CUDA中统一虚拟寻址（UVA）的信息/示例 [英] Information/example on Unified Virtual Addressing (UVA) in CUDA

问题描述

我想了解CUDA中统一虚拟寻址（UVA）的概念。我有两个问题：

I'm trying to understand the concept of Unified Virtual Addressing (UVA) in CUDA. I have two questions:

1. 是否有任何示例（psudo）代码演示了这个概念？

1. Is there any sample (psudo)code available that demonstrates this concept?

我在CUDA C编程指南中读到，UVA只能用于64位操作系统。为什么会这样？

I read in the CUDA C Programming Guide that UVA can be used only with 64 bit operating systems. Why it is so?

推荐答案

A 将指针（值）和设备代码中使用的分配映射与指针组合在一起（值）和主机代码中使用的分配映射到单个统一空间中。

A unified virtual address space combines the pointer (values) and allocation mappings used in device code with the pointer (values) and allocation mappings used in host code into a single unified space.

1 - 是否有可用的展示这个概念？

1-Is there any sample (psudo)code available that demonstrate this concept?

这个方法的一个例子是使用 cudaMemcpyDefault 转移类型说明符，代替 cudaMemcpyDeviceToHost 或 cudaMemcpyHostToDevice 在 cudaMemcpy 操作。

One example of how this may be used is in the usage of the cudaMemcpyDefault transfer kind specifier, in place of cudaMemcpyDeviceToHost or cudaMemcpyHostToDevice in cudaMemcpy operations.

2 - 该编程指南仅与64位操作系统相关。为什么会这样？

2-Also I read in the programming guide that it is associated only with 64 bit operating system. Why it is so?

32位计算机的最大可寻址空间为4 GB。在现代GPU（可能具有多吉字节的存储器）和现代PC（可能具有多吉字节的存储器和I / O空间）中，这不是方便地具有在其中定义主机和设备存储器空间的足够空间。

A 32 bit machine has a maximum addressable space of 4 Gigabytes. In modern GPUs (with perhaps multiple gigabytes of memory) and modern PCs (with perhaps multiple gigabytes of memory and I/O space) this is not conveniently enough space to have both the host and device memory spaces defined in it.

这篇关于CUDA中统一虚拟寻址（UVA）的信息/示例的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！