By Mike Abrash
Last night, my daughter asked me if I could take her to Borders to get a Connie Willis book. One of the longest-standing rules in our house is that reading is good, and any book she wants we''ll get for her, so we set off for the store.
Now, I knew that Christmas was only four days away, but I hadn''t really considered the implications, which turned out to be dire, to say the least. The line from the cash registers stretched back past Literature and History all the way into Social Sciences—more than 100 feet in all. They had all the cash registers going, but unfortunately that only amounted to four or five all told. So my daughter picked out her book, then browsed around the store while I waited in line. And waited. And waited. At least I got to read the first part of her book, which wasn''t bad.
And so we learn something about bottlenecks. All the books in the world wouldn''t have enabled Borders to handle any more customers, but one more register would have made a huge difference. The secret to performance is putting resources in the right place to ease the bottlenecks—which is a very appropriate way to kick off a discussion of data flow through the T&L pipeline. The T&L pipeline bristles with potential bottlenecks—and also with caches that, properly applied, shrink those bottlenecks almost magically.
And with that, onward to T&L pipeline data flow and a variety of associated performance issues!
上面的这段是XDK中的White Papers中的一篇文章的开头,这篇文章是我见过的关于顶点Transform and Lighting部分讲得最详细的。这篇文章不是讲顶点该如何Transform该如何Lighting,而是将上面的过程作为black boxes一带而过,重点是讲GPU采取什么样子的方式读取数据,数据采取什么方式流动,并且详细分析了整个数据流每个部分的协作和如何达到最大的效率。
通过阅读这篇文章,你将对XBOX(也能参考于PC)的顶点处理有了一个全面而详细地了解,并知道如何最大的利用硬件的功能。
PS:XDK可以在EDonkey上下载到。
