Metal编程指南：GPU加速3D图形与计算

需积分: 50 78 浏览量更新于2024-07-22 收藏 1.06MB PDF 举报

"Metal编程指南翻译——iOS Metal 3D技术详解" 在深入探讨Metal编程之前，我们首先理解一下Metal是什么。Metal是Apple开发的一个高性能图形和计算框架，专为iOS、macOS以及tvOS等平台设计，它提供了一个现代且高效的API，用于3D图形渲染和数据并行计算。Metal的主要目的是通过减少CPU在执行GPU任务时的开销，来提高GPU的工作效率和应用程序性能。一、核心概念 1. **命令提交模型**：在Metal中，应用程序通过创建命令缓冲区（Command Buffer）来组织和提交一系列GPU指令。这些指令可以包括设置渲染状态、执行图形或计算任务等。命令缓冲区允许开发者批量提交工作，以减少CPU与GPU之间的通信次数，提高效率。 2. **内存管理模型**：Metal采用了一种基于资源生命周期管理的内存模型。开发者需要明确地分配、复制、释放内存，确保GPU能高效地访问数据。同时，Metal支持纹理和缓冲区对象，它们可以自动在GPU内存和系统内存之间进行优化的交换。 3. **独立编译代码**：Metal支持图形着色器函数和数据并行计算函数，这些函数通常以SPIR-V（可移植着色语言）形式存在，可以被独立编译，便于跨平台使用。开发者可以自定义着色器来实现复杂的图形效果和并行计算任务。二、如何使用Metal API 1. **设置硬件状态**：为了执行图形或数据并行计算任务，开发者需要设置相应的硬件状态，如管线状态对象（Pipeline State Object）、着色器函数、纹理和缓冲区等。 2. **提交命令**：创建命令编码器（Command Encoder），如图形命令编码器或计算命令编码器，然后在这些编码器上记录GPU指令。完成后，将编码器的命令提交到命令缓冲区。 3. **管理内存**：有效地管理内存是优化Metal应用的关键。这包括分配和释放设备内存，以及在适当的时候同步CPU和GPU对内存的访问。 4. **渲染和计算**：通过图形命令编码器，可以设置渲染目标、绘制物体、执行顶点和片段着色器等；而计算命令编码器则用于执行自定义的计算任务，如图像处理、物理模拟等。 5. **资源调度**：Metal允许开发者控制资源的生命周期，包括纹理、缓冲区和其他GPU对象。这包括创建、销毁、保留和释放资源，以及在多个命令缓冲区之间共享资源。 6. **错误处理和调试**：Metal提供了一套工具，如Xcode的GPU框架，用于调试和分析Metal应用程序的性能，帮助开发者找出并解决潜在的问题。 Metal编程指南详细介绍了如何利用这个框架构建高性能的3D图形和并行计算应用。通过理解并掌握上述核心概念和API用法，开发者可以充分利用iOS设备的GPU能力，创造出流畅、精细的图形体验和高效的数据处理应用。

如果你希望并为command buffer执指令编码，那么可以同时创建多个command buffer，使多个线程，每个

线程单独为个command buffer编码指令。如果你事先知道command buffer应该以如何的顺序执，那么

MTLCommandBuffer的enqueue法可以在command queue中声明执顺序，不必等待执编码和提交操作。

要不然就只有等到command buffer被提交，这时在command queue中它就被指定了个位置，位于所有先提交的

command buffer后。

Only one CPU thread can access a command buffer at time. Multithreaded apps can use one thread per command buffer

to create multiple command buffers in parallel.

Figure 2-2 (page 17) shows an example with three threads. Each thread has its own command buffer. For each thread,

one command encoder at a time has access to its associated command buffer. Figure 2-2 (page 17) also shows each

command buffer receiving commands from different command encoders. When you are done encoding, call the

endEncoding method of the command encoder, and a new command encoder object can then begin encoding

commands to the command buffer.

A MTLParallelRenderCommandEncoder object allows a single rendering pass to be broken up across multiple

command encoders and assigned to separate threads. For more information about

MTLParallelRenderCommandEncoder, see Encoding a Single Rendering Pass Using Multiple Threads (page 55).

任时刻只有个GPU线程在访问command buffer，多线程的应可以为每个command buffer准备个线程，如

此实现并创建多个command buffer。

图2-2例了个3线程应，每个线程都操作个command buffer，每个线程中，任时刻只有个Encoder在

访问它对应的command buffer。当你完成了编码，调Encoder的endEncoding法，然后个新的Encoder才可

以开始为command buffer编码指令。

种MTLParallelRenderCommandEncoder对象可以允许个绘制pass被打散成多个Encoder并分配到多个线程

进操作。

Figure 2-2 Metal Command Buffers with Multiple Threads

Resource Objects: Buffers and Textures

资源对象：缓存和纹理

This chapter describes Metal resource objects (MTLResource) for storing unformatted memory and formatted image

data. There are two types of MTLResource objects:

MTLBuffer represents an allocation of unformatted memory that can contain any type of data. Buffers are often used for

vertex, shader, and compute state data.

MTLTexture represents an allocation of formatted image data with a speciﬁed texture type and pixel format. Texture

objects are used as source textures for vertex, fragment, or compute functions, as well as to store graphics rendering

output (i.e., an attachment).

MTLSamplerState objects are also discussed in this chapter. Although samplers are not resources themselves, they are

used when performing lookup calculations with a texture object.

本章节描述Metal的资源对象(MTLResource)，这些对象可以存储格式化的内存数据或是格式化的图像数据，

有下两种MTLResource类型的对象。

MTLBuffer，它表块格式化的内存可以存放任何类型的数据。它通常于存放顶点数据，着和计算状

态数据。

MTLTexture，它表块有格式的图像数据，有特定的纹理类型和像素格式。它通常被作顶点着程序，

元着程序，或是并计算着程序的纹理源。它也可以存放图形绘制的结果输出。

MTLSamplerState类型的对象也将在这个章节讲述，虽然采样器不是资源，但它总是于纹理对象的查找计

算。

Buffers are Typeless Allocations of Memory

缓存是类型的内存段

A MTLBuffer object represents an allocation of memory that can contain any type of data.

个MTLBuffer对象表了个可以装载任类型数据的内存段。

Creating a Buffer Object

创建个缓存对象

The following MTLDevice methods create and return a MTLBuffer object:

The newBufferWithLength:options: method creates a MTLBuffer object with a new storage allocation.

The newBufferWithBytes:length:options: method creates a MTLBuffer object by copying data from existing

storage (located at the CPU address pointer) into a new storage allocation.

The newBufferWithBytesNoCopy:length:options:deallocator: method creates a MTLBuffer object with an

existing storage allocation and does not allocate any new storage for this object.

All buffer creation methods have the input value length to indicate the size of the storage allocation, in bytes. All the

methods also accept a MTLResourceOptions object for options that can modify the behavior of the created buffer. If

the value for options is 0, the default values are used for resource options.

下的MTLDevice法户创建并返回类型MTLBuffer对象：

newBufferWithLength:options:，该法新分配内存创建个MTLBuffer对象。

newBufferWithBytes:length:options:，该法新分配内存创建个MTLBuffer对象，并且从已有的存储（由

CPU内存地址指针指定）拷贝数据到新分配的内存中。

newBufferWithBytesNoCopy:length:options:deallocator:，该法创建个MTLBuffer对象，但是不会为

这个缓存对象分配新内存，是使个已经存在的内存。

以上所有的创建法都有叫做length的参，它来指定分配存储的多少（按byte）；他们还有叫做的options

参，接受个MTLResourceOptions类型的值，来设定缓存创建的为。options的默认值为0。

Buffer Methods

缓存对象法

The MTLBuffer protocol has the following methods:

The contents method returns the CPU address of the buffer’s storage allocation.!

The newTextureWithDescriptor:offset:bytesPerRow: method creates a special kind of texture

object that references the buffer's data. This method is detailed in Creating a Texture Object (page 19).

缓存对象有如下的法：

contents，该法返回缓存对象对应的内存的CPU地址。

newTextureWithDescriptor:offset:bytesPerRow:，该法创建某种特定类型的纹理。

Textures are Formatted Image Data

纹理是格式化的图像数据

A MTLTexture object represents an allocation of formatted image data that can be used as a resource for a vertex

shader, fragment shader, or compute function, or as an attachment to be used as a rendering destination. A MTLTexture

object can have one of the following structures:

a 1D, 2D, or 3D image, arrays of 1D or 2D images, or a cube of six 2D images.

MTLPixelFormat speciﬁes the organization of individual pixels in a MTLTexture object. Pixel formats are discussed

further in Pixel Formats for Textures (page 23).

个MTLTexture对象代表了个格式化的图像数据内存段，它在顶点着程序、元着程序、并计算着

程序中被作输源，或是作为绘制操作的标输出。个对象有如下结构：

个1维，2维，或是3维图像

个含有1维或是2维的图像的数组

拥有6个2维图像的体

MTLPixelFormat，该属性描述了个MTLTexture对象中的每个单独像素的组织排列。

Creating a Texture Object

创建纹理对象

The following methods create and return a MTLTexture object:

The newTextureWithDescriptor: method of MTLDevice creates a MTLTexture object with a new storage

allocation for the texture image data, using a MTLTextureDescriptor object to describe the texture’s properties.

The newTextureViewWithPixelFormat: method of MTLTexture creates a MTLTexture object that shares the same

storage allocation as the calling MTLTexture object. Since they share the same storage, any changes to the pixels of the

new texture object are reﬂected in the calling texture object, and vice versa. For the newly created texture, the

newTextureViewWithPixelFormat: method reinterprets the existing texture image data of the storage allocation of

the calling MTLTexture object as if the data was stored in the speciﬁed pixel format. The MTLPixelFormat of the new

texture object must be compatible with the MTLPixelFormat of the original texture object. (See Pixel Formats for

Textures (page 23) for details about the ordinary, packed, and compressed pixel formats.)

The newTextureWithDescriptor:offset:bytesPerRow: method of MTLBuffer creates a MTLTexture object that

shares the storage allocation of the calling MTLBuffer object as its texture image data. As they share the same storage,

any changes to the pixels of the new texture object are reﬂected in the calling texture object, and vice versa. Sharing

storage between a texture and a buffer may prevent applying certain texturing optimizations (e.g., pixel swizzling or

tiling).

剩余63页未读，继续阅读

普通网友

粉丝: 4
资源:
2

Metal编程指南：GPU加速3D图形与计算

iOS12 Metal by Tutorials-2018 .pdf

Metal-Shading-Language-Specification.pdf

Metal

metal编程指南 pdf

metal编程指南和参考(中文).pdf

metal.programming.guide.tutorial.and.reference.via.swift

ios metal 书

metal有什么神经网络

base层和metal层

用metal深度学习

最新资源