performance/rendering/profile-gpu.jd

page.title=Analyzing with Profile GPU Rendering
page.metaDescription=Use the Profile GPU tool to help you optimize your app's rendering performance.

meta.tags="power"
page.tags="power"

@jd:body

<div id="qv-wrapper">
<div id="qv">

<h2>In this document</h2>
    <ol>
      <li>
        <a href="#visrep">Visual Representation</a></li>
      </li>

      <li>
       <a href="#sam">Stages and Their Meanings</a>

      <ul>
         <li>
           <a href="#sv">Input Handling</a>
         </li>
         <li>
           <a href="#asd">Animation</a>
         </li>
         <li>
           <a href="#asd">Measurement/Layout</a>
         </li>
         <li>
           <a href="#asd">Drawing</a>
         </li>
         </li>
         <li>
           <a href="#asd">Sync/Upload</a>
         </li>
         <li>
           <a href="#asd">Issuing Commands</a>
         </li>
         <li>
           <a href="#asd">Processing/Swapping Buffer</a>
         </li>
         <li>
           <a href="#asd">Miscellaneous</a>
         </li>
      </ul>
      </li>
     </ol>
  </div>
</div>

<p>
The <a href="/studio/profile/dev-options-rendering.html">
Profile GPU Rendering</a> tool indicates the relative time that each stage of
the rendering pipeline takes to render the previous frame. This knowledge
can help you identify bottlenecks in the pipeline, so that you
can know what to optimize to improve your app's rendering performance.
</p>

<p>
This page briefly explains what happens during each pipeline stage, and
discusses issues that can cause bottlenecks there. Before reading
this page, you should be familiar with the information presented in the
<a href="/studio/profile/dev-options-rendering.html">Profile GPU
Rendering Walkthrough</a>. In addition, to understand how all of the
stages fit together, it may be helpful to review
<a href="https://www.youtube.com/watch?v=we6poP0kw6E&index=64&list=PLWz5rJ2EKKc9CBxr3BVjPTPoDPLdPIFCE">
how the rendering pipeline works.</a>
</p>

<h2 id="#visrep">Visual Representation</h2>

<p>
The Profile GPU Rendering tool displays stages and their relative times in the
form of a graph: a color-coded histogram. Figure 1 shows an example of
such a display.
</p>

  <img src="{@docRoot}topic/performance/images/bars.png">
  <p class="img-caption">
<strong>Figure 1.</strong> Profile GPU Rendering Graph
  </p>

</p>

<p>
Each segment of each vertical bar displayed in the Profile GPU Rendering
graph represents a stage of the pipeline and is highlighted using a specific
color in
the bar graph. Figure 2 shows a key to the meaning of each displayed color.
</p>

  <img src="{@docRoot}topic/performance/images/s-profiler-legend.png">
  <p class="img-caption">
<strong>Figure 2.</strong> Profile GPU Rendering Graph Legend
  </p>

<p>
Once you understand what each color signfiies,
you can target specific aspects of your
app to try to optimize its rendering performance.
</p>

<h2 id="sam">Stages and Their Meanings</a></h2>

<p>
This section explains what happens during each stage corresponding
to a color in Figure 2, as well as bottleneck causes to look out for.
</p>


<h3 id="ih">Input Handling</h3>

<p>
The input handling stage of the pipeline measures how long the app
spent handling input events. This metric indicates how long the app
spent executing code called as a result of input event callbacks.
</p>

<h4>When this segment is large</h4>

<p>
High values in this area are typically a result of too much work, or
too-complex work, occurring inside the input-handler event callbacks.
Since these callbacks always occur on the main thread, solutions to this
problem focus on optimizing the work directly, or offloading the work to a
different thread.
</p>

<p>
Its also worth noting that {@link android.support.v7.widget.RecyclerView}
scrolling can appear in this phase.
{@link android.support.v7.widget.RecyclerView} scrolls immediately when it
consumes the touch event. As a result,
it can inflate or populate new item views. For this reason, its important to
make this operation as fast as possible. Profiling tools like Traceview or
Systrace can help you investigate further.
</p>

<h3 id="at">Animation</h3>

<p>
The Animations phase shows you just how long it took to evaluate all the
animators that were running in that frame. The most common animators are
{@link android.animation.ObjectAnimator},
{@link android.view.ViewPropertyAnimator}, and
<a href="/training/transitions/overview.html">Transitions</a>.
</p>

<h4>When this segment is large</h4>

<p>
High values in this area are typically a result of work thats executing due
to some property change of the animation. For example, a fling animation,
which scrolls your {@link android.widget.ListView} or
{@link android.support.v7.widget.RecyclerView}, causes large amounts of view
inflation and population.
</p>

<h3 id="ml">Measurement/Layout</h3>

<p>
In order for Android to draw your view items on the screen, it executes
two specific operations across layouts and views in your view hierarchy.
</p>

<p>
First, the system measures the view items. Every view and layout has
specific data that describes the size of the object on the screen. Some views
can have a specific size; others have a size that adapts to the size
of the parent layout container
</p>

<p>
Second, the system lays out the view items. Once the system calculates
the sizes of children views, the system can proceed with layout, sizing
and positioning the views on the screen.
</p>

<p>
The system performs measurement and layout not only for the views to be drawn,
but also for the parent hierarchies of those views, all the way up to the root
view.
</p>

<h4>When this segment is large</h4>

<p>
If your app spends a lot of time per frame in this area, it is
usually either because of the sheer volume of views that need to be
laid out, or problems such as
<a href="/topic/performance/optimizing-view-hierarchies.html#double">
double taxation</a> at the wrong spot in your
hierarchy. In either of these cases, addressing performance involves
<a href="/topic/performance/optimizing-view-hierarchies.html">improving
the performance of your view hierarchies</a>.
</p>

<p>
Code that youve added to
{@link android.view.View#onLayout(boolean, int, int, int, int)} or
{@link android.view.View#onMeasure(int, int)}
can also cause performance
issues. <a href="/studio/profile/traceview.html">Traceview</a> and
<a href="/studio/profile/systrace.html">Systrace</a> can help you examine
the callstacks to identify problems your code may have.
</p>

<h3 id="draw">Drawing</h3>

<p>
The draw stage translates a views rendering operations, such as drawing
a background or drawing text, into a sequence of native drawing commands.
The system captures these commands into a display list.
</p>

<p>
The Draw bar records how much time it takes to complete capturing the commands
into the display list, for all the views that needed to be updated on the screen
this frame. The measured time applies to any code that you have added to the UI
objects in your app. Examples of such code may be the
{@link android.view.View#onDraw(android.graphics.Canvas) onDraw()},
{@link android.view.View#dispatchDraw(android.graphics.Canvas) dispatchDraw()},
and the various <code>draw ()methods</code> belonging to the subclasses of the
{@link android.graphics.drawable.Drawable} class.
</p>

<h4>When this segment is large</h4>

<p>
In simplified terms, you can understand this metric as showing how long it took
to run all of the calls to
{@link android.view.View#onDraw(android.graphics.Canvas) onDraw()}
for each invalidated view. This
measurement includes any time spent dispatching draw commands to children and
drawables that may be present. For this reason, when you see this bar spike, the
cause could be that a bunch of views suddenly became invalidated. Invalidation
makes it necessary to regenerate views' display lists. Alternatively, a
lengthy time may be the result of a few custom views that have some extremely
complex logic in their
{@link android.view.View#onDraw(android.graphics.Canvas) onDraw()} methods.
</p>

<h3 id="su">Sync/Upload</h3>

<p>
The Sync & Upload metric represents the time it takes to transfer
bitmap objects from CPU memory to GPU memory during the current frame.
</p>

<p>
As different processors, the CPU and the GPU have different RAM areas
dedicated to processing. When you draw a bitmap on Android, the system
transfers the bitmap to GPU memory before the GPU can render it to the
screen. Then, the GPU caches the bitmap so that the system doesnt need to
transfer the data again unless the texture gets evicted from the GPU texture
cache.
</p>

<p class="note"><strong>Note:</strong> On Lollipop devices, this stage is
purple.
</p>

<h4>When this segment is large</h4>

<p>
All resources for a frame need to reside in GPU memory before they can be
used to draw a frame. This means that a high value for this metric could mean
either a large number of small resource loads or a small number of very large
resources. A common case is when an app displays a single bitmap thats
close to the size of the screen. Another case is when an app displays a
large number of thumbnails.
</p>

<p>
To shrink this bar, you can employ techniques such as:
</p>

<ul>
   <li>
Ensuring your bitmap resolutions are not much larger than the size at which they
will be displayed. For example, your app should avoid displaying a 1024x1024
image as a 48x48 image.
   </li>

   <li>
Taking advantage of {@link android.graphics.Bitmap#prepareToDraw()}
to asynchronously pre-upload a bitmap before the next sync phase.
   </li>
</ul>

<h3 id="ic">Issuing Commands</h3>

<p>
The <em>Issue Commands</em> segment represents the time it takes to issue all
of the commands necessary for drawing display lists to the screen.
</p>

<p>
For the system to draw display lists to the screen, it sends the
necessary commands to the GPU. Typically, it performs this action through the
<a href="/guide/topics/graphics/opengl.html">OpenGL ES</a> API.
</p>

<p>
This process takes some time, as the system performs final transformation
and clipping for each command before sending the command to the GPU. Additional
overhead then arises on the GPU side, which computes the final commands. These
commands include final transformations, and additional clipping.
</p>

<h4>When this segment is large</h4>

<p>
The time spent in this stage is a direct measure of the complexity and
quantity of display lists that the system renders in a given
frame. For example, having many draw operations, especially in cases where
there's a small inherent cost to each draw primitive, could inflate this time.
For example:
</p>

<pre>
for (int i = 0; i < 1000; i++)
canvas.drawPoint()
</pre>

<p>
is a lot more expensive to issue than:
</p>

<pre>
canvas.drawPoints(mThousandPointArray);
</pre>

<p>
There isnt always a 1:1 correlation between issuing commands and
actually drawing display lists. Unlike <em>Issue Commands</em>,
which captures the time it takes to send drawing commands to the GPU,
the <em>Draw</em> metric represents the time that it took to capture the issued
commands into the display list.
</p>

<p>
This difference arises because the display lists are cached by
the system wherever possible. As a result, there are situations where a
scroll, transform, or animation requires the system to re-send a display
list, but not have to actually rebuild it&mdash;recapture the drawing
commands&mdash;from scratch. As a result, you can see a high Issue
commands bar without seeing a high <em>Draw commands</em> bar.
</p>

<h3 id="psb">Processing/Swapping Buffers</h3>

<p>
Once Android finishes submitting all its display list to the GPU,
the system issues one final command to tell the graphics driver that it's
done with the current frame. At this point, the driver can finally present
the updated image to the screen.
</p>

<h4>When this segment is large</h4>

<p>
Its important to understand that the GPU executes work in parallel with the
CPU. The Android system issues draw commands to the GPU, and then moves on to
the next task. The GPU reads those draw commands from a queue and processes
them.
</p>

<p>
In situations where the CPU issues commands faster than the GPU
consumes them, the communications queue between the processors can become
full. When this occurs, the CPU blocks, and waits until there is space in the
queue to place the next command. This full-queue state arises often during the
<em>Swap Buffers</em> stage, because at that point, a whole frames worth of
commands have been submitted.
</p>

</p>
The key to mitigating this problem is to reduce the complexity of work occurring
on the GPU, in similar fashion to what you would do for the Issue Commands
phase.
</p>


<h3 id="mt">Miscellaneous</h3>

<p>
In addition to the time it takes the rendering system to perform its work,
theres an additional set of work that occurs on the main thread and has
nothing to do with rendering. Time that this work consumes is reported as
<em>misc time</em>. Misc time generally represents work that might be occurring
on the UI thread between two consecutive frames of rendering.
</p>

<h4>When this segment is large</h4>

<p>
If this value is high, it is likely that your app has callbacks, intents, or
other work that should be happening on another thread. Tools such as
<a href="/studio/profile/traceview.html">Method
Tracing</a> or <a href="/studio/profile/systrace.html">Systrace</a> can provide
visibility into the tasks that are running on
the main thread. This information can help you target performance improvements.
</p>