copy from http://developer.amd.com/community/blog/2014/10/31/opencl-2-0-pipes/

OpenCL™ 2.0 – Pipes

In the previous post, we saw one of the important features of OpenCL™ 2.0, Shared Virtual Memory (SVM). In this blog, we will see another feature of OpenCL 2.0 called “pipes”.

To get the most from our discussions, we recommend the same approach as in the previous post:

Review the code snippets in each blog post along with the explanatory text.

Download the AMD OpenCL 2.0 Driver located here. This page also has a complete list of supported platforms.

Download the sample code from our OpenCL 2.0 Samples page here.

Write your own OpenCL 2.0 samples and share your results with the OpenCL community.

The code will run on a variety of AMD platforms, such as the Radeon HD8000 series. The driver page has a complete list of supported product families.

Overview

OpenCL 2.0 introduces a new mechanism for passing data between kernels, called “pipes.” A pipe is essentially a structured buffer containing some space for a set of “packets”—kernel-specified type objects. As the name suggests, these packets of data are ordered in the pipe. There is a write end of the pipe into which the data is written and a read end of the pipe from which the data is read. A pipe is essentially an addition to buffer objects, such as buffers and images. Pipes can only be accessed using the built-in functions provided by the kernel and cannot be accessed from the host.

Special built-in functions, read_pipe and write_pipe, provide access to pipes from the kernel. A given kernel can either read from or write to a pipe, but not both. Pipes are only coherent at the standard synchronization points; the result of concurrent accesses to the same pipe by multiple kernels (even hardware permitting) is undefined. The host side cannot access a pipe.

It is easy to create pipes. On the host, invoke clCreatePipe, and you are done.

You can use the pipes for a variety of functions. You can pass the pipes between kernels. Even better, combine pipes with the device-size enqueue feature in OpenCL 2.0 to dynamically construct computational data flow graphs.

Pipes come in two types: a read pipe, from which a number of packets can be read, and a write pipe, to which a number of packets can be written.

Note: You cannot write to a pipe specified as read-only, nor can you read from a pipe specified as write-only. You cannot both read from and write to a pipe at the same time.

Functions for Accessing Pipes

OpenCL 2.0 adds a new host API function to create a pipe.

cl_mem clCreatePipe ( cl_context context, cl_mem_flags flags,
cl_uint pipe_packet_size, cl_uint pipe_max_packets,
const cl_pipe_properties * properties,
cl_int *errcode_ret)

The memory allocated in this function can pass to kernels as read-only or write-only pipes. The pipe objects can only be passed as kernel arguments or kernel functions and cannot be declared inside a kernel or as program-scoped objects.

Also, the OpenCL 2.0 spec adds a set of built-in functions for operating on the pipes. The important ones are:

  • read_pipe (pipe p, gentype *ptr): for reading packet from pipe p into ptr.
  • write_pipe (pipe p, gentype *ptr): for writing packet pointed to by ptr to pipe p.

To ensure you have enough space in the pipe structure for reading and writing (before you actually do it), you can use built-in functions to “reserve” enough space. For example, you could reserve room by calling reserve_read_pipe or reserve_write_pipe. These functions return a reservation ID, which can be used when the actual operations are performed. Similarly, the standard has built-in functions for workgroup level reservations, such as work_group_reserve_read_pipe and work_group_reserve_write_pipe and for the workgroup order (in the program). These workgroup built-in functions operate at the workgroup level. Ordering across workgroups is undefined. Calls to commit_read_pipe and commit_write_pipe, as the names suggest, commit the actual operations (read/write).

Using Pipes—A Simple Example

Let’s look at a typical usage of pipes in the example code (SimplePipe). The code contains two kernels: producer_kernel, which writes to the pipe, and consumer_kernel, which reads from the same pipe. In the example, the producer writes a sequence of random numbers; the consumer reads them and creates a histogram.

The host creates the pipe, which both kernels will use, as follows:

rngPipe = clCreatePipe(context,
CL_MEM_READ_WRITE,
szPipePkt,
szPipe,
NULL,
&status);

This code makes a pipe that the program kernels can access (read/write). The host creates two kernels, producer_kernel and consumer_kernel. The producer kernel first reserves enough space for the write pipe:

//reserve space in pipe for writing random numbers.
reserve_id_t rid = work_group_reserve_write_pipe(rng_pipe, szgr);

Next, the kernel writes and commits to the pipe by invoking the following functions:

write_pipe(rng_pipe,rid,lid, &gfrn);
work_group_commit_write_pipe(rng_pipe, rid);

Similarly, the consumer kernel reads from the pipe:

//reserve pipe for reading
reserve_id_t rid = work_group_reserve_read_pipe(rng_pipe, szgr);
if(is_valid_reserve_id(rid)) {
//read random number from the pipe.
read_pipe(rng_pipe,rid,lid, &rn);
work_group_commit_read_pipe(rng_pipe, rid);
}

The consumer_kernel then uses this set of random number and constructs the histogram. The CPU creates the same histogram and verifies whether the histogram created by the kernel is correct. Here, lid is the local id of the work item, obtained by get_local_id(0).

The example code demonstrates how you can use a pipe as a convenient data structure that allows two kernels to communicate. It’s really pretty simple.

In OpenCL 1.2, this kind of communication typically involves the host – although kernels can communicate without returning control to the host. Pipes, however, ease programming by reducing the amount of code that some applications require. There are additional examples of pipes used in conjunction with device enqueue, which we will explore in later blogs in this series.

To conclude, using pipes in OpenCL 2.0 can make your code simpler and more readable. Don’t believe us? Write your own OpenCL 2.0 programs and tell us about the difference!

Sample code and readme

The sample code demonstrates the use of the pipes feature of OpenCL.2.0 using the SimplePipe (producer/consumer kernels) sample:

  • The sample code and readme are provided here.
  • Install AMD OpenCL 2.0 Driver located here. This page also has a complete list of supported platforms.
  • Build the sample using the instructions in the readme.
  • Let us know your feedback and comments on the Developer Central OpenCL forum.

– Prakash Raghavendra

Dr. Prakash Raghavendra is a technical lead at AMD. He has several years of experience in developing compilers and run-time. His postings are his own opinions and may not represent AMD’s positions, strategies or opinions. Links to third party sites, and references to third party trademarks, are provided for convenience and illustrative purposes only. Unless explicitly stated, AMD is not responsible for the contents of such links, and no third party endorsement of AMD or any of its products is implied.

OpenCL™ 2.0 – Pipes的更多相关文章

  1. OpenCL介绍

    OpenCL(全称Open Computing Language,开放运算语言)是第一个面向异构系统通用目的并行编程的开放式.免费标准,也是一个统一的编程环境,便于软件开发人员为高性能计算服务器.桌面 ...

  2. OpenCL Kernel设计优化

    使用Intel® FPGA SDK for OpenCL™ 离线编译器,不需要调整kernel代码便可以将其最佳的适应于固定的硬件设备,而是离线编译器会根据kernel的要求自适应调整硬件的结构. 通 ...

  3. 基于SoCkit的opencl实验1-基础例程

    基于SoCkit的opencl实验1-基础例程 准备软硬件 Arrow SoCkit Board 4GB or larger microSD Card Quartus II v14.1 SoCEDS ...

  4. 面向OPENCL的ALTERA SDK

    面向OPENCL的ALTERA SDK 使用面向开放计算语言 (OpenCL™) 的 Altera® SDK,用户可以抽象出传统的硬件 FPGA 开发流程,采用更快.更高层面的软件开发流程.在基于 x ...

  5. OpenCL中三种内存创建image的效率对比

    第一种:使用ION: cl_mem_ion_host_ptr ion_host_ptr1; ion_host_ptr1.ext_host_ptr.allocation_type = CL_MEM_IO ...

  6. OpenCL科普及在ubuntu 16.04 LTS上的安装

    OpenCL(Open Computing Language,开放计算语言)是一个为异构平台编写程序的框架,此异构平台可由CPU.GPU.DSP.FPGA或其他类型的处理器與硬體加速器所组成.Open ...

  7. OpenCL 查询平台和设备

    ▶ 查询平台和设备的代码以结果,放在这里方便以后逐渐扩充和查询(没有营养) #include <stdio.h> #include <stdlib.h> #include &l ...

  8. OpenCL Hello World

    ▶ OpenCL 的环境配置与第一个程序 ● CUDA 中自带 OpenCL 需要的头文件和库,直接拉近项目里边去就行:AMD 需要下载 AMD APP SDK(https://community.a ...

  9. OpenCL双边滤波实现美颜功能

    OpenCL是一个并行异构计算的框架,包括intel,AMD,英伟达等等许多厂家都有对它的支持,不过英伟达只到1.2版本,主要发展自己的CUDA去了.虽然没有用过CUDA,但个人感觉CUDA比Open ...

随机推荐

  1. 文件与IO-字节输入/输出

    package IoDemo; import java.io.File; import java.io.FileInputStream; import java.io.FileNotFoundExce ...

  2. java 与大数据学习较好的网站

    C# C#中 Thread,Task,Async/Await,IAsyncResult 的那些事儿!https://www.cnblogs.com/doforfuture/p/6293926.html ...

  3. FckEditor 配置手册中文教程详细说明

    http://www.jb51.net/article/17965.htm 首先,FCKEDITOR的性能是非常好的,用户只需很少的时间就可以载入 FCKEDITOR所需文件.对于其他在线编辑器来说, ...

  4. Gamma函数相关matlab代码

    1.Gamma函数: Gamma函数matlab代码: x=:0.5:5syms t y=)*exp(-t),,inf) y=double(y) plot(x,y,) 图像如下: 2.lgΓ(x)函数 ...

  5. tomcat_下载

    1. http://tomcat.apache.org/ 2. 3.

  6. 用HyperMesh(7.0)手工修改网格

    检查网格质量并修改 2D->qulifyindex:其中View/Edit/Optimize工具栏功能常用,尤其是Edit页中的Place Node,可以动态地拖动节点并观察单元质量是否符合要求 ...

  7. JQuery中选择元素的方法:

    document.getElementById('div1');document.getElementsByTagName('div');getByClass(document,'box'); $(' ...

  8. jQuery中hover和blur使用代理delegate无效的解决方法

    今天就遇到了这样的小问题: $(document).ready(function(){ $('.status_on').hover(function(){ $(this).html('点击禁用'); ...

  9. Selenium with Python 003 - 页面元素定位

    WebUI自动化,首先需要定位页面中待操作的元素,然后进行各种事件操作,这里我们首先介绍Selenium Python 如何定位页面元素,WebDriver 提供了一系列的方法. 定位单个页面元素(返 ...

  10. ddt读取json文件测试用例的执行顺序

    一. 源码的说明 在源码中,ddt的file_data函数下有这样一段话 意思是说,如果json文件的内容是字典,字典的键名将会作为测试用例名的后缀,字典的值将会作为测试数据,如果这样的话,如果键名字 ...