Question

我正在尝试使用gpu.js运行一维元胞自动机，但是没有gpu.js的代码看起来要快20倍。我不明白问题出在哪里... 下面的代码从一个ON单元开始运行规则30的单元自动机。自动机的长度为2001个细胞。该代码使用CPU和GPU来计算前10000代，对于CPU来说需要0.16秒，而对于GPU则需要3.2秒。

<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN" "http://www.w3.org/TR/html4/loose.dtd">
<html lang="en"> 
	<head> 
		<meta http-equiv="content-type" content="text/html; charset=utf-8"> 
		<title>Title Goes Here</title> 

<script src="https://cdnjs.cloudflare.com/ajax/libs/gpu.js/1.10.4/gpu.min.js"></script>

	</head> 
	<body>
	 	<script>

			 gen = [];
			 for (var i = 0; i<1000; i++) gen.push(0);
			 gen.push(1);	
			 for (var i = 0; i<1000; i++) gen.push(0);

		    const gpu = new GPU();
			const kernel = gpu.createKernel(function(x) {
				const x1 = x[this.thread.x-1];
				const x2 = x[this.thread.x];
				const x3 = x[this.thread.x+1];
			    return x1+x2+x3-2*x1*x2-x2*x3-2*x1*x3+2*x1*x2*x3;
			}).setOutput([gen.length]);

			kernelCPU = function(x){
				var res = [];
				x.unshift(0);
				x.push(0);
				for (var i = 1; i < x.length-1; i++) {
					x1 = x[i-1];
				 	x2 = x[i];
				 	x3 = x[i+1];
				 	res.push(x1+x2+x3-2*x1*x2-x2*x3-2*x1*x3+2*x1*x2*x3);
				}
				return res;

			}
			var t0 = performance.now();
			for (var i = 0; i< 10000; i++) {
			
				gen = kernelCPU(gen);
			}
			var t1 = performance.now();
			console.log("Took " + (t1 - t0)/1000 + " seconds for CPU")
			var t0 = performance.now();
			for (var i = 0; i< 10000; i++) {
			
				gen = kernel(gen);
			}
			var t1 = performance.now();
			console.log("Took " + (t1 - t0)/1000 + " seconds for GPU.")
		</script>
	</body> 
</html>

Answer 1

您的问题出在循环中

for (var i = 0; i< 10000; i++) {
    gen = kernel(gen);
}

调用GPU内核函数会产生一些非零的开销。您只有2001个可以并行计算的单元-在调用GPU内核函数的初始开销之后，这些单元的计算速度要比CPU上的计算速度快得多。但是，因为遇到GPU调用开销 10,000 次，您所付出的不仅仅是获得的任何性能提升。

使用gpu.js进行细胞自动机

1 个答案: