page.title=Vulkan Design Guidelines
@jd:body

<div id="qv-wrapper">
    <div id="qv">
      <h2>On this page</h2>

      <ol>
        <li><a href="#apply">Apply Display Rotation During Rendering</a></li>
        <li><a href="#minimize">Minimize Render Passes Per Frame</a></li>
        <li><a href="#choose">Choose Appropriate Memory Types</a></li>
        <li><a href="#group">Group Descriptor Sets by Frequency</a></li>
      </ol>
    </div>
  </div>

<p>
Vulkan is unlike earlier graphics APIs in that drivers do not perform certain
optimizations, such as pipeline reuse, for apps. Instead, apps using Vulkan must
implement such optimizations themselves. If they do not, they may exhibit worse
performance than apps running OpenGL ES.
</p>

<p>
When apps implement these optimizations themselves, they have the potential
to do so more successfully than the driver can, because they have access to
more specific information for a given use case. As a result, skillfully
optimizing an app that uses Vulkan can yield better performance than if the
app were using OpenGL ES.
</p>

<p>
This page introduces several optimizations that your Android app can implement
to gain performance boosts from Vulkan.
</p>

<h2 id="apply">Apply Display Rotation During Rendering</h2>

<p>
When the upward-facing direction of an app doesn’t match the orientation of the device’s
display, the compositor rotates the application’s swapchain images so that it
does match. It performs this rotation as it displays the images, which results
in more power consumption&mdash;sometimes significantly more&mdash;than if it were not
rotating them.
</p>

<p>
By contrast, rotating swapchain images while generating them results in
little, if any, additional power consumption. The
{@code VkSurfaceCapabilitiesKHR::currentTransform} field indicates the rotation
that the compositor applies to the window. After an app applies that rotation
during rendering, the app uses the {@code VkSwapchainCreateInfoKHR::preTransform}
field to report that the rotation is complete.
</p>

<h2 id="minimize">Minimize Render Passes Per Frame</h2>

<p>
On most mobile GPU architectures, beginning and ending a render pass is an
expensive operation. Your app can improve performance by organizing rendering operations into
as few render passes as possible.
</p>

<p>
Different attachment-load and attachment-store ops offer different levels of
performance. For example, if you do not need to preserve the contents of an attachment, you
can use the much faster {@code VK_ATTACHMENT_LOAD_OP_CLEAR} or
{@code VK_ATTACHMENT_LOAD_OP_DONT_CARE} instead of {@code VK_ATTACHMENT_LOAD_OP_LOAD}. Similarly, if
you don't need to write the attachment's final values to memory for later use, you can use
{@code VK_ATTACHMENT_STORE_OP_DONT_CARE} to attain much better performance than
{@code VK_ATTACHMENT_STORE_OP_STORE}.
</p>

<p>
Also, in most render passes, your app doesn’t need to load or store the
depth/stencil attachment. In such cases, you can avoid having to allocate physical memory for
the attachment by using the {@code VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT}
flag when creating the attachment image. This bit provides the same benefits as does
{@code glFramebufferDiscard} in OpenGL ES.
</p>

<h2 id="choose">Choose Appropriate Memory Types</h2>

<p>
When allocating device memory, apps must choose a memory type. Memory type
determines how an app can use the memory, and also describes caching and
coherence properties of the memory.  Different devices have different memory
types available; different memory types exhibit different performance
characteristics.
</p>

<p>
An app can use a simple algorithm to pick the best memory type for a given
use. This algorithm picks the first memory type in the
{@code VkPhysicalDeviceMemoryProperties::memoryTypes} array that meets two criteria:
The memory type must be allowed for the buffer
or image, and must have the minimum properties that the app requires.
</p>

<p>
Mobile systems generally don’t have separate physical memory heaps for the
CPU and GPU. On such systems, {@code VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT} is not as
significant as it is on systems that have discrete GPUs with their own, dedicated
memory. An app should not assume this property is required.
</p>

<h2 id="group">Group Descriptor Sets by Frequency</h2>

<p>
If you have resource bindings that change at different frequencies, use
multiple descriptor sets per pipeline rather than rebinding all resources for
each draw. For example, you can have one set of descriptors for per-scene
bindings, another set for per-material bindings, and a third set for
per-mesh-instance bindings.
</p>

<p>
Use immediate constants for the highest-frequency changes, such as changes
executed with each draw call.
</p>