mirror of
https://github.com/GPUOpen-LibrariesAndSDKs/VulkanMemoryAllocator
synced 2024-11-05 12:20:07 +00:00
137 lines
13 KiB
HTML
137 lines
13 KiB
HTML
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "https://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
|
|
<html xmlns="http://www.w3.org/1999/xhtml">
|
|
<head>
|
|
<meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/>
|
|
<meta http-equiv="X-UA-Compatible" content="IE=9"/>
|
|
<meta name="generator" content="Doxygen 1.9.1"/>
|
|
<meta name="viewport" content="width=device-width, initial-scale=1"/>
|
|
<title>Vulkan Memory Allocator: Recommended usage patterns</title>
|
|
<link href="tabs.css" rel="stylesheet" type="text/css"/>
|
|
<script type="text/javascript" src="jquery.js"></script>
|
|
<script type="text/javascript" src="dynsections.js"></script>
|
|
<link href="search/search.css" rel="stylesheet" type="text/css"/>
|
|
<script type="text/javascript" src="search/searchdata.js"></script>
|
|
<script type="text/javascript" src="search/search.js"></script>
|
|
<link href="doxygen.css" rel="stylesheet" type="text/css" />
|
|
</head>
|
|
<body>
|
|
<div id="top"><!-- do not remove this div, it is closed by doxygen! -->
|
|
<div id="titlearea">
|
|
<table cellspacing="0" cellpadding="0">
|
|
<tbody>
|
|
<tr style="height: 56px;">
|
|
<td id="projectalign" style="padding-left: 0.5em;">
|
|
<div id="projectname">Vulkan Memory Allocator
|
|
</div>
|
|
</td>
|
|
</tr>
|
|
</tbody>
|
|
</table>
|
|
</div>
|
|
<!-- end header part -->
|
|
<!-- Generated by Doxygen 1.9.1 -->
|
|
<script type="text/javascript">
|
|
/* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */
|
|
var searchBox = new SearchBox("searchBox", "search",false,'Search','.html');
|
|
/* @license-end */
|
|
</script>
|
|
<script type="text/javascript" src="menudata.js"></script>
|
|
<script type="text/javascript" src="menu.js"></script>
|
|
<script type="text/javascript">
|
|
/* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */
|
|
$(function() {
|
|
initMenu('',true,false,'search.php','Search');
|
|
$(document).ready(function() { init_search(); });
|
|
});
|
|
/* @license-end */</script>
|
|
<div id="main-nav"></div>
|
|
<!-- window showing the filter options -->
|
|
<div id="MSearchSelectWindow"
|
|
onmouseover="return searchBox.OnSearchSelectShow()"
|
|
onmouseout="return searchBox.OnSearchSelectHide()"
|
|
onkeydown="return searchBox.OnSearchSelectKey(event)">
|
|
</div>
|
|
|
|
<!-- iframe showing the search results (closed by default) -->
|
|
<div id="MSearchResultsWindow">
|
|
<iframe src="javascript:void(0)" frameborder="0"
|
|
name="MSearchResults" id="MSearchResults">
|
|
</iframe>
|
|
</div>
|
|
|
|
<div id="nav-path" class="navpath">
|
|
<ul>
|
|
<li class="navelem"><a class="el" href="index.html">Vulkan Memory Allocator</a></li> </ul>
|
|
</div>
|
|
</div><!-- top -->
|
|
<div class="PageDoc"><div class="header">
|
|
<div class="headertitle">
|
|
<div class="title">Recommended usage patterns </div> </div>
|
|
</div><!--header-->
|
|
<div class="contents">
|
|
<div class="textblock"><p>See also slides from talk: <a href="https://www.gdcvault.com/play/1025458/Advanced-Graphics-Techniques-Tutorial-New">Sawicki, Adam. Advanced Graphics Techniques Tutorial: Memory management in Vulkan and DX12. Game Developers Conference, 2018</a></p>
|
|
<h1><a class="anchor" id="usage_patterns_common_mistakes"></a>
|
|
Common mistakes</h1>
|
|
<p><b>Use of CPU_TO_GPU instead of CPU_ONLY memory</b></p>
|
|
<p><a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca9066b52c5a7079bb74a69aaf8b92ff67">VMA_MEMORY_USAGE_CPU_TO_GPU</a> is recommended only for resources that will be mapped and written by the CPU, as well as read directly by the GPU - like some buffers or textures updated every frame (dynamic). If you create a staging copy of a resource to be written by CPU and then used as a source of transfer to another resource placed in the GPU memory, that staging resource should be created with <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a>. Please read the descriptions of these enums carefully for details.</p>
|
|
<p><b>Unnecessary use of custom pools</b></p>
|
|
<p><a class="el" href="custom_memory_pools.html">Custom memory pools</a> may be useful for special purposes - when you want to keep certain type of resources separate e.g. to reserve minimum amount of memory for them, limit maximum amount of memory they can occupy, or make some of them push out the other through the mechanism of <a class="el" href="lost_allocations.html">Lost allocations</a>. For most resources this is not needed and so it is not recommended to create <a class="el" href="struct_vma_pool.html" title="Represents custom memory pool.">VmaPool</a> objects and allocations out of them. Allocating from the default pool is sufficient.</p>
|
|
<h1><a class="anchor" id="usage_patterns_simple"></a>
|
|
Simple patterns</h1>
|
|
<h2><a class="anchor" id="usage_patterns_simple_render_targets"></a>
|
|
Render targets</h2>
|
|
<p><b>When:</b> Any resources that you frequently write and read on GPU, e.g. images used as color attachments (aka "render targets"), depth-stencil attachments, images/buffers used as storage image/buffer (aka "Unordered Access View (UAV)").</p>
|
|
<p><b>What to do:</b> Create them in video memory that is fastest to access from GPU using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>.</p>
|
|
<p>Consider using <a class="el" href="vk_khr_dedicated_allocation.html">VK_KHR_dedicated_allocation</a> extension and/or manually creating them as dedicated allocations using <a class="el" href="vk__mem__alloc_8h.html#ad9889c10c798b040d59c92f257cae597a3fc311d855c2ff53f1090ef5c722b38f" title="Set this flag if the allocation should have its own memory block.">VMA_ALLOCATION_CREATE_DEDICATED_MEMORY_BIT</a>, especially if they are large or if you plan to destroy and recreate them e.g. when display resolution changes. Prefer to create such resources first and all other GPU resources (like textures and vertex buffers) later.</p>
|
|
<h2><a class="anchor" id="usage_patterns_simple_immutable_resources"></a>
|
|
Immutable resources</h2>
|
|
<p><b>When:</b> Any resources that you fill on CPU only once (aka "immutable") or infrequently and then read frequently on GPU, e.g. textures, vertex and index buffers, constant buffers that don't change often.</p>
|
|
<p><b>What to do:</b> Create them in video memory that is fastest to access from GPU using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>.</p>
|
|
<p>To initialize content of such resource, create a CPU-side (aka "staging") copy of it in system memory - <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a>, map it, fill it, and submit a transfer from it to the GPU resource. You can keep the staging copy if you need it for another upload transfer in the future. If you don't, you can destroy it or reuse this buffer for uploading different resource after the transfer finishes.</p>
|
|
<p>Prefer to create just buffers in system memory rather than images, even for uploading textures. Use <code>vkCmdCopyBufferToImage()</code>. Dont use images with <code>VK_IMAGE_TILING_LINEAR</code>.</p>
|
|
<h2><a class="anchor" id="usage_patterns_dynamic_resources"></a>
|
|
Dynamic resources</h2>
|
|
<p><b>When:</b> Any resources that change frequently (aka "dynamic"), e.g. every frame or every draw call, written on CPU, read on GPU.</p>
|
|
<p><b>What to do:</b> Create them using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca9066b52c5a7079bb74a69aaf8b92ff67">VMA_MEMORY_USAGE_CPU_TO_GPU</a>. You can map it and write to it directly on CPU, as well as read from it on GPU.</p>
|
|
<p>This is a more complex situation. Different solutions are possible, and the best one depends on specific GPU type, but you can use this simple approach for the start. Prefer to write to such resource sequentially (e.g. using <code>memcpy</code>). Don't perform random access or any reads from it on CPU, as it may be very slow. Also note that textures written directly from the host through a mapped pointer need to be in LINEAR not OPTIMAL layout.</p>
|
|
<h2><a class="anchor" id="usage_patterns_readback"></a>
|
|
Readback</h2>
|
|
<p><b>When:</b> Resources that contain data written by GPU that you want to read back on CPU, e.g. results of some computations.</p>
|
|
<p><b>What to do:</b> Create them using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca7b586d2fdaf82a463b58f581ed72be27">VMA_MEMORY_USAGE_GPU_TO_CPU</a>. You can write to them directly on GPU, as well as map and read them on CPU.</p>
|
|
<h1><a class="anchor" id="usage_patterns_advanced"></a>
|
|
Advanced patterns</h1>
|
|
<h2><a class="anchor" id="usage_patterns_integrated_graphics"></a>
|
|
Detecting integrated graphics</h2>
|
|
<p>You can support integrated graphics (like Intel HD Graphics, AMD APU) better by detecting it in Vulkan. To do it, call <code>vkGetPhysicalDeviceProperties()</code>, inspect <code>VkPhysicalDeviceProperties::deviceType</code> and look for <code>VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU</code>. When you find it, you can assume that memory is unified and all memory types are comparably fast to access from GPU, regardless of <code>VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT</code>.</p>
|
|
<p>You can then sum up sizes of all available memory heaps and treat them as useful for your GPU resources, instead of only <code>DEVICE_LOCAL</code> ones. You can also prefer to create your resources in memory types that are <code>HOST_VISIBLE</code> to map them directly instead of submitting explicit transfer (see below).</p>
|
|
<h2><a class="anchor" id="usage_patterns_direct_vs_transfer"></a>
|
|
Direct access versus transfer</h2>
|
|
<p>For resources that you frequently write on CPU and read on GPU, many solutions are possible:</p>
|
|
<ol type="1">
|
|
<li>Create one copy in video memory using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>, second copy in system memory using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a> and submit explicit transfer each time.</li>
|
|
<li>Create just a single copy using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca9066b52c5a7079bb74a69aaf8b92ff67">VMA_MEMORY_USAGE_CPU_TO_GPU</a>, map it and fill it on CPU, read it directly on GPU.</li>
|
|
<li>Create just a single copy using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca40bdf4cddeffeb12f43d45ca1286e0a5">VMA_MEMORY_USAGE_CPU_ONLY</a>, map it and fill it on CPU, read it directly on GPU.</li>
|
|
</ol>
|
|
<p>Which solution is the most efficient depends on your resource and especially on the GPU. It is best to measure it and then make the decision. Some general recommendations:</p>
|
|
<ul>
|
|
<li>On integrated graphics use (2) or (3) to avoid unnecessary time and memory overhead related to using a second copy and making transfer.</li>
|
|
<li>For small resources (e.g. constant buffers) use (2). Discrete AMD cards have special 256 MiB pool of video memory that is directly mappable. Even if the resource ends up in system memory, its data may be cached on GPU after first fetch over PCIe bus.</li>
|
|
<li>For larger resources (e.g. textures), decide between (1) and (2). You may want to differentiate NVIDIA and AMD, e.g. by looking for memory type that is both <code>DEVICE_LOCAL</code> and <code>HOST_VISIBLE</code>. When you find it, use (2), otherwise use (1).</li>
|
|
</ul>
|
|
<p>Similarly, for resources that you frequently write on GPU and read on CPU, multiple solutions are possible:</p>
|
|
<ol type="1">
|
|
<li>Create one copy in video memory using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305ccac6b5dc1432d88647aa4cd456246eadf7">VMA_MEMORY_USAGE_GPU_ONLY</a>, second copy in system memory using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca7b586d2fdaf82a463b58f581ed72be27">VMA_MEMORY_USAGE_GPU_TO_CPU</a> and submit explicit tranfer each time.</li>
|
|
<li>Create just single copy using <a class="el" href="vk__mem__alloc_8h.html#aa5846affa1e9da3800e3e78fae2305cca7b586d2fdaf82a463b58f581ed72be27">VMA_MEMORY_USAGE_GPU_TO_CPU</a>, write to it directly on GPU, map it and read it on CPU.</li>
|
|
</ol>
|
|
<p>You should take some measurements to decide which option is faster in case of your specific resource.</p>
|
|
<p>Note that textures accessed directly from the host through a mapped pointer need to be in LINEAR layout, which may slow down their usage on the device. Textures accessed only by the device and transfer operations can use OPTIMAL layout.</p>
|
|
<p>If you don't want to specialize your code for specific types of GPUs, you can still make an simple optimization for cases when your resource ends up in mappable memory to use it directly in this case instead of creating CPU-side staging copy. For details see <a class="el" href="memory_mapping.html#memory_mapping_finding_if_memory_mappable">Finding out if memory is mappable</a>. </p>
|
|
</div></div><!-- contents -->
|
|
</div><!-- PageDoc -->
|
|
<!-- start footer part -->
|
|
<hr class="footer"/><address class="footer"><small>
|
|
Generated by <a href="https://www.doxygen.org/index.html"><img class="footer" src="doxygen.svg" width="104" height="31" alt="doxygen"/></a> 1.9.1
|
|
</small></address>
|
|
</body>
|
|
</html>
|