使用 JCuda 计算 Skintone 未给出正确的百分比

Question

我正在计算 java 中图像的肤色。

转换yCbCR中Image的像素。
检查图像像素是否在特定范围内，然后它是肤色。
通过将百分比除以总像素来计算百分比。

它在 CPU 代码中工作正常，但是当我将它转换为 GPU 代码时，像素百分比不正确。

让我感到困惑的部分是将像素数据发送到 GPU 并在 GPU 中获取其 r、g、b 值。

所以我按照 JCuda Pixel Invert Example 示例发送像素数据。不同之处在于示例在 int[] 数组中发送像素数据，而我在 byte[] 数组中发送它。

这里是代码。

import static jcuda.driver.JCudaDriver.cuCtxCreate;
import static jcuda.driver.JCudaDriver.cuCtxSynchronize;
import static jcuda.driver.JCudaDriver.cuDeviceGet;
import static jcuda.driver.JCudaDriver.cuInit;
import static jcuda.driver.JCudaDriver.cuLaunchKernel;
import static jcuda.driver.JCudaDriver.cuMemAlloc;
import static jcuda.driver.JCudaDriver.cuMemFree;
import static jcuda.driver.JCudaDriver.cuMemcpyDtoH;
import static jcuda.driver.JCudaDriver.cuMemcpyHtoD;

import java.awt.image.BufferedImage;
import java.awt.image.DataBuffer;
import java.awt.image.DataBufferByte;
import java.awt.image.Raster;
import java.io.File;
import java.io.IOException;

import javax.imageio.ImageIO;

import ij.IJ;
import jcuda.Pointer;
import jcuda.Sizeof;
import jcuda.driver.CUcontext;
import jcuda.driver.CUdevice;
import jcuda.driver.CUdeviceptr;
import jcuda.driver.CUfunction;
import jcuda.driver.JCudaDriver;
import jcuda.nvrtc.JNvrtc;

public class SkinTone {

public static void CalculateSKintoneGPU(File file) throws IOException {
    BufferedImage bufferedImage = ImageIO.read(file);
    if (bufferedImage == null || bufferedImage.getData() == null)
        return;
    Raster raster = bufferedImage.getData();

    DataBuffer dataBuffer = raster.getDataBuffer();
    DataBufferByte dataBufferInt = (DataBufferByte)dataBuffer;
    byte[] pixels =  dataBufferInt.getData();

    int totalPixels = raster.getHeight() * raster.getWidth();

    CUfunction kernelFunction = initlize();

    int output[] = execute(kernelFunction, pixels, raster.getWidth(), raster.getHeight());
    // Flushing memory
    raster = null;
    bufferedImage.flush();
    bufferedImage = null;

    long skintoneThreshold = Math.round(output[0] / (double) totalPixels * 100.0);

    System.err.println("Skintone Using GPU: " + output[0]);
    System.err.println("Total Pixel Of GPU: " + totalPixels);
    System.err.println("SKinTone Percentage Using GPU: " + skintoneThreshold + "%");
}

static int[] execute(CUfunction kernelFunction, byte[] pixels, int w, int h) {
    // Allocate memory on the device, and copy the host data to the device
    int size = w * h * Sizeof.BYTE;
    CUdeviceptr pointer = new CUdeviceptr();
    cuMemAlloc(pointer, size);
    cuMemcpyHtoD(pointer, Pointer.to(pixels), size);

    int numElements = 9;
    int s = 0;
    // Allocate device output memory
    CUdeviceptr deviceOutput = new CUdeviceptr();
    cuMemAlloc(deviceOutput, numElements * Sizeof.INT);

    // Set up the kernel parameters: A pointer to an array
    // of pointers which point to the actual values.
    Pointer kernelParameters = Pointer.to(Pointer.to(pointer), Pointer.to(new int[] { w }),
            Pointer.to(new int[] { h }), Pointer.to(deviceOutput));

    // Call the kernel function
    int blockSize = 16;
    int gridSize = (Math.max(w, h) + blockSize - 1) / blockSize;
    cuLaunchKernel(kernelFunction, gridSize, gridSize, 1, // Grid dimension
            blockSize, blockSize, 1, // Block dimension
            0, null, // Shared memory size and stream
            kernelParameters, null // Kernel- and extra parameters
    );
    cuCtxSynchronize();

    // Allocate host output memory and copy the device output
    // to the host.
    int hostOutput[] = new int[numElements];
    cuMemcpyDtoH(Pointer.to(hostOutput), deviceOutput, numElements * Sizeof.INT);

    // Clean up.
    cuMemFree(deviceOutput);
    cuMemFree(pointer);

    return hostOutput;
}

public static CUfunction initlize() {

    // Enable exceptions and omit all subsequent error checks
    JCudaDriver.setExceptionsEnabled(true);
    JNvrtc.setExceptionsEnabled(true);

    // Initialize the driver and create a context for the first device.
    cuInit(0);
    CUdevice device = new CUdevice();
    cuDeviceGet(device, 0);
    CUcontext context = new CUcontext();
    cuCtxCreate(context, 0, device);

    // Obtain the CUDA source code from the CUDA file
    String cuFileName = "Skintone.cu";
    String sourceCode = CudaUtils.readResourceAsString(cuFileName);
    if (sourceCode == null) {
        IJ.showMessage("Error", "Could not read the kernel source code");
    }

    // Create the kernel function
    return CudaUtils.createFunction(sourceCode, "skintone");
}

public static void CalculateSKintoneCPU(File file) throws IOException {
    BufferedImage bufferedImage = ImageIO.read(file);
    if (bufferedImage == null || bufferedImage.getData() == null)
        return;
    Raster raster = bufferedImage.getData();
    float[] rgb = new float[4];
    int totalPixels = raster.getHeight() * raster.getWidth();

    int skinTonePixels = 0;

    for (int x = 0; x < raster.getWidth(); x++) {
        for (int y = 0; y < raster.getHeight(); y++) {
            raster.getPixel(x, y, rgb);
            if (skintone(rgb)) {
                skinTonePixels++;
            }
        }
    }

    // Flushing memory
    raster = null;
    rgb = null;
    bufferedImage.flush();
    bufferedImage = null;

    long skintoneThreshold = Math.round(skinTonePixels / (double) totalPixels * 100.0);

    System.err.println("Skintone Using CPU: " + skinTonePixels);
    System.err.println("Total Pixel Of CPU: " + totalPixels);
    System.err.println("SKinTone Percentage Using CPU: " + skintoneThreshold + "%");
}

private static boolean skintone(float[] rgb) {
    float yCbCr[] = (float[]) convertRGBtoYUV(rgb);
    if ((yCbCr[1] >= 80 && yCbCr[1] <= 120) && (yCbCr[2] >= 133 && yCbCr[2] <= 173)) {
        return true;
    }
    return false;
}

private static float[] convertRGBtoYUV(float[] rgb) {
    final float[] yCbCr = new float[3];
    float r = rgb[0];
    float g = rgb[1];
    float b = rgb[2];

    yCbCr[0] = 16 + (0.299f * r) + (0.587f * g) + (0.144f * b);
    yCbCr[1] = 128 + (-0.169f * r) - (0.331f * g) + (0.5f * b);
    yCbCr[2] = 128 + (0.5f * r) - (0.419f * g) - (0.081f * b);

    return yCbCr;
}

public static void main(String[] args) throws IOException {
    File file = new File("C:\Users\Aqeel\git\jcuda-imagej-example\src\test\resources\lena512color.png");
    CalculateSKintoneCPU(file);
    CalculateSKintoneGPU(file);
}

}

内核文件

    extern "C"
__global__ void skintone(uchar4* data, int w, int h, int* output)
{
    int x = threadIdx.x+blockIdx.x*blockDim.x;
    int y = threadIdx.y+blockIdx.y*blockDim.y;

if (x < w && y < h)
{
    float r, g, b;
    float cb, cr;

    int index = y*w+x;
    uchar4 pixel = data[index];

    r = pixel.x;
    g = pixel.y;
    b = pixel.z;

    cb = 128 + (-0.169f * r) - (0.331f * g) + (0.5f * b);
    cr = 128 + (0.5f * r) - (0.419f * g) - (0.081f * b);


    if((cb >= 80 &&  cb <= 120) && (cr >= 133 &&  cr <= 173)) {
        atomicAdd(&output[0], 1);
    }
}
}

Complete Example src, Machine Need Nvida Card, Cuda Toolkit V9 and Graphics Drivers

Answer 1

我用命中和试炼的方法解决了这个问题。在内核中，我用 b 改变了 r 的位置，问题解决了，我也必须在 java 中发送 int 数组中的代码而不是字节。

extern "C"
__global__ void skintone(uchar4* data, int w, int h, int* output)
{
    int x = threadIdx.x+blockIdx.x*blockDim.x;
    int y = threadIdx.y+blockIdx.y*blockDim.y;

if (x < w && y < h)
{
    float b, g, r;
    float cb, cr;

    int index = y*w+x;
    uchar4 pixel = data[index];

    b = (float)pixel.x;
    g = (float)pixel.y;
    r = (float)pixel.z;

    cb = 128 + (-0.169f * r) - (0.331f * g) + (0.5f * b);
    cr = 128 + (0.5f * r) - (0.419f * g) - (0.081f * b);


    if((cb >= 80 &&  cb <= 120) && (cr >= 133 &&  cr <= 173)) {
        atomicAdd(&output[0], 1);
    }
}
}

Java 代码更改。

public static void calculateSkintoneGPU() throws IOException {
    BufferedImage img = ImageIO.read(SkinTone.class.getClassLoader().getResource("images.jpg"));
    if (img == null || img.getData() == null)
        return;

    int width = img.getWidth(null);
    int height = img.getHeight(null);
    int[] pixels = new int[width * height];
    PixelGrabber pg = new PixelGrabber(img, 0, 0, width, height, pixels, 0, width);
    try {
        pg.grabPixels();
    } catch (InterruptedException e){};

    int totalPixels = width * height;

    CUfunction kernelFunction = initlize();

    int output[] = execute(kernelFunction, pixels, width, height);
    // Flushing memory
    img.flush();
    img = null;

    long skintoneThreshold = Math.round(output[0] / (double) totalPixels * 100.0);

    System.err.println("Skintone Using GPU: " + output[0]);
    System.err.println("Total Pixel Of GPU: " + totalPixels);
    System.err.println("SKinTone Percentage Using GPU: " + skintoneThreshold + "%");
}

使用 JCuda 计算 Skintone 未给出正确的百分比

Calculate Skintone using JCuda is not giving the right percentage

c

java

cuda

gpu

jcuda