescriptor(
MPSImageFeatureChannelFormat.Unorm8,
ImageSize, ImageSize, 1,
1,
MTLTextureUsage.ShaderWrite | MTLTextureUsage.ShaderRead);
var trainBatch = new List <MPSImage> ();
var lossStateBatch = new List <MPSState> ();
unsafe
{
fixed(byte *imagesPointer = imagesData)
fixed(byte *labelsPointer = labelsData)
{
for (var i = 0; i < batchSize; i++)
{
var randomIndex = random.Next(numImages);
var trainImage = new MPSImage(device, trainImageDesc)
{
Label = "TrainImage" + i
};
trainBatch.Add(trainImage);
var trainImagePointer = imagesPointer + ImagesPrefixSize + randomIndex * ImageSize * ImageSize;
trainImage.WriteBytes((IntPtr)trainImagePointer, MPSDataLayout.HeightPerWidthPerFeatureChannels, 0);
var labelPointer = labelsPointer + LabelsPrefixSize + randomIndex;
var labelsValues = new float[12];
labelsValues[*labelPointer] = 1;
fixed(void *p = labelsValues)
{
using var data = NSData.FromBytes((IntPtr)p, 12 * sizeof(float));
var desc = MPSCnnLossDataDescriptor.Create(
data, MPSDataLayout.HeightPerWidthPerFeatureChannels, new MTLSize(1, 1, 12));
var lossState = new MPSCnnLossLabels(device, desc);
lossStateBatch.Add(lossState);
}
}
}
}
return(NSArray <MPSImage> .FromNSObjects(trainBatch.ToArray()),
NSArray <MPSState> .FromNSObjects(lossStateBatch.ToArray()));
}
/// <summary>
/// This function encodes all the layers of the network into given commandBuffer, it calls subroutines for each piece of the network
/// Returns: Guess of the network as to what the digit is as uint
/// </summary>
/// <param name="inputImage">Image coming in on which the network will run</param>
/// <param name="imageNum">If the test set is being used we will get a value between 0 and 9999 for which of the 10,000 images is being evaluated</param>
/// <param name="correctLabel">The correct label for the inputImage while testing</param>
public override uint Forward(MPSImage inputImage = null, int imageNum = 9999, int correctLabel = 10)
{
uint label = 99;
// Get command buffer to use in MetalPerformanceShaders.
using (var commandBuffer = commandQueue.CommandBuffer()) {
// output will be stored in this image
var finalLayer = new MPSImage(commandBuffer.Device, DID);
// encode layers to metal commandBuffer
if (inputImage == null)
{
conv1.EncodeToCommandBuffer(commandBuffer, SrcImage, c1Image);
}
else
{
conv1.EncodeToCommandBuffer(commandBuffer, inputImage, c1Image);
}
pool.EncodeToCommandBuffer(commandBuffer, c1Image, p1Image);
conv2.EncodeToCommandBuffer(commandBuffer, p1Image, c2Image);
pool.EncodeToCommandBuffer(commandBuffer, c2Image, p2Image);
fc1.EncodeToCommandBuffer(commandBuffer, p2Image, fc1Image);
fc2.EncodeToCommandBuffer(commandBuffer, fc1Image, dstImage);
softmax.EncodeToCommandBuffer(commandBuffer, dstImage, finalLayer);
// add a completion handler to get the correct label the moment GPU is done and compare it to the correct output or return it
commandBuffer.AddCompletedHandler(buffer => {
label = GetLabel(finalLayer);
if (correctLabel == label)
{
Atomics.Increment();
}
});
// commit commandbuffer to run on GPU and wait for completion
commandBuffer.Commit();
if (imageNum == 9999 || inputImage == null)
{
commandBuffer.WaitUntilCompleted();
}
}
return(label);
}
/// <summary>
/// This function runs the inference network on the test set
/// </summary>
/// <param name="imageNum">If the test set is being used we will get a value between 0 and 9999 for which of the 10,000 images is being evaluated</param>
/// <param name="correctLabel">The correct label for the inputImage while testing</param>
void Inference(int imageNum, int correctLabel)
{
// get the correct image pixels from the test set
int startIndex = imageNum * mnistInputNumPixels;
// create a source image for the network to forward
var inputImage = new MPSImage(device, runningNet.SID);
// put image in source texture (input layer)
inputImage.Texture.ReplaceRegion(region: new MTLRegion(new MTLOrigin(0, 0, 0), new MTLSize((nint)mnistInputWidth, mnistInputHeight, 1)),
level: 0,
slice: 0,
pixelBytes: Mnistdata.Images + startIndex,
bytesPerRow: mnistInputWidth,
bytesPerImage: 0);
// run the network forward pass
runningNet.Forward(inputImage, imageNum, correctLabel);
}
public MnistFullLayerNeuralNetwork(IMTLCommandQueue commandQueueIn)
{
// CommandQueue to be kept around
commandQueue = commandQueueIn;
device = commandQueueIn.Device;
// Initialize MPSImage from descriptors
SrcImage = new MPSImage(device, SID);
dstImage = new MPSImage(device, DID);
// setup convolution layer (which is a fully-connected layer)
// cliprect, offset is automatically set
layer = SlimMPSCnnFullyConnected.Create(kernelWidth: 28, kernelHeight: 28,
inputFeatureChannels: 1, outputFeatureChannels: 10,
neuronFilter: null, device: device,
kernelParamsBinaryName: "NN");
// prepare softmax layer to be applied at the end to get a clear label
softmax = new MPSCnnSoftMax(device);
}
public MnistFullLayerNeuralNetwork (IMTLCommandQueue commandQueueIn)
{
// CommandQueue to be kept around
commandQueue = commandQueueIn;
device = commandQueueIn.Device;
// Initialize MPSImage from descriptors
SrcImage = new MPSImage (device, SID);
dstImage = new MPSImage (device, DID);
// setup convolution layer (which is a fully-connected layer)
// cliprect, offset is automatically set
layer = SlimMPSCnnFullyConnected.Create (kernelWidth: 28, kernelHeight: 28,
inputFeatureChannels: 1, outputFeatureChannels: 10,
neuronFilter: null, device: device,
kernelParamsBinaryName: "NN");
// prepare softmax layer to be applied at the end to get a clear label
softmax = new MPSCnnSoftMax (device);
}
public MnistDeepConvNeuralNetwork(IMTLCommandQueue commandQueueIn)
: base(commandQueueIn)
{
// use device for a little while to initialize
var device = commandQueueIn.Device;
pool = new MPSCnnPoolingMax(device, 2, 2, 2, 2)
{
Offset = new MPSOffset {
X = 1, Y = 1, Z = 0
},
EdgeMode = MPSImageEdgeMode.Clamp
};
relu = new MPSCnnNeuronReLU(device, 0);
// Initialize MPSImage from descriptors
c1Image = new MPSImage(device, c1id);
p1Image = new MPSImage(device, p1id);
c2Image = new MPSImage(device, c2id);
p2Image = new MPSImage(device, p2id);
fc1Image = new MPSImage(device, fc1id);
// setup convolution layers
conv1 = SlimMPSCnnConvolution.Create(kernelWidth: 5,
kernelHeight: 5,
inputFeatureChannels: 1,
outputFeatureChannels: 32,
neuronFilter: relu,
device: device,
kernelParamsBinaryName: "conv1",
padding: true,
strideX: 1,
strideY: 1,
destinationFeatureChannelOffset: 0,
groupNum: 1);
conv2 = SlimMPSCnnConvolution.Create(kernelWidth: 5,
kernelHeight: 5,
inputFeatureChannels: 32,
outputFeatureChannels: 64,
neuronFilter: relu,
device: device,
kernelParamsBinaryName: "conv2",
padding: true,
strideX: 1,
strideY: 1,
destinationFeatureChannelOffset: 0,
groupNum: 1);
// same as a 1x1 convolution filter to produce 1x1x10 from 1x1x1024
fc1 = SlimMPSCnnFullyConnected.Create(kernelWidth: 7,
kernelHeight: 7,
inputFeatureChannels: 64,
outputFeatureChannels: 1024,
neuronFilter: null,
device: device,
kernelParamsBinaryName: "fc1",
destinationFeatureChannelOffset: 0);
fc2 = SlimMPSCnnFullyConnected.Create(kernelWidth: 1,
kernelHeight: 1,
inputFeatureChannels: 1024,
outputFeatureChannels: 10,
neuronFilter: null,
device: device,
kernelParamsBinaryName: "fc2");
}
public static unsafe UIKit.UIImage GetUIImage(MPSImage mpsImage)
{
var width = (int)mpsImage.Width;
var height = (int)mpsImage.Height;
var nfc = (int)mpsImage.FeatureChannels;
var obytesPerRow = 4 * width;
var cellSize = 44;
using var cs = CoreGraphics.CGColorSpace.CreateDeviceRGB();
//Console.WriteLine ((width, height, mpsImage.Precision, mpsImage.PixelSize, mpsImage.FeatureChannels, mpsImage.PixelFormat, mpsImage.FeatureChannelFormat));
if (mpsImage.FeatureChannelFormat == MPSImageFeatureChannelFormat.Float32 && nfc == 3)
{
var data = new float[width * height * nfc];
fixed(float *dataPointer = data)
{
mpsImage.ReadBytes((IntPtr)dataPointer, MPSDataLayout.HeightPerWidthPerFeatureChannels, 0);
}
using var bc = new CoreGraphics.CGBitmapContext(null, width, height, 8, obytesPerRow, cs, CoreGraphics.CGImageAlphaInfo.NoneSkipFirst);
var pixels = (byte *)bc.Data;
var p = pixels;
for (var y = 0; y < height; y++)
{
for (var x = 0; x < width; x++)
{
*p++ = 255;
*p++ = ClampRGBA32Float(data[y * (width * 3) + x * 3 + 2]);
*p++ = ClampRGBA32Float(data[y * (width * 3) + x * 3 + 1]);
*p++ = ClampRGBA32Float(data[y * (width * 3) + x * 3 + 0]);
}
}
var cgimage = bc.ToImage();
//Console.WriteLine ($"pixels f32 = " + string.Join (", ", data.Skip (data.Length / 2).Take (12)));
return(UIImage.FromImage(cgimage));
}
else if (mpsImage.FeatureChannelFormat == MPSImageFeatureChannelFormat.Float32 && nfc == 1)
{
var data = new float[width * height * nfc];
fixed(float *dataPointer = data)
{
mpsImage.ReadBytes((IntPtr)dataPointer, MPSDataLayout.HeightPerWidthPerFeatureChannels, 0);
}
using var bc = new CoreGraphics.CGBitmapContext(null, width, height, 8, obytesPerRow, cs, CoreGraphics.CGImageAlphaInfo.NoneSkipFirst);
var pixels = (byte *)bc.Data;
var p = pixels;
for (var y = 0; y < height; y++)
{
for (var x = 0; x < width; x++)
{
var g = ClampRGBA32Float(data[y * width + x]);
* p++ = 255;
* p++ = g;
* p++ = g;
* p++ = g;
}
}
var cgimage = bc.ToImage();
//Console.WriteLine ($"pixels f32 = " + string.Join (", ", data.Skip (data.Length / 2).Take (12)));
return(UIImage.FromImage(cgimage));
}
else if (mpsImage.FeatureChannelFormat == MPSImageFeatureChannelFormat.Unorm8 && nfc == 3)
{
var data = new byte[width * height * (int)mpsImage.FeatureChannels];
fixed(byte *dataPointer = data)
{
mpsImage.ReadBytes((IntPtr)dataPointer, MPSDataLayout.HeightPerWidthPerFeatureChannels, 0);
//mpsImage.Texture.GetBytes ((IntPtr)dataPointer, (nuint)(4 * width), MTLRegion.Create3D (0, 0, 0, width, height, 1), 0);
}
using var bc = new CoreGraphics.CGBitmapContext(null, width, height, 8, obytesPerRow, cs, CoreGraphics.CGImageAlphaInfo.NoneSkipFirst);
var pixels = (byte *)bc.Data;
var p = pixels;
for (var y = 0; y < height; y++)
{
for (var x = 0; x < width; x++)
{
*p++ = 255;
*p++ = data[y * (width * 3) + x * 3 + 2]; // Red
*p++ = data[y * (width * 3) + x * 3 + 1]; // Green
*p++ = data[y * (width * 3) + x * 3 + 0]; // Blue
}
}
var cgimage = bc.ToImage();
//Console.WriteLine ($"pixels 3 unorm8 = " + string.Join (", ", data.Skip (data.Length / 2).Take (12)));
return(UIImage.FromImage(cgimage));
}
else if (mpsImage.FeatureChannelFormat == MPSImageFeatureChannelFormat.Unorm8 && nfc == 1)
{
var data = new byte[width * height * (int)mpsImage.FeatureChannels];
fixed(byte *dataPointer = data)
{
mpsImage.ReadBytes((IntPtr)dataPointer, MPSDataLayout.HeightPerWidthPerFeatureChannels, 0);
//mpsImage.Texture.GetBytes ((IntPtr)dataPointer, (nuint)(4 * width), MTLRegion.Create3D (0, 0, 0, width, height, 1), 0);
}
using var bc = new CoreGraphics.CGBitmapContext(null, width, height, 8, obytesPerRow, cs, CoreGraphics.CGImageAlphaInfo.NoneSkipFirst);
var pixels = (byte *)bc.Data;
var p = pixels;
for (var y = 0; y < height; y++)
{
for (var x = 0; x < width; x++)
{
var g = data[y * width + x]; // Red
* p++ = 255;
* p++ = g;
* p++ = g;
* p++ = g;
}
}
var cgimage = bc.ToImage();
//Console.WriteLine ($"pixels 1 unorm8 = " + string.Join (", ", data.Skip (data.Length / 2).Take (12)));
return(UIImage.FromImage(cgimage));
}
else if (mpsImage.FeatureChannelFormat == MPSImageFeatureChannelFormat.Float32 && width == 1 && height == 1)
{
var data = new float[width * height * nfc];
fixed(void *dataPointer = data)
{
mpsImage.ReadBytes((IntPtr)dataPointer, MPSDataLayout.HeightPerWidthPerFeatureChannels, 0);
}
return(DrawCells(nfc, cellSize, data));
}
else if (mpsImage.FeatureChannelFormat == MPSImageFeatureChannelFormat.Unorm8 && width == 1 && height == 1)
{
var data = new byte[width * height * nfc];
fixed(void *dataPointer = data)
{
mpsImage.ReadBytes((IntPtr)dataPointer, MPSDataLayout.HeightPerWidthPerFeatureChannels, 0);
}
return(DrawCells(nfc, cellSize, data.Select(x => x / 255.0f).ToArray()));
}
else
{
if (width == 1 && height == 1)
{
width = cellSize;
height = cellSize;
}
UIGraphics.BeginImageContext(new CoreGraphics.CGSize(width, height));
UIColor.Red.SetColor();
var m = $"{mpsImage.FeatureChannels}{mpsImage.FeatureChannelFormat}?";
m.DrawString(new CoreGraphics.CGPoint(0, 0), UIFont.SystemFontOfSize(8));
var image = UIGraphics.GetImageFromCurrentImageContext();
UIGraphics.EndImageContext();
return(image);
}
}
/// <summary>
/// This function runs the inference network on the test set
/// </summary>
/// <param name="imageNum">If the test set is being used we will get a value between 0 and 9999 for which of the 10,000 images is being evaluated</param>
/// <param name="correctLabel">The correct label for the inputImage while testing</param>
void Inference (int imageNum, int correctLabel)
{
// get the correct image pixels from the test set
int startIndex = imageNum * mnistInputNumPixels;
// create a source image for the network to forward
var inputImage = new MPSImage (device, runningNet.SID);
// put image in source texture (input layer)
inputImage.Texture.ReplaceRegion (region: new MTLRegion (new MTLOrigin (0, 0, 0), new MTLSize ((nint)mnistInputWidth, mnistInputHeight, 1)),
level: 0,
slice: 0,
pixelBytes: Mnistdata.Images + startIndex,
bytesPerRow: mnistInputWidth,
bytesPerImage: 0);
// run the network forward pass
runningNet.Forward (inputImage, imageNum, correctLabel);
}
/// <summary>
/// Encode a MPSCnnKernel into a command Buffer. The operation shall proceed out-of-place.
/// We calculate the appropriate offset as per how TensorFlow calculates its padding using input image size and stride here.
/// This [Link](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/nn.py) has an explanation in header comments how tensorFlow pads its convolution input images.
/// </summary>
/// <param name="commandBuffer">A valid MTLCommandBuffer to receive the encoded filter</param>
/// <param name="sourceImage">A valid MPSImage object containing the source image.</param>
/// <param name="destinationImage">A valid MPSImage to be overwritten by result image. destinationImage may not alias sourceImage</param>
public override void EncodeToCommandBuffer (IMTLCommandBuffer commandBuffer, MPSImage sourceImage, MPSImage destinationImage)
{
// select offset according to padding being used or not
if (padding) {
var pad_along_height = ((destinationImage.Height - 1) * StrideInPixelsY + KernelHeight - sourceImage.Height);
var pad_along_width = ((destinationImage.Width - 1) * StrideInPixelsX + KernelWidth - sourceImage.Width);
var pad_top = pad_along_height / 2;
var pad_left = pad_along_width / 2;
Offset = new MPSOffset {
X = (nint)(KernelWidth / 2 - pad_left),
Y = (nint)(KernelHeight / 2 - pad_top),
Z = 0
};
} else {
Offset = new MPSOffset {
X = (nint)(KernelWidth / 2),
Y = (nint)(KernelHeight / 2),
Z = 0
};
}
base.EncodeToCommandBuffer (commandBuffer, sourceImage, destinationImage);
}
/// <summary>
/// This function reads the output probabilities from finalLayer to CPU, sorts them and gets the label with heighest probability
/// </summary>
/// <param name="finalLayer">output image of the network this has probabilities of each digit</param>
/// <returns>Guess of the network as to what the digit is as uint</returns>
public uint GetLabel (MPSImage finalLayer)
{
// even though we have 10 labels outputed the MTLTexture format used is RGBAFloat16 thus 3 slices will have 3*4 = 12 outputs
var resultHalfArray = Enumerable.Repeat ((ushort)6, 12).ToArray ();
var resultHalfArrayHandle = GCHandle.Alloc (resultHalfArray, GCHandleType.Pinned);
var resultHalfArrayPtr = resultHalfArrayHandle.AddrOfPinnedObject ();
var resultFloatArray = Enumerable.Repeat (0.3f, 10).ToArray ();
var resultFloatArrayHandle = GCHandle.Alloc (resultFloatArray, GCHandleType.Pinned);
var resultFloatArrayPtr = resultFloatArrayHandle.AddrOfPinnedObject ();
for (uint i = 0; i <= 2; i++) {
finalLayer.Texture.GetBytes (resultHalfArrayPtr + 4 * (int)i * sizeof (ushort),
sizeof (ushort) * 1 * 4, sizeof (ushort) * 1 * 1 * 4,
new MTLRegion (new MTLOrigin (0, 0, 0), new MTLSize (1, 1, 1)),
0, i);
}
// we use vImage to convert our data to float16, Metal GPUs use float16 and swift float is 32-bit
var fullResultVImagebuf = new vImageBuffer {
Data = resultFloatArrayPtr,
Height = 1,
Width = 10,
BytesPerRow = 10 * 4
};
var halfResultVImagebuf = new vImageBuffer {
Data = resultHalfArrayPtr,
Height = 1,
Width = 10,
BytesPerRow = 10 * 2
};
if (Planar16FtoPlanarF (ref halfResultVImagebuf, ref fullResultVImagebuf, 0) != vImageError.NoError)
Console.WriteLine ("Error in vImage");
// poll all labels for probability and choose the one with max probability to return
float max = 0f;
uint mostProbableDigit = 10;
for (uint i = 0; i <= 9; i++) {
if (max < resultFloatArray [i]) {
max = resultFloatArray [i];
mostProbableDigit = i;
}
}
resultHalfArrayHandle.Free ();
resultFloatArrayHandle.Free ();
return mostProbableDigit;
}