热点科技

标题: [高阶] 微软的GPGPU计划：The Accelerator Project [打印本页]

作者: xjb785 时间: 2007-2-21 12:39
标题: [高阶] 微软的GPGPU计划：The Accelerator Project
微软的GPGPU计划：The Accelerator Project
Summary: Accelerator provides a high-level data-parallel programming model as a library that is available for all .Net programming languages. The library translates the data-parallel operations on-the-fly to optimized GPU pixel shader code and API calls. Future versions will target multi-core cpus.

How to use this Wiki:A wiki is a Web site that lets you edit and update the content on it yourself. To be able to update this Wiki, you must get an account on Channel 9. Just click on the "Login" button on the upper right-hand side of the screen and then click on the "Create New Account link. Once you have that, login and go back to this page. There will now be an edit button on the left-hand side of the screen. Click on that if you want to add, modify, or update content.

Links:We have replaced our older technical report with a more up-to-date paper. The new paper can be found at [ftp://ftp.research.microsoft.com/pub/tr/TR-2005-184.pdf]. We will be presenting the paper at ASPLOS '06 in October, 2006. The paper is copyright ACM, 2006. It is the author's version of the work and is posted here by permission of ACM for your personal use.

Download:Version 1.1 of Accelerator is now available. It was uploaded October 25, 2006. Please upgrade to this version. It has significantly improved performance as well as recent bug fixes.

You can download it from http://research.microsoft.com/research/downloads/Details/25e1bea3-142e-4694-bde5-f0d44f9d8709/Details.aspx?CategoryID. Alternatively, you can go to http://research.microsoft.com/research/downloads and search for 'Microsoft Research Accelerator Project'.

Announcements

arties interested in inquiring about commercial licensing of the Microsoft Research Accelerator software should contact msrlg@microsoft.com for more information.

If you are having problems running the samples after installation, please download the DirectX end user runtime from http://www.microsoft.com/downloads/details.aspx?FamilyID=2da43d38-db71-4c1b-bc6a-9b6652cd92a3
Activities:Help us build up Click to read this topic 1/30/2007 3:40:01 PM - 000300008245882B

OurFAQ by posting your questions.
Consult our Click to read this topic 10/17/2006 5:00:54 PM - 0003BFFD833E5504

TroubleShooting page if you are having problems.
Check the Click to read this topic 6/2/2006 6:27:43 PM -

WishList to see if your desired feature is on the drawing board.
Read a Click to read this topic 6/1/2006 4:15:01 PM -

ShortIntro to the library.
Send email to the development team : mailto:msraccde@microsoft.com

We are eager to get feedback on the API and to hear about any problems. We are also interested in keeping track of the different application areas where the library is being used.

这个东西号称"provides a high-level data-parallel programming model as a library that is available for all .Net programming languages. The library translates the data-parallel operations on-the-fly to optimized GPU pixel shader code and API calls. Future versions will target multi-core cpus"，如果你正准备弄一些GPGPU的应用又认为NVIDIA的CUDA方案有控制同行嫌疑的话，可以考虑这个东西。

作者: yuguaiguai 时间: 2007-2-22 10:25
谢谢分享

顺便提一下, 里面的sample就有一个简单的Pi算例。

作者: ly9733 时间: 2007-2-22 15:05
有什么办法能让这个computePi可以算到1M位呢，更改矩阵的尺寸结果有时候差别很大？

作者: kele876 时间: 2007-2-23 03:05
雖然跑起GPGPU like的工作，CELL很多時候有著比GPU更好的效率，但是考慮普及性的話CELL踏入x86市場的機會真的很低，GPU卻已經無法從PC市場中除外了....

所以即使高階市場上CELL應該可以持續領先下去（CELL持續有擴充規模的計畫），但是真正要比普及的話GPU應該還是會佔上風。PC上的軟體資源也應該會比較顯著地支援GPGPU市場。

作者: sheng128 时间: 2007-2-23 18:43

原帖由 Edison 于 2007-2-22 15:05 发表
有什么办法能让这个computePi可以算到1M位呢，更改矩阵的尺寸结果有时候差别很大？

那个算例也是不可能的, 顶多算到单精度。矩阵的尺寸我没有试验过, 可能原因是算例太简单, 没有考虑GPU的texture cache的配置。

GPGPU方面我觉得目前还是玩具而已...

作者: mirdy 时间: 2007-2-24 12:03
矩阵规模还是受制于texture size的上限，的确麻烦。

作者: ZSL1119 时间: 2007-2-25 14:11

原帖由 Edison 于 2007-2-24 12:03 发表
矩阵规模还是受制于texture size的上限，的确麻烦。

如果透過Mega Texture相關的發展，讓GPU走向超大texture size的話，應該會改善這部份。
某種意味上GPU也該支援CPU等級的記憶體容量了。

作者: chenwudi 时间: 2007-2-26 11:54
GPU支持CPU等级的内存管理恐怕需要WGF2.0比较有希望。
至于texture input size, 可以通过shader based paging或者3d texture来绕过去的。但是output size的限制恐怕就还没什么办法通过图形API来解决。

欢迎光临热点科技 (http://www.itheat.com/activity/)