1.对于一个标准的3*3 均值滤波，kernel代码如下：
使用buffer/image缓冲对象
for(i = x - k; i <= x + k; i++)
{
finalcolor = finalcolor + convert_uint4(inputImage[i + j * width]);
}
}
outputImage[x + y * width] = convert_uchar4(finalcolor/n);
}
{
finalcolor = finalcolor + convert_uint4(inputImage[i + j * width]);
}
}
outputImage[x + y * width] = convert_uchar4(finalcolor/n);
}
for(i = x - k; i <= x + k; i++)
{
finalcolor = finalcolor + read_imageui(inputImage, imageSampler, (int2)(i,j));
}
}
finalcolor = finalcolor/n;
write_imageui(outputImage, (int2)(x,y), finalcolor);
}
对一个2048*2048的图像执行filter操作，
global work size = {2048, 2048, 1}, group work size = {16, 16}, 一般group work size应该为64的倍数，因为对于AMD显卡，wave是基本的硬件线程调度单位。
使用了6个GPRs，没有使用ScratchRegs,ScratchRregs是指用vedio meory来模拟GPR，但是线程执行的速度会大大降低，应尽量减少ScratchRegs的数量。
可以看到，使用image对象kernel执行时间要短，但奇怪的是各项性能参数都是buffer对象领先，除了alu busy和alu指令数目。
改为下面的kernel代码，性能会有所提高
/* k*k area */
uint4 finalcolor = (uint4)(0);
finalcolor = finalcolor + read_imageui(inputImage, imageSampler, (int2)(x-1,y-1));
finalcolor = finalcolor + read_imageui(inputImage, imageSampler, (int2)(x,y-1));
finalcolor = finalcolor + read_imageui(inputImage, imageSampler, (int2)(x+1,y-1));
finalcolor = finalcolor + read_imageui(inputImage, imageSampler, (int2)(x-1,y));
finalcolor = finalcolor + read_imageui(inputImage, imageSampler, (int2)(x,y));
finalcolor = finalcolor + read_imageui(inputImage, imageSampler, (int2)(x+1,y));
finalcolor = finalcolor + read_imageui(inputImage, imageSampler, (int2)(x-1,y+1));
finalcolor = finalcolor + read_imageui(inputImage, imageSampler, (int2)(x,y+1));
finalcolor = finalcolor + read_imageui(inputImage, imageSampler, (int2)(x+1,y+1));
finalcolor = finalcolor/9;
write_imageui(outputImage, (int2)(x,y), finalcolor);
}
