1 page.title=Audio Latency 2 @jd:body 3 4 <!-- 5 Copyright 2013 The Android Open Source Project 6 7 Licensed under the Apache License, Version 2.0 (the "License"); 8 you may not use this file except in compliance with the License. 9 You may obtain a copy of the License at 10 11 http://www.apache.org/licenses/LICENSE-2.0 12 13 Unless required by applicable law or agreed to in writing, software 14 distributed under the License is distributed on an "AS IS" BASIS, 15 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 16 See the License for the specific language governing permissions and 17 limitations under the License. 18 --> 19 <div id="qv-wrapper"> 20 <div id="qv"> 21 <h2>In this document</h2> 22 <ol id="auto-toc"> 23 </ol> 24 </div> 25 </div> 26 27 <p>Audio latency is the time delay as an audio signal passes through a system. 28 For a complete description of audio latency for the purposes of Android 29 compatibility, see <em>Section 5.5 Audio Latency</em> 30 in the <a href="http://source.android.com/compatibility/index.html">Android CDD</a>. 31 See <a href="latency_design.html">Design For Reduced Latency</a> for an 32 understanding of Android's audio latency-reduction efforts. 33 </p> 34 35 <p> 36 This page focuses on the contributors to output latency, 37 but a similar discussion applies to input latency. 38 </p> 39 <p> 40 Assuming the analog circuitry does not contribute significantly, then the major 41 surface-level contributors to audio latency are the following: 42 </p> 43 44 <ul> 45 <li>Application</li> 46 <li>Total number of buffers in pipeline</li> 47 <li>Size of each buffer, in frames</li> 48 <li>Additional latency after the app processor, such as from a DSP</li> 49 </ul> 50 51 <p> 52 As accurate as the above list of contributors may be, it is also misleading. 53 The reason is that buffer count and buffer size are more of an 54 <em>effect</em> than a <em>cause</em>. What usually happens is that 55 a given buffer scheme is implemented and tested, but during testing, an audio 56 underrun is heard as a "click" or "pop." To compensate, the 57 system designer then increases buffer sizes or buffer counts. 58 This has the desired result of eliminating the underruns, but it also 59 has the undesired side effect of increasing latency. 60 </p> 61 62 <p> 63 A better approach is to understand the causes of the 64 underruns and then correct those. This eliminates the 65 audible artifacts and may even permit even smaller or fewer buffers 66 and thus reduce latency. 67 </p> 68 69 <p> 70 In our experience, the most common causes of underruns include: 71 </p> 72 <ul> 73 <li>Linux CFS (Completely Fair Scheduler)</li> 74 <li>high-priority threads with SCHED_FIFO scheduling</li> 75 <li>long scheduling latency</li> 76 <li>long-running interrupt handlers</li> 77 <li>long interrupt disable time</li> 78 </ul> 79 80 <h3>Linux CFS and SCHED_FIFO scheduling</h3> 81 <p> 82 The Linux CFS is designed to be fair to competing workloads sharing a common CPU 83 resource. This fairness is represented by a per-thread <em>nice</em> parameter. 84 The nice value ranges from -19 (least nice, or most CPU time allocated) 85 to 20 (nicest, or least CPU time allocated). In general, all threads with a given 86 nice value receive approximately equal CPU time and threads with a 87 numerically lower nice value should expect to 88 receive more CPU time. However, CFS is "fair" only over relatively long 89 periods of observation. Over short-term observation windows, 90 CFS may allocate the CPU resource in unexpected ways. For example, it 91 may take the CPU away from a thread with numerically low niceness 92 onto a thread with a numerically high niceness. In the case of audio, 93 this can result in an underrun. 94 </p> 95 96 <p> 97 The obvious solution is to avoid CFS for high-performance audio 98 threads. Beginning with Android 4.1, such threads now use the 99 <code>SCHED_FIFO</code> scheduling policy rather than the <code>SCHED_NORMAL</code> (also called 100 <code>SCHED_OTHER</code>) scheduling policy implemented by CFS. 101 </p> 102 103 <p> 104 Though the high-performance audio threads now use <code>SCHED_FIFO</code>, they 105 are still susceptible to other higher priority <code>SCHED_FIFO</code> threads. 106 These are typically kernel worker threads, but there may also be a few 107 non-audio user threads with policy <code>SCHED_FIFO</code>. The available <code>SCHED_FIFO</code> 108 priorities range from 1 to 99. The audio threads run at priority 109 2 or 3. This leaves priority 1 available for lower priority threads, 110 and priorities 4 to 99 for higher priority threads. We recommend 111 you use priority 1 whenever possible, and reserve priorities 4 to 99 for 112 those threads that are guaranteed to complete within a bounded amount 113 of time and are known to not interfere with scheduling of audio threads. 114 </p> 115 116 <h3>Scheduling latency</h3> 117 <p> 118 Scheduling latency is the time between when a thread becomes 119 ready to run, and when the resulting context switch completes so that the 120 thread actually runs on a CPU. The shorter the latency the better, and 121 anything over two milliseconds causes problems for audio. Long scheduling 122 latency is most likely to occur during mode transitions, such as 123 bringing up or shutting down a CPU, switching between a security kernel 124 and the normal kernel, switching from full power to low-power mode, 125 or adjusting the CPU clock frequency and voltage. 126 </p> 127 128 <h3>Interrupts</h3> 129 <p> 130 In many designs, CPU 0 services all external interrupts. So a 131 long-running interrupt handler may delay other interrupts, in particular 132 audio direct memory access (DMA) completion interrupts. Design interrupt handlers 133 to finish quickly and defer any lengthy work to a thread (preferably 134 a CFS thread or <code>SCHED_FIFO</code> thread of priority 1). 135 </p> 136 137 <p> 138 Equivalently, disabling interrupts on CPU 0 for a long period 139 has the same result of delaying the servicing of audio interrupts. 140 Long interrupt disable times typically happen while waiting for a kernel 141 <i>spin lock</i>. Review these spin locks to ensure that 142 they are bounded. 143 </p> 144 145