audio HAL: Optimize for Spatial Audio
Add CAP_SYS_NICE.
Allow util_clamp_min for Spatializer Effect worker thread.
Reduce glitches caused by core migration.
Reduce power consumption as Spatializer Effect is DSP compute bound.
This change only affects Spatial Audio Effect processing.
Test: instrumented
Test: adb shell 'uclampset -a -p $(pgrep -of android.hardware.audio.service)'
Test: adb shell cat "/proc/$(adb shell pgrep -of android.hardware.audio.service)/status"
Test: adb shell 'ps -Tl -p $(pgrep -of android.hardware.audio.service)'
Bug: 181148259
Bug: 260918856
Bug: 261228892
Bug: 261686532
Bug: 262803152
Change-Id: I59ea99e8a1720f61954747f6c29e14cfea9a0f56
diff --git a/audio/common/all-versions/default/service/android.hardware.audio.service.rc b/audio/common/all-versions/default/service/android.hardware.audio.service.rc
index f859f21..0de4eea 100644
--- a/audio/common/all-versions/default/service/android.hardware.audio.service.rc
+++ b/audio/common/all-versions/default/service/android.hardware.audio.service.rc
@@ -3,7 +3,7 @@
user audioserver
# media gid needed for /dev/fm (radio) and for /data/misc/media (tee)
group audio camera drmrpc inet media mediadrm net_bt net_bt_admin net_bw_acct wakelock context_hub
- capabilities BLOCK_SUSPEND
+ capabilities BLOCK_SUSPEND SYS_NICE
# setting RLIMIT_RTPRIO allows binder RT priority inheritance
rlimit rtprio 10 10
ioprio rt 4
diff --git a/audio/effect/all-versions/default/Effect.cpp b/audio/effect/all-versions/default/Effect.cpp
index 5dc42dc..87e1ab7 100644
--- a/audio/effect/all-versions/default/Effect.cpp
+++ b/audio/effect/all-versions/default/Effect.cpp
@@ -25,8 +25,11 @@
#define ATRACE_TAG ATRACE_TAG_AUDIO
#include <HidlUtils.h>
#include <android/log.h>
+#include <cutils/properties.h>
#include <media/EffectsFactoryApi.h>
#include <mediautils/ScopedStatistics.h>
+#include <sys/syscall.h>
+#include <system/audio_effects/effect_spatializer.h>
#include <util/EffectUtils.h>
#include <utils/Trace.h>
@@ -47,6 +50,160 @@
namespace {
+/**
+ * Some basic scheduling tools.
+ */
+namespace scheduler {
+
+int getCpu() {
+ return sched_getcpu();
+}
+
+uint64_t getAffinity(pid_t tid) {
+ cpu_set_t set;
+ CPU_ZERO_S(sizeof(set), &set);
+
+ if (sched_getaffinity(tid, sizeof(set), &set)) {
+ ALOGW("%s: for tid:%d returning 0, failed %s", __func__, tid, strerror(errno));
+ return 0;
+ }
+ const int count = CPU_COUNT_S(sizeof(set), &set);
+ uint64_t mask = 0;
+ for (int i = 0; i < CPU_SETSIZE; ++i) {
+ if (CPU_ISSET_S(i, sizeof(set), &set)) {
+ mask |= 1 << i;
+ }
+ }
+ ALOGV("%s: for tid:%d returning cpu count %d mask %llu", __func__, tid, count,
+ (unsigned long long)mask);
+ return mask;
+}
+
+status_t setAffinity(pid_t tid, uint64_t mask) {
+ cpu_set_t set;
+ CPU_ZERO_S(sizeof(set), &set);
+
+ for (uint64_t m = mask; m != 0;) {
+ uint64_t tz = __builtin_ctz(m);
+ CPU_SET_S(tz, sizeof(set), &set);
+ m &= ~(1 << tz);
+ }
+ if (sched_setaffinity(tid, sizeof(set), &set)) {
+ ALOGW("%s: for tid:%d setting cpu mask %llu failed %s", __func__, tid,
+ (unsigned long long)mask, strerror(errno));
+ return -errno;
+ }
+ ALOGV("%s: for tid:%d setting cpu mask %llu", __func__, tid, (unsigned long long)mask);
+ return OK;
+}
+
+__unused status_t setPriority(pid_t tid, int policy, int priority) {
+ struct sched_param param {
+ .sched_priority = priority,
+ };
+ if (sched_setscheduler(tid, policy, ¶m) != 0) {
+ ALOGW("%s: Cannot set FIFO priority for tid %d to policy %d priority %d %s", __func__, tid,
+ policy, priority, strerror(errno));
+ return -errno;
+ }
+ ALOGV("%s: Successfully set priority for tid %d to policy %d priority %d", __func__, tid,
+ policy, priority);
+ return NO_ERROR;
+}
+
+status_t setUtilMin(pid_t tid, uint32_t utilMin) {
+ // Currently, there is no wrapper in bionic: b/183240349.
+ struct {
+ uint32_t size;
+ uint32_t sched_policy;
+ uint64_t sched_flags;
+ int32_t sched_nice;
+ uint32_t sched_priority;
+ uint64_t sched_runtime;
+ uint64_t sched_deadline;
+ uint64_t sched_period;
+ uint32_t sched_util_min;
+ uint32_t sched_util_max;
+ } attr{
+ .size = sizeof(attr),
+ .sched_flags = SCHED_FLAG_KEEP_ALL | SCHED_FLAG_UTIL_CLAMP_MIN,
+ .sched_util_min = utilMin,
+ };
+
+ if (syscall(__NR_sched_setattr, tid, &attr, 0 /* flags */)) {
+ ALOGW("%s: Cannot set sched_util_min for pid %d to %u %s", __func__, tid, utilMin,
+ strerror(errno));
+ return -errno;
+ }
+ ALOGV("%s: Successfully set sched_util_min for pid %d to %u", __func__, tid, utilMin);
+ return NO_ERROR;
+}
+
+/*
+ Attempts to raise the priority and usage of tid for spatialization.
+ Returns OK if everything works.
+*/
+status_t updateSpatializerPriority(pid_t tid) {
+ status_t status = OK;
+
+ const int cpu = getCpu();
+ ALOGV("%s: current CPU:%d", __func__, cpu);
+
+ const auto currentAffinity = getAffinity(tid);
+ ALOGV("%s: current Affinity:%llx", __func__, (unsigned long long)currentAffinity);
+
+ // Set the desired CPU core affinity.
+ // Typically this would be done to move the Spatializer effect off of the little cores.
+ // The mid cores and large cores typically have more FP/NEON units
+ // and will advantageously reduce power and prevent glitches due CPU limitations.
+ //
+ // Since this is SOC dependent, we do not set the core affinity here but
+ // prefer to set the util_clamp_min below.
+ //
+ constexpr uint64_t kDefaultAffinity = 0;
+ const int32_t desiredAffinity =
+ property_get_int32("audio.spatializer.effect.affinity", kDefaultAffinity);
+ if (desiredAffinity != 0 && (desiredAffinity & ~currentAffinity) == 0) {
+ const status_t localStatus = setAffinity(tid, desiredAffinity);
+ status = status ? status : localStatus;
+ }
+
+ // Set the util_clamp_min.
+ // This is beneficial to reduce glitches when starting up, or due to scheduler
+ // thread statistics reset (e.g. core migration), which cause the CPU frequency to drop
+ // to minimum.
+ //
+ // Experimentation has found that moving to a mid core over a little core reduces
+ // power if the mid core (e.g. A76/78) has more (e.g. 2x) FP/NEON units
+ // than the little core (e.g. A55).
+ // A possible value is 300.
+ //
+ constexpr uint32_t kUtilMin = 0;
+ const int32_t utilMin = property_get_int32("audio.spatializer.effect.util_clamp_min", kUtilMin);
+ if (utilMin > 0 && utilMin <= 1024) {
+ const status_t localStatus = setUtilMin(tid, utilMin);
+ status = status ? status : localStatus;
+ }
+
+#if 0
+ // Provided for local vendor testing but not enabled as audioserver does this for us.
+ //
+ // Set priority if specified.
+ constexpr int32_t kRTPriorityMin = 1;
+ constexpr int32_t kRTPriorityMax = 3;
+ const int32_t priorityBoost =
+ property_get_int32("audio.spatializer.priority", kRTPriorityMin);
+ if (priorityBoost >= kRTPriorityMin && priorityBoost <= kRTPriorityMax) {
+ const status_t localStatus = scheduler::setPriority(threadId, SCHED_FIFO, priorityBoost);
+ status = status ? status : localStatus;
+ }
+#endif
+
+ return status;
+}
+
+} // namespace scheduler
+
#define SCOPED_STATS() \
::android::mediautils::ScopedStatistics scopedStatistics { \
std::string("EffectHal::").append(__func__), mEffectHal->mStatistics \
@@ -83,6 +240,16 @@
};
bool ProcessThread::threadLoop() {
+ // For a spatializer effect, we perform scheduler adjustments to reduce glitches and power.
+ {
+ effect_descriptor_t halDescriptor{};
+ if ((*mEffect)->get_descriptor(mEffect, &halDescriptor) == NO_ERROR &&
+ memcmp(&halDescriptor.type, FX_IID_SPATIALIZER, sizeof(effect_uuid_t)) == 0) {
+ const status_t status = scheduler::updateSpatializerPriority(gettid());
+ ALOGW_IF(status != OK, "Failed to update Spatializer priority");
+ }
+ }
+
// This implementation doesn't return control back to the Thread until it decides to stop,
// as the Thread uses mutexes, and this can lead to priority inversion.
while (!std::atomic_load_explicit(mStop, std::memory_order_acquire)) {