diff --git a/libs/gui/SurfaceComposerClient.cpp b/libs/gui/SurfaceComposerClient.cpp
index 47d801a..9358e29 100644
--- a/libs/gui/SurfaceComposerClient.cpp
+++ b/libs/gui/SurfaceComposerClient.cpp
@@ -897,6 +897,10 @@
     mApplyToken = nullptr;
 }
 
+uint64_t SurfaceComposerClient::Transaction::getId() {
+    return mId;
+}
+
 void SurfaceComposerClient::doUncacheBufferTransaction(uint64_t cacheId) {
     sp<ISurfaceComposer> sf(ComposerService::getComposerService());
 
diff --git a/libs/gui/include/gui/SurfaceComposerClient.h b/libs/gui/include/gui/SurfaceComposerClient.h
index efbdb36..b598b43 100644
--- a/libs/gui/include/gui/SurfaceComposerClient.h
+++ b/libs/gui/include/gui/SurfaceComposerClient.h
@@ -461,6 +461,10 @@
         // Clears the contents of the transaction without applying it.
         void clear();
 
+        // Returns the current id of the transaction.
+        // The id is updated every time the transaction is applied.
+        uint64_t getId();
+
         status_t apply(bool synchronous = false, bool oneWay = false);
         // Merge another transaction in to this one, clearing other
         // as if it had been applied.
diff --git a/services/surfaceflinger/CompositionEngine/include/compositionengine/Output.h b/services/surfaceflinger/CompositionEngine/include/compositionengine/Output.h
index db2fd1b..2203639 100644
--- a/services/surfaceflinger/CompositionEngine/include/compositionengine/Output.h
+++ b/services/surfaceflinger/CompositionEngine/include/compositionengine/Output.h
@@ -27,6 +27,7 @@
 #include <compositionengine/LayerFE.h>
 #include <renderengine/LayerSettings.h>
 #include <ui/Fence.h>
+#include <ui/FenceTime.h>
 #include <ui/GraphicTypes.h>
 #include <ui/LayerStack.h>
 #include <ui/Region.h>
@@ -311,6 +312,8 @@
             const Region& flashRegion,
             std::vector<LayerFE::LayerSettings>& clientCompositionLayers) = 0;
     virtual void setExpensiveRenderingExpected(bool enabled) = 0;
+    virtual void setHintSessionGpuFence(std::unique_ptr<FenceTime>&& gpuFence) = 0;
+    virtual bool isPowerHintSessionEnabled() = 0;
     virtual void cacheClientCompositionRequests(uint32_t cacheSize) = 0;
     virtual bool canPredictCompositionStrategy(const CompositionRefreshArgs&) = 0;
 };
diff --git a/services/surfaceflinger/CompositionEngine/include/compositionengine/impl/Display.h b/services/surfaceflinger/CompositionEngine/include/compositionengine/impl/Display.h
index 61a0e6a..fa7bc5d 100644
--- a/services/surfaceflinger/CompositionEngine/include/compositionengine/impl/Display.h
+++ b/services/surfaceflinger/CompositionEngine/include/compositionengine/impl/Display.h
@@ -89,6 +89,8 @@
     std::unique_ptr<compositionengine::OutputLayer> createOutputLayer(const sp<LayerFE>&) const;
 
 private:
+    bool isPowerHintSessionEnabled() override;
+    void setHintSessionGpuFence(std::unique_ptr<FenceTime>&& gpuFence) override;
     DisplayId mId;
     bool mIsDisconnected = false;
     Hwc2::PowerAdvisor* mPowerAdvisor = nullptr;
diff --git a/services/surfaceflinger/CompositionEngine/include/compositionengine/impl/Output.h b/services/surfaceflinger/CompositionEngine/include/compositionengine/impl/Output.h
index 31c51e6..df721cd 100644
--- a/services/surfaceflinger/CompositionEngine/include/compositionengine/impl/Output.h
+++ b/services/surfaceflinger/CompositionEngine/include/compositionengine/impl/Output.h
@@ -140,6 +140,8 @@
           std::vector<LayerFE*> &outLayerFEs) override;
     void appendRegionFlashRequests(const Region&, std::vector<LayerFE::LayerSettings>&) override;
     void setExpensiveRenderingExpected(bool enabled) override;
+    void setHintSessionGpuFence(std::unique_ptr<FenceTime>&& gpuFence) override;
+    bool isPowerHintSessionEnabled() override;
     void dumpBase(std::string&) const;
 
     // Implemented by the final implementation for the final state it uses.
diff --git a/services/surfaceflinger/CompositionEngine/include/compositionengine/mock/Output.h b/services/surfaceflinger/CompositionEngine/include/compositionengine/mock/Output.h
index cb9fbad..2a04949 100644
--- a/services/surfaceflinger/CompositionEngine/include/compositionengine/mock/Output.h
+++ b/services/surfaceflinger/CompositionEngine/include/compositionengine/mock/Output.h
@@ -133,6 +133,8 @@
     MOCK_METHOD1(canPredictCompositionStrategy, bool(const CompositionRefreshArgs&));
     MOCK_METHOD1(setPredictCompositionStrategy, void(bool));
     MOCK_METHOD1(setTreat170mAsSrgb, void(bool));
+    MOCK_METHOD(void, setHintSessionGpuFence, (std::unique_ptr<FenceTime> && gpuFence));
+    MOCK_METHOD(bool, isPowerHintSessionEnabled, ());
 };
 
 } // namespace android::compositionengine::mock
diff --git a/services/surfaceflinger/CompositionEngine/src/Display.cpp b/services/surfaceflinger/CompositionEngine/src/Display.cpp
index b79b46b..cccf907 100644
--- a/services/surfaceflinger/CompositionEngine/src/Display.cpp
+++ b/services/surfaceflinger/CompositionEngine/src/Display.cpp
@@ -243,11 +243,14 @@
         return false;
     }
 
+    const nsecs_t startTime = systemTime();
+
     // Get any composition changes requested by the HWC device, and apply them.
     std::optional<android::HWComposer::DeviceRequestedChanges> changes;
     auto& hwc = getCompositionEngine().getHwComposer();
+    const bool requiresClientComposition = anyLayersRequireClientComposition();
     if (status_t result =
-                hwc.getDeviceCompositionChanges(*halDisplayId, anyLayersRequireClientComposition(),
+                hwc.getDeviceCompositionChanges(*halDisplayId, requiresClientComposition,
                                                 getState().earliestPresentTime,
                                                 getState().previousPresentFence,
                                                 getState().expectedPresentTime, outChanges);
@@ -257,6 +260,11 @@
         return false;
     }
 
+    if (isPowerHintSessionEnabled()) {
+        mPowerAdvisor->setValidateTiming(mId, startTime, systemTime());
+        mPowerAdvisor->setRequiresClientComposition(mId, requiresClientComposition);
+    }
+
     return true;
 }
 
@@ -356,9 +364,24 @@
     }
 
     auto& hwc = getCompositionEngine().getHwComposer();
+
+    const nsecs_t startTime = systemTime();
+
+    if (isPowerHintSessionEnabled()) {
+        if (!getCompositionEngine().getHwComposer().getComposer()->isSupported(
+                    Hwc2::Composer::OptionalFeature::ExpectedPresentTime) &&
+            getState().previousPresentFence->getSignalTime() != Fence::SIGNAL_TIME_PENDING) {
+            mPowerAdvisor->setPresentDelayedTime(mId, getState().earliestPresentTime);
+        }
+    }
+
     hwc.presentAndGetReleaseFences(*halDisplayIdOpt, getState().earliestPresentTime,
                                    getState().previousPresentFence);
 
+    if (isPowerHintSessionEnabled()) {
+        mPowerAdvisor->setPresentTiming(mId, startTime, systemTime());
+    }
+
     fences.presentFence = hwc.getPresentFence(*halDisplayIdOpt);
 
     // TODO(b/121291683): Change HWComposer call to return entire map
@@ -384,6 +407,14 @@
     }
 }
 
+bool Display::isPowerHintSessionEnabled() {
+    return mPowerAdvisor != nullptr && mPowerAdvisor->usePowerHintSession();
+}
+
+void Display::setHintSessionGpuFence(std::unique_ptr<FenceTime>&& gpuFence) {
+    mPowerAdvisor->setGpuFenceTime(mId, std::move(gpuFence));
+}
+
 void Display::finishFrame(const compositionengine::CompositionRefreshArgs& refreshArgs,
                           GpuCompositionResult&& result) {
     // We only need to actually compose the display if:
@@ -396,6 +427,13 @@
     }
 
     impl::Output::finishFrame(refreshArgs, std::move(result));
+
+    if (isPowerHintSessionEnabled()) {
+        auto& hwc = getCompositionEngine().getHwComposer();
+        if (auto halDisplayId = HalDisplayId::tryCast(mId)) {
+            mPowerAdvisor->setSkippedValidate(mId, hwc.getValidateSkipped(*halDisplayId));
+        }
+    }
 }
 
 } // namespace android::compositionengine::impl
diff --git a/services/surfaceflinger/CompositionEngine/src/Output.cpp b/services/surfaceflinger/CompositionEngine/src/Output.cpp
index c3385a8..070b285 100644
--- a/services/surfaceflinger/CompositionEngine/src/Output.cpp
+++ b/services/surfaceflinger/CompositionEngine/src/Output.cpp
@@ -1099,6 +1099,10 @@
         return;
     }
 
+    if (isPowerHintSessionEnabled()) {
+        // get fence end time to know when gpu is complete in display
+        setHintSessionGpuFence(std::make_unique<FenceTime>(new Fence(dup(optReadyFence->get()))));
+    }
     // swap buffers (presentation)
     mRenderSurface->queueBuffer(std::move(*optReadyFence));
 }
@@ -1403,6 +1407,14 @@
     // The base class does nothing with this call.
 }
 
+void Output::setHintSessionGpuFence(std::unique_ptr<FenceTime>&&) {
+    // The base class does nothing with this call.
+}
+
+bool Output::isPowerHintSessionEnabled() {
+    return false;
+}
+
 void Output::postFramebuffer() {
     ATRACE_CALL();
     ALOGV(__FUNCTION__);
diff --git a/services/surfaceflinger/CompositionEngine/tests/DisplayTest.cpp b/services/surfaceflinger/CompositionEngine/tests/DisplayTest.cpp
index 0e5a7b6..344fea3 100644
--- a/services/surfaceflinger/CompositionEngine/tests/DisplayTest.cpp
+++ b/services/surfaceflinger/CompositionEngine/tests/DisplayTest.cpp
@@ -169,6 +169,7 @@
         EXPECT_CALL(mCompositionEngine, getRenderEngine()).WillRepeatedly(ReturnRef(mRenderEngine));
         EXPECT_CALL(mRenderEngine, supportsProtectedContent()).WillRepeatedly(Return(false));
         EXPECT_CALL(mRenderEngine, isProtected()).WillRepeatedly(Return(false));
+        EXPECT_CALL(mPowerAdvisor, usePowerHintSession()).WillRepeatedly(Return(false));
     }
 
     DisplayCreationArgs getDisplayCreationArgsForPhysicalDisplay() {
diff --git a/services/surfaceflinger/CompositionEngine/tests/MockHWComposer.h b/services/surfaceflinger/CompositionEngine/tests/MockHWComposer.h
index 9b12b08..d7704a8 100644
--- a/services/surfaceflinger/CompositionEngine/tests/MockHWComposer.h
+++ b/services/surfaceflinger/CompositionEngine/tests/MockHWComposer.h
@@ -137,6 +137,7 @@
     MOCK_METHOD(bool, hasDisplayIdleTimerCapability, (PhysicalDisplayId), (const, override));
     MOCK_METHOD(Hwc2::AidlTransform, getPhysicalDisplayOrientation, (PhysicalDisplayId),
                 (const, override));
+    MOCK_METHOD(bool, getValidateSkipped, (HalDisplayId), (const, override));
 };
 
 } // namespace mock
diff --git a/services/surfaceflinger/CompositionEngine/tests/MockPowerAdvisor.h b/services/surfaceflinger/CompositionEngine/tests/MockPowerAdvisor.h
index 50adcfb..c8b6d44 100644
--- a/services/surfaceflinger/CompositionEngine/tests/MockPowerAdvisor.h
+++ b/services/surfaceflinger/CompositionEngine/tests/MockPowerAdvisor.h
@@ -38,11 +38,31 @@
     MOCK_METHOD(bool, usePowerHintSession, (), (override));
     MOCK_METHOD(bool, supportsPowerHintSession, (), (override));
     MOCK_METHOD(bool, isPowerHintSessionRunning, (), (override));
-    MOCK_METHOD(void, setTargetWorkDuration, (int64_t targetDurationNanos), (override));
-    MOCK_METHOD(void, sendActualWorkDuration, (int64_t actualDurationNanos, nsecs_t timestamp),
-                (override));
+    MOCK_METHOD(void, setTargetWorkDuration, (int64_t targetDuration), (override));
+    MOCK_METHOD(void, sendActualWorkDuration, (), (override));
+    MOCK_METHOD(void, sendPredictedWorkDuration, (), (override));
     MOCK_METHOD(void, enablePowerHint, (bool enabled), (override));
     MOCK_METHOD(bool, startPowerHintSession, (const std::vector<int32_t>& threadIds), (override));
+    MOCK_METHOD(void, setGpuFenceTime,
+                (DisplayId displayId, std::unique_ptr<FenceTime>&& fenceTime), (override));
+    MOCK_METHOD(void, setValidateTiming,
+                (DisplayId displayId, nsecs_t valiateStartTime, nsecs_t validateEndTime),
+                (override));
+    MOCK_METHOD(void, setPresentTiming,
+                (DisplayId displayId, nsecs_t presentStartTime, nsecs_t presentEndTime),
+                (override));
+    MOCK_METHOD(void, setSkippedValidate, (DisplayId displayId, bool skipped), (override));
+    MOCK_METHOD(void, setRequiresClientComposition,
+                (DisplayId displayId, bool requiresClientComposition), (override));
+    MOCK_METHOD(void, setExpectedPresentTime, (nsecs_t expectedPresentTime), (override));
+    MOCK_METHOD(void, setPresentDelayedTime,
+                (DisplayId displayId,
+                 std::chrono::steady_clock::time_point earliestFrameStartTime));
+    MOCK_METHOD(void, setFrameDelay, (nsecs_t frameDelayDuration), (override));
+    MOCK_METHOD(void, setCommitStart, (nsecs_t commitStartTime), (override));
+    MOCK_METHOD(void, setCompositeEnd, (nsecs_t compositeEndtime), (override));
+    MOCK_METHOD(void, setDisplays, (std::vector<DisplayId> & displayIds), (override));
+    MOCK_METHOD(void, setTotalFrameTargetWorkDuration, (int64_t targetDuration), (override));
 };
 
 } // namespace mock
diff --git a/services/surfaceflinger/CompositionEngine/tests/OutputTest.cpp b/services/surfaceflinger/CompositionEngine/tests/OutputTest.cpp
index 063726b..505f94e 100644
--- a/services/surfaceflinger/CompositionEngine/tests/OutputTest.cpp
+++ b/services/surfaceflinger/CompositionEngine/tests/OutputTest.cpp
@@ -3319,6 +3319,9 @@
         MOCK_METHOD2(appendRegionFlashRequests,
                      void(const Region&, std::vector<LayerFE::LayerSettings>&));
         MOCK_METHOD1(setExpensiveRenderingExpected, void(bool));
+        MOCK_METHOD(void, setHintSessionGpuFence, (std::unique_ptr<FenceTime> && gpuFence),
+                    (override));
+        MOCK_METHOD(bool, isPowerHintSessionEnabled, (), (override));
     };
 
     OutputComposeSurfacesTest() {
diff --git a/services/surfaceflinger/DisplayDevice.cpp b/services/surfaceflinger/DisplayDevice.cpp
index a915b61..8680900 100644
--- a/services/surfaceflinger/DisplayDevice.cpp
+++ b/services/surfaceflinger/DisplayDevice.cpp
@@ -211,6 +211,7 @@
               to_string(getId()).c_str());
         return BAD_VALUE;
     }
+    mNumModeSwitchesInPolicy++;
     mUpcomingActiveMode = info;
     ATRACE_INT(mActiveModeFPSHwcTrace.c_str(), info.mode->getFps().getIntValue());
     return mHwComposer.setActiveModeWithConstraints(getPhysicalId(), info.mode->getHwcId(),
@@ -537,6 +538,27 @@
     mDesiredActiveModeChanged = false;
 }
 
+status_t DisplayDevice::setRefreshRatePolicy(
+        const std::optional<scheduler::RefreshRateConfigs::Policy>& policy, bool overridePolicy) {
+    const auto oldPolicy = mRefreshRateConfigs->getCurrentPolicy();
+    const status_t setPolicyResult = overridePolicy
+            ? mRefreshRateConfigs->setOverridePolicy(policy)
+            : mRefreshRateConfigs->setDisplayManagerPolicy(*policy);
+
+    if (setPolicyResult == OK) {
+        const int numModeChanges = mNumModeSwitchesInPolicy.exchange(0);
+
+        ALOGI("Display %s policy changed\n"
+              "Previous: {%s}\n"
+              "Current:  {%s}\n"
+              "%d mode changes were performed under the previous policy",
+              to_string(getId()).c_str(), oldPolicy.toString().c_str(),
+              policy ? policy->toString().c_str() : "null", numModeChanges);
+    }
+
+    return setPolicyResult;
+}
+
 std::atomic<int32_t> DisplayDeviceState::sNextSequenceId(1);
 
 }  // namespace android
diff --git a/services/surfaceflinger/DisplayDevice.h b/services/surfaceflinger/DisplayDevice.h
index d5d87b4..2161436 100644
--- a/services/surfaceflinger/DisplayDevice.h
+++ b/services/surfaceflinger/DisplayDevice.h
@@ -249,6 +249,10 @@
     nsecs_t getVsyncPeriodFromHWC() const;
     nsecs_t getRefreshTimestamp() const;
 
+    status_t setRefreshRatePolicy(
+            const std::optional<scheduler::RefreshRateConfigs::Policy>& policy,
+            bool overridePolicy);
+
     // release HWC resources (if any) for removable displays
     void disconnect();
 
@@ -303,6 +307,8 @@
     TracedOrdinal<bool> mDesiredActiveModeChanged
             GUARDED_BY(mActiveModeLock) = {"DesiredActiveModeChanged", false};
     ActiveModeInfo mUpcomingActiveMode GUARDED_BY(kMainThreadContext);
+
+    std::atomic_int mNumModeSwitchesInPolicy = 0;
 };
 
 struct DisplayDeviceState {
diff --git a/services/surfaceflinger/DisplayHardware/HWComposer.cpp b/services/surfaceflinger/DisplayHardware/HWComposer.cpp
index 0da8ece..a6aee1f 100644
--- a/services/surfaceflinger/DisplayHardware/HWComposer.cpp
+++ b/services/surfaceflinger/DisplayHardware/HWComposer.cpp
@@ -738,6 +738,13 @@
             });
 }
 
+bool HWComposer::getValidateSkipped(HalDisplayId displayId) const {
+    if (mDisplayData.count(displayId) == 0) {
+        return false;
+    }
+    return mDisplayData.at(displayId).validateWasSkipped;
+}
+
 status_t HWComposer::setBootDisplayMode(PhysicalDisplayId displayId,
                                         hal::HWConfigId displayModeId) {
     RETURN_IF_INVALID_DISPLAY(displayId, BAD_INDEX);
diff --git a/services/surfaceflinger/DisplayHardware/HWComposer.h b/services/surfaceflinger/DisplayHardware/HWComposer.h
index 4c0ecd8..92a8f30 100644
--- a/services/surfaceflinger/DisplayHardware/HWComposer.h
+++ b/services/surfaceflinger/DisplayHardware/HWComposer.h
@@ -199,6 +199,9 @@
             PhysicalDisplayId, float brightness, float brightnessNits,
             const Hwc2::Composer::DisplayBrightnessOptions&) = 0;
 
+    // Get whether the display skipped validation on the latest present
+    virtual bool getValidateSkipped(HalDisplayId displayId) const = 0;
+
     // Events handling ---------------------------------------------------------
 
     // Returns stable display ID (and display name on connection of new or previously disconnected
@@ -397,6 +400,8 @@
 
     status_t setActiveColorMode(PhysicalDisplayId, ui::ColorMode, ui::RenderIntent) override;
 
+    bool getValidateSkipped(HalDisplayId displayId) const override;
+
     // Composer 2.4
     ui::DisplayConnectionType getDisplayConnectionType(PhysicalDisplayId) const override;
     bool isVsyncPeriodSwitchSupported(PhysicalDisplayId) const override;
diff --git a/services/surfaceflinger/DisplayHardware/PowerAdvisor.cpp b/services/surfaceflinger/DisplayHardware/PowerAdvisor.cpp
index b5678b4..4c0a942 100644
--- a/services/surfaceflinger/DisplayHardware/PowerAdvisor.cpp
+++ b/services/surfaceflinger/DisplayHardware/PowerAdvisor.cpp
@@ -196,7 +196,7 @@
     return mPowerHintSessionRunning;
 }
 
-void PowerAdvisor::setTargetWorkDuration(int64_t targetDurationNanos) {
+void PowerAdvisor::setTargetWorkDuration(int64_t targetDuration) {
     if (!usePowerHintSession()) {
         ALOGV("Power hint session target duration cannot be set, skipping");
         return;
@@ -205,26 +205,44 @@
         std::lock_guard lock(mPowerHalMutex);
         HalWrapper* const halWrapper = getPowerHal();
         if (halWrapper != nullptr) {
-            halWrapper->setTargetWorkDuration(targetDurationNanos - kTargetSafetyMargin.count());
+            halWrapper->setTargetWorkDuration(targetDuration);
         }
     }
 }
 
-void PowerAdvisor::sendActualWorkDuration(int64_t actualDurationNanos, nsecs_t timeStampNanos) {
+void PowerAdvisor::sendActualWorkDuration() {
     if (!mBootFinished || !usePowerHintSession()) {
         ALOGV("Actual work duration power hint cannot be sent, skipping");
         return;
     }
-    {
+    const std::optional<nsecs_t> actualDuration = estimateWorkDuration(false);
+    if (actualDuration.has_value()) {
         std::lock_guard lock(mPowerHalMutex);
         HalWrapper* const halWrapper = getPowerHal();
         if (halWrapper != nullptr) {
-            halWrapper->sendActualWorkDuration(actualDurationNanos, timeStampNanos);
+            halWrapper->sendActualWorkDuration(*actualDuration + kTargetSafetyMargin.count(),
+                                               systemTime());
         }
     }
 }
 
-// needs to be set after the flag is known but before PowerAdvisor enters onBootFinished
+void PowerAdvisor::sendPredictedWorkDuration() {
+    if (!mBootFinished || !usePowerHintSession()) {
+        ALOGV("Actual work duration power hint cannot be sent, skipping");
+        return;
+    }
+
+    const std::optional<nsecs_t> predictedDuration = estimateWorkDuration(true);
+
+    if (predictedDuration.has_value()) {
+        std::lock_guard lock(mPowerHalMutex);
+        HalWrapper* const halWrapper = getPowerHal();
+        if (halWrapper != nullptr) {
+            halWrapper->sendActualWorkDuration(*predictedDuration, systemTime());
+        }
+    }
+}
+
 void PowerAdvisor::enablePowerHint(bool enabled) {
     mPowerHintEnabled = enabled;
 }
@@ -244,6 +262,282 @@
     return mPowerHintSessionRunning;
 }
 
+void PowerAdvisor::setGpuFenceTime(DisplayId displayId, std::unique_ptr<FenceTime>&& fenceTime) {
+    DisplayTimingData& displayData = mDisplayTimingData[displayId];
+    if (displayData.gpuEndFenceTime) {
+        nsecs_t signalTime = displayData.gpuEndFenceTime->getSignalTime();
+        if (signalTime != Fence::SIGNAL_TIME_INVALID && signalTime != Fence::SIGNAL_TIME_PENDING) {
+            for (auto&& [_, otherDisplayData] : mDisplayTimingData) {
+                // If the previous display started before us but ended after we should have
+                // started, then it likely delayed our start time and we must compensate for that.
+                // Displays finishing earlier should have already made their way through this call
+                // and swapped their timing into "lastValid" from "latest", so we check that here.
+                if (!otherDisplayData.lastValidGpuStartTime.has_value()) continue;
+                if ((*otherDisplayData.lastValidGpuStartTime < *displayData.gpuStartTime) &&
+                    (*otherDisplayData.lastValidGpuEndTime > *displayData.gpuStartTime)) {
+                    displayData.lastValidGpuStartTime = *otherDisplayData.lastValidGpuEndTime;
+                    break;
+                }
+            }
+            displayData.lastValidGpuStartTime = displayData.gpuStartTime;
+            displayData.lastValidGpuEndTime = signalTime;
+        }
+    }
+    displayData.gpuEndFenceTime = std::move(fenceTime);
+    displayData.gpuStartTime = systemTime();
+}
+
+void PowerAdvisor::setValidateTiming(DisplayId displayId, nsecs_t validateStartTime,
+                                     nsecs_t validateEndTime) {
+    DisplayTimingData& displayData = mDisplayTimingData[displayId];
+    displayData.validateStartTime = validateStartTime;
+    displayData.validateEndTime = validateEndTime;
+}
+
+void PowerAdvisor::setPresentTiming(DisplayId displayId, nsecs_t presentStartTime,
+                                    nsecs_t presentEndTime) {
+    DisplayTimingData& displayData = mDisplayTimingData[displayId];
+    displayData.presentStartTime = presentStartTime;
+    displayData.presentEndTime = presentEndTime;
+}
+
+void PowerAdvisor::setSkippedValidate(DisplayId displayId, bool skipped) {
+    mDisplayTimingData[displayId].skippedValidate = skipped;
+}
+
+void PowerAdvisor::setRequiresClientComposition(DisplayId displayId,
+                                                bool requiresClientComposition) {
+    mDisplayTimingData[displayId].usedClientComposition = requiresClientComposition;
+}
+
+void PowerAdvisor::setExpectedPresentTime(nsecs_t expectedPresentTime) {
+    mExpectedPresentTimes.append(expectedPresentTime);
+}
+
+void PowerAdvisor::setFrameDelay(nsecs_t frameDelayDuration) {
+    mFrameDelayDuration = frameDelayDuration;
+}
+
+void PowerAdvisor::setPresentDelayedTime(
+        DisplayId displayId, std::chrono::steady_clock::time_point earliestFrameStartTime) {
+    mDisplayTimingData[displayId].presentDelayedTime =
+            (earliestFrameStartTime - std::chrono::steady_clock::now()).count() + systemTime();
+}
+
+void PowerAdvisor::setCommitStart(nsecs_t commitStartTime) {
+    mCommitStartTimes.append(commitStartTime);
+}
+
+void PowerAdvisor::setCompositeEnd(nsecs_t compositeEnd) {
+    mLastCompositeEndTime = compositeEnd;
+    // calculate the postcomp time here as well
+    std::vector<DisplayId>&& displays = getOrderedDisplayIds(&DisplayTimingData::presentEndTime);
+    DisplayTimingData& timingData = mDisplayTimingData[displays.back()];
+    mLastPostcompDuration = compositeEnd -
+            (timingData.skippedValidate ? *timingData.validateEndTime : *timingData.presentEndTime);
+}
+
+void PowerAdvisor::setDisplays(std::vector<DisplayId>& displayIds) {
+    mDisplayIds = displayIds;
+}
+
+void PowerAdvisor::setTotalFrameTargetWorkDuration(nsecs_t targetDuration) {
+    mTotalFrameTargetDuration = targetDuration;
+}
+
+std::vector<DisplayId> PowerAdvisor::getOrderedDisplayIds(
+        std::optional<nsecs_t> DisplayTimingData::*sortBy) {
+    std::vector<DisplayId> sortedDisplays;
+    std::copy_if(mDisplayIds.begin(), mDisplayIds.end(), std::back_inserter(sortedDisplays),
+                 [&](DisplayId id) {
+                     return mDisplayTimingData.count(id) &&
+                             (mDisplayTimingData[id].*sortBy).has_value();
+                 });
+    std::sort(sortedDisplays.begin(), sortedDisplays.end(), [&](DisplayId idA, DisplayId idB) {
+        return *(mDisplayTimingData[idA].*sortBy) < *(mDisplayTimingData[idB].*sortBy);
+    });
+    return sortedDisplays;
+}
+
+std::optional<nsecs_t> PowerAdvisor::estimateWorkDuration(bool earlyHint) {
+    if (earlyHint && (!mExpectedPresentTimes.isFull() || !mCommitStartTimes.isFull())) {
+        return std::nullopt;
+    }
+
+    // Tracks when we finish presenting to hwc
+    nsecs_t estimatedEndTime = mCommitStartTimes[0];
+
+    // How long we spent this frame not doing anything, waiting for fences or vsync
+    nsecs_t idleDuration = 0;
+
+    // Most recent previous gpu end time in the current frame, probably from a prior display, used
+    // as the start time for the next gpu operation if it ran over time since it probably blocked
+    std::optional<nsecs_t> previousValidGpuEndTime;
+
+    // The currently estimated gpu end time for the frame,
+    // used to accumulate gpu time as we iterate over the active displays
+    std::optional<nsecs_t> estimatedGpuEndTime;
+
+    // If we're predicting at the start of the frame, we use last frame as our reference point
+    // If we're predicting at the end of the frame, we use the current frame as a reference point
+    nsecs_t referenceFrameStartTime = (earlyHint ? mCommitStartTimes[-1] : mCommitStartTimes[0]);
+
+    // We need an idea of when the last present fence fired and how long it made us wait
+    // If we're predicting at the start of the frame, we want frame n-2's present fence time
+    // If we're predicting at the end of the frame we want frame n-1's present time
+    nsecs_t referenceFenceTime =
+            (earlyHint ? mExpectedPresentTimes[-2] : mExpectedPresentTimes[-1]);
+    // The timing info for the previously calculated display, if there was one
+    std::optional<DisplayTimeline> previousDisplayReferenceTiming;
+    std::vector<DisplayId>&& displayIds =
+            getOrderedDisplayIds(&DisplayTimingData::presentStartTime);
+    DisplayTimeline referenceTiming, estimatedTiming;
+
+    // Iterate over the displays in the same order they are presented
+    for (DisplayId displayId : displayIds) {
+        if (mDisplayTimingData.count(displayId) == 0) {
+            continue;
+        }
+
+        auto& displayData = mDisplayTimingData.at(displayId);
+        referenceTiming = displayData.calculateDisplayTimeline(referenceFenceTime);
+
+        // If this is the first display, add the pre-present time to the total
+        if (!previousDisplayReferenceTiming.has_value()) {
+            estimatedEndTime += referenceTiming.prePresentTime - referenceFrameStartTime;
+        } else { // Otherwise add last display's postprocessing time to the total
+            estimatedEndTime += referenceTiming.prePresentTime -
+                    previousDisplayReferenceTiming->postPresentTime;
+        }
+
+        estimatedTiming = referenceTiming.estimateTimelineFromReference(mExpectedPresentTimes[-1],
+                                                                        estimatedEndTime);
+        // Update predicted present finish time with this display's present time
+        estimatedEndTime = estimatedTiming.postPresentTime;
+
+        // Track how long we spent waiting for the fence, can be excluded from the timing estimate
+        idleDuration += estimatedTiming.probablyWaitsForFence
+                ? mExpectedPresentTimes[-1] - estimatedTiming.preFenceWaitTime
+                : 0;
+
+        // Track how long we spent waiting to present, can be excluded from the timing estimate
+        idleDuration +=
+                !earlyHint ? referenceTiming.presentStartTime - referenceTiming.prePresentTime : 0;
+
+        // Estimate the reference frame's gpu timing
+        auto gpuTiming = displayData.estimateGpuTiming(previousValidGpuEndTime);
+        if (gpuTiming.has_value()) {
+            previousValidGpuEndTime = gpuTiming->startTime + gpuTiming->duration;
+
+            // Estimate the prediction frame's gpu end time from the reference frame
+            estimatedGpuEndTime =
+                    std::max(estimatedTiming.prePresentTime, estimatedGpuEndTime.value_or(0)) +
+                    gpuTiming->duration;
+        }
+        previousDisplayReferenceTiming = referenceTiming;
+    }
+    ATRACE_INT64("Idle duration", idleDuration);
+
+    // Don't count time spent idly waiting in the estimate as we could do more work in that time
+    estimatedEndTime -= idleDuration;
+
+    // We finish the frame when both present and the gpu are done, so wait for the later of the two
+    // Also add the frame delay duration since the target did not move while we were delayed
+    nsecs_t totalDuration = mFrameDelayDuration +
+            std::max(estimatedEndTime, estimatedGpuEndTime.value_or(0)) - mCommitStartTimes[0];
+
+    // We finish SurfaceFlinger when post-composition finishes, so add that in here
+    nsecs_t flingerDuration = estimatedEndTime + mLastPostcompDuration - mCommitStartTimes[0];
+    nsecs_t combinedDuration = combineTimingEstimates(totalDuration, flingerDuration);
+
+    return std::make_optional(combinedDuration);
+}
+
+nsecs_t PowerAdvisor::combineTimingEstimates(nsecs_t totalDuration, nsecs_t flingerDuration) {
+    nsecs_t targetDuration;
+    {
+        std::lock_guard lock(mPowerHalMutex);
+        targetDuration = *getPowerHal()->getTargetWorkDuration();
+    }
+    if (!mTotalFrameTargetDuration.has_value()) return flingerDuration;
+
+    // Normalize total to the flinger target (vsync period) since that's how often we actually send
+    // hints
+    nsecs_t normalizedTotalDuration = (targetDuration * totalDuration) / *mTotalFrameTargetDuration;
+    return std::max(flingerDuration, normalizedTotalDuration);
+}
+
+PowerAdvisor::DisplayTimeline PowerAdvisor::DisplayTimeline::estimateTimelineFromReference(
+        nsecs_t fenceTime, nsecs_t displayStartTime) {
+    DisplayTimeline estimated;
+    estimated.prePresentTime = displayStartTime;
+
+    // We don't predict waiting for vsync alignment yet
+    estimated.presentStartTime = estimated.prePresentTime;
+
+    // For now just re-use last frame's post-present duration and assume it will not change much
+    // How long we expect to run before we start waiting for the fence
+    estimated.preFenceWaitTime = estimated.presentStartTime + (preFenceWaitTime - presentStartTime);
+    estimated.probablyWaitsForFence = fenceTime > estimated.preFenceWaitTime;
+    estimated.postPresentTime = postFenceDuration +
+            (estimated.probablyWaitsForFence ? fenceTime : estimated.preFenceWaitTime);
+    return estimated;
+}
+
+PowerAdvisor::DisplayTimeline PowerAdvisor::DisplayTimingData::calculateDisplayTimeline(
+        nsecs_t fenceTime) {
+    DisplayTimeline timeline;
+    // How long between calling present from flinger and trying to wait on the fence in HWC
+    const nsecs_t preFenceWaitDelay =
+            (skippedValidate ? kPrefenceDelaySkippedValidate : kPrefenceDelayValidated).count();
+
+    // Did our reference frame wait for an earliest present time before calling the HWC
+    const bool waitedOnPresentTime = presentDelayedTime.has_value() &&
+            *presentDelayedTime > *presentStartTime && *presentDelayedTime < *presentEndTime;
+
+    // Use validate start here if we skipped it because we did validate + present together
+    timeline.prePresentTime = skippedValidate ? *validateStartTime : *presentStartTime;
+
+    // Use validate end here if we skipped it because we did validate + present together
+    timeline.postPresentTime = skippedValidate ? *validateEndTime : *presentEndTime;
+
+    // When we think we started waiting for the fence after calling into present
+    // This is after any time spent waiting for the earliest present time
+    timeline.presentStartTime =
+            (waitedOnPresentTime ? *presentDelayedTime : timeline.prePresentTime);
+    timeline.preFenceWaitTime = timeline.presentStartTime + preFenceWaitDelay;
+    timeline.probablyWaitsForFence =
+            fenceTime > timeline.preFenceWaitTime && fenceTime < timeline.postPresentTime;
+
+    // How long we ran after we finished waiting for the fence but before present happened
+    timeline.postFenceDuration = timeline.postPresentTime -
+            (timeline.probablyWaitsForFence ? fenceTime : timeline.preFenceWaitTime);
+    return timeline;
+}
+
+std::optional<PowerAdvisor::GpuTimeline> PowerAdvisor::DisplayTimingData::estimateGpuTiming(
+        std::optional<nsecs_t> previousEnd) {
+    if (!(usedClientComposition && lastValidGpuStartTime.has_value() && gpuEndFenceTime)) {
+        return std::nullopt;
+    }
+    const nsecs_t latestGpuStartTime = std::max(previousEnd.value_or(0), *gpuStartTime);
+    const nsecs_t latestGpuEndTime = gpuEndFenceTime->getSignalTime();
+    nsecs_t gpuDuration = 0;
+    if (latestGpuEndTime != Fence::SIGNAL_TIME_INVALID &&
+        latestGpuEndTime != Fence::SIGNAL_TIME_PENDING) {
+        // If we know how long the most recent gpu duration was, use that
+        gpuDuration = latestGpuEndTime - latestGpuStartTime;
+    } else if (lastValidGpuEndTime.has_value()) {
+        // If we don't have the fence data, use the most recent information we do have
+        gpuDuration = *lastValidGpuEndTime - *lastValidGpuStartTime;
+        if (latestGpuEndTime == Fence::SIGNAL_TIME_PENDING) {
+            // If pending but went over the previous duration, use current time as the end
+            gpuDuration = std::max(gpuDuration, systemTime() - latestGpuStartTime);
+        }
+    }
+    return GpuTimeline{.duration = gpuDuration, .startTime = latestGpuStartTime};
+}
+
 class HidlPowerHalWrapper : public PowerAdvisor::HalWrapper {
 public:
     HidlPowerHalWrapper(sp<V1_3::IPower> powerHal) : mPowerHal(std::move(powerHal)) {}
@@ -325,6 +619,10 @@
     }
 
     mSupportsPowerHint = checkPowerHintSessionSupported();
+
+    mAllowedActualDeviation =
+            base::GetIntProperty<nsecs_t>("debug.sf.allowed_actual_deviation",
+                                          std::chrono::nanoseconds(250us).count());
 }
 
 AidlPowerHalWrapper::~AidlPowerHalWrapper() {
@@ -332,7 +630,7 @@
         mPowerHintSession->close();
         mPowerHintSession = nullptr;
     }
-};
+}
 
 std::unique_ptr<PowerAdvisor::HalWrapper> AidlPowerHalWrapper::connect() {
     // This only waits if the service is actually declared
@@ -370,7 +668,7 @@
     return ret.isOk();
 }
 
-// only version 2+ of the aidl supports power hint sessions, hidl has no support
+// Only version 2+ of the aidl supports power hint sessions, hidl has no support
 bool AidlPowerHalWrapper::supportsPowerHintSession() {
     return mSupportsPowerHint;
 }
@@ -424,30 +722,14 @@
     return isPowerHintSessionRunning();
 }
 
-bool AidlPowerHalWrapper::shouldSetTargetDuration(int64_t targetDurationNanos) {
-    if (targetDurationNanos <= 0) {
-        return false;
-    }
-    // report if the change in target from our last submission to now exceeds the threshold
-    return abs(1.0 -
-               static_cast<double>(mLastTargetDurationSent) /
-                       static_cast<double>(targetDurationNanos)) >= kAllowedTargetDeviationPercent;
-}
-
-void AidlPowerHalWrapper::setTargetWorkDuration(int64_t targetDurationNanos) {
+void AidlPowerHalWrapper::setTargetWorkDuration(int64_t targetDuration) {
     ATRACE_CALL();
-    mTargetDuration = targetDurationNanos;
-    if (sTraceHintSessionData) ATRACE_INT64("Time target", targetDurationNanos);
-    if (!sNormalizeTarget && isPowerHintSessionRunning() &&
-        shouldSetTargetDuration(targetDurationNanos)) {
-        if (mLastActualDurationSent.has_value()) {
-            // update the error term here since we are actually sending an update to powerhal
-            if (sTraceHintSessionData)
-                ATRACE_INT64("Target error term", targetDurationNanos - *mLastActualDurationSent);
-        }
-        ALOGV("Sending target time: %" PRId64 "ns", targetDurationNanos);
-        mLastTargetDurationSent = targetDurationNanos;
-        auto ret = mPowerHintSession->updateTargetWorkDuration(targetDurationNanos);
+    mTargetDuration = targetDuration;
+    if (sTraceHintSessionData) ATRACE_INT64("Time target", targetDuration);
+    if (isPowerHintSessionRunning() && (targetDuration != mLastTargetDurationSent)) {
+        ALOGV("Sending target time: %" PRId64 "ns", targetDuration);
+        mLastTargetDurationSent = targetDuration;
+        auto ret = mPowerHintSession->updateTargetWorkDuration(targetDuration);
         if (!ret.isOk()) {
             ALOGW("Failed to set power hint target work duration with error: %s",
                   ret.exceptionMessage().c_str());
@@ -456,8 +738,8 @@
     }
 }
 
-bool AidlPowerHalWrapper::shouldReportActualDurationsNow() {
-    // report if we have never reported before or are approaching a stale session
+bool AidlPowerHalWrapper::shouldReportActualDurations() {
+    // Report if we have never reported before or are approaching a stale session
     if (!mLastActualDurationSent.has_value() ||
         (systemTime() - mLastActualReportTimestamp) > kStaleTimeout.count()) {
         return true;
@@ -466,65 +748,42 @@
     if (!mActualDuration.has_value()) {
         return false;
     }
-
-    // duration of most recent timing
-    const double mostRecentActualDuration = static_cast<double>(*mActualDuration);
-    // duration of the last timing actually reported to the powerhal
-    const double lastReportedActualDuration = static_cast<double>(*mLastActualDurationSent);
-
-    // report if the change in duration from then to now exceeds the threshold
-    return abs(1.0 - mostRecentActualDuration / lastReportedActualDuration) >=
-            kAllowedActualDeviationPercent;
+    // Report if the change in actual duration exceeds the threshold
+    return abs(*mActualDuration - *mLastActualDurationSent) > mAllowedActualDeviation;
 }
 
-void AidlPowerHalWrapper::sendActualWorkDuration(int64_t actualDurationNanos,
-                                                 nsecs_t timeStampNanos) {
+void AidlPowerHalWrapper::sendActualWorkDuration(int64_t actualDuration, nsecs_t timestamp) {
     ATRACE_CALL();
 
-    if (actualDurationNanos < 0 || !isPowerHintSessionRunning()) {
+    if (actualDuration < 0 || !isPowerHintSessionRunning()) {
         ALOGV("Failed to send actual work duration, skipping");
         return;
     }
-    nsecs_t reportedDuration = actualDurationNanos;
+    const nsecs_t reportedDuration = actualDuration;
 
-    // normalize the sent values to a pre-set target
-    if (sNormalizeTarget) {
-        reportedDuration += mLastTargetDurationSent - mTargetDuration;
-    } else {
-        // when target duration change is within deviation and not updated, adjust the actual
-        // duration proportionally based on the difference, e.g. if new target is 5ms longer than
-        // last reported but actual duration is the same as last target, we want to report a smaller
-        // actual work duration now to indicate that we are overshooting
-        if (mLastTargetDurationSent != kDefaultTarget.count() && mTargetDuration != 0) {
-            reportedDuration =
-                    static_cast<int64_t>(static_cast<long double>(mLastTargetDurationSent) /
-                                         mTargetDuration * actualDurationNanos);
-            mActualDuration = reportedDuration;
-        }
-    }
     mActualDuration = reportedDuration;
     WorkDuration duration;
     duration.durationNanos = reportedDuration;
-    duration.timeStampNanos = timeStampNanos;
+    duration.timeStampNanos = timestamp;
     mPowerHintQueue.push_back(duration);
 
     if (sTraceHintSessionData) {
-        ATRACE_INT64("Measured duration", actualDurationNanos);
-        ATRACE_INT64("Target error term", mTargetDuration - actualDurationNanos);
+        ATRACE_INT64("Measured duration", actualDuration);
+        ATRACE_INT64("Target error term", actualDuration - mTargetDuration);
 
         ATRACE_INT64("Reported duration", reportedDuration);
         ATRACE_INT64("Reported target", mLastTargetDurationSent);
-        ATRACE_INT64("Reported target error term", mLastTargetDurationSent - reportedDuration);
+        ATRACE_INT64("Reported target error term", reportedDuration - mLastTargetDurationSent);
     }
 
     ALOGV("Sending actual work duration of: %" PRId64 " on reported target: %" PRId64
           " with error: %" PRId64,
-          reportedDuration, mLastTargetDurationSent, mLastTargetDurationSent - reportedDuration);
+          reportedDuration, mLastTargetDurationSent, reportedDuration - mLastTargetDurationSent);
 
     // This rate limiter queues similar duration reports to the powerhal into
     // batches to avoid excessive binder calls. The criteria to send a given batch
     // are outlined in shouldReportActualDurationsNow()
-    if (shouldReportActualDurationsNow()) {
+    if (shouldReportActualDurations()) {
         ALOGV("Sending hint update batch");
         mLastActualReportTimestamp = systemTime();
         auto ret = mPowerHintSession->reportActualWorkDuration(mPowerHintQueue);
@@ -534,8 +793,8 @@
             mShouldReconnectHal = true;
         }
         mPowerHintQueue.clear();
-        // we save the non-normalized value here to detect % changes
-        mLastActualDurationSent = reportedDuration;
+        // We save the actual duration here for rate limiting
+        mLastActualDurationSent = actualDuration;
     }
 }
 
@@ -551,12 +810,13 @@
     return mTargetDuration;
 }
 
+void AidlPowerHalWrapper::setAllowedActualDeviation(nsecs_t allowedDeviation) {
+    mAllowedActualDeviation = allowedDeviation;
+}
+
 const bool AidlPowerHalWrapper::sTraceHintSessionData =
         base::GetBoolProperty(std::string("debug.sf.trace_hint_sessions"), false);
 
-const bool AidlPowerHalWrapper::sNormalizeTarget =
-        base::GetBoolProperty(std::string("debug.sf.normalize_hint_session_durations"), false);
-
 PowerAdvisor::HalWrapper* PowerAdvisor::getPowerHal() {
     static std::unique_ptr<HalWrapper> sHalWrapper = nullptr;
     static bool sHasHal = true;
@@ -565,7 +825,7 @@
         return nullptr;
     }
 
-    // grab old hint session values before we destroy any existing wrapper
+    // Grab old hint session values before we destroy any existing wrapper
     std::vector<int32_t> oldPowerHintSessionThreadIds;
     std::optional<int64_t> oldTargetWorkDuration;
 
@@ -582,7 +842,7 @@
 
     if (sHalWrapper != nullptr) {
         auto wrapper = sHalWrapper.get();
-        // if the wrapper is fine, return it, but if it indicates a reconnect, remake it
+        // If the wrapper is fine, return it, but if it indicates a reconnect, remake it
         if (!wrapper->shouldReconnectHAL()) {
             return wrapper;
         }
@@ -590,7 +850,7 @@
         sHalWrapper = nullptr;
     }
 
-    // at this point, we know for sure there is no running session
+    // At this point, we know for sure there is no running session
     mPowerHintSessionRunning = false;
 
     // First attempt to connect to the AIDL Power HAL
@@ -601,13 +861,12 @@
         sHalWrapper = HidlPowerHalWrapper::connect();
     } else {
         ALOGD("Successfully connecting AIDL Power HAL");
-        // if AIDL, pass on any existing hint session values
-        // thread ids always safe to set
+        // If AIDL, pass on any existing hint session values
         sHalWrapper->setPowerHintSessionThreadIds(oldPowerHintSessionThreadIds);
-        // only set duration and start if duration is defined
+        // Only set duration and start if duration is defined
         if (oldTargetWorkDuration.has_value()) {
             sHalWrapper->setTargetWorkDuration(*oldTargetWorkDuration);
-            // only start if possible to run and both threadids and duration are defined
+            // Only start if possible to run and both threadids and duration are defined
             if (usePowerHintSession() && !oldPowerHintSessionThreadIds.empty()) {
                 mPowerHintSessionRunning = sHalWrapper->startPowerHintSession();
             }
diff --git a/services/surfaceflinger/DisplayHardware/PowerAdvisor.h b/services/surfaceflinger/DisplayHardware/PowerAdvisor.h
index 7c10e19..8e1e33f 100644
--- a/services/surfaceflinger/DisplayHardware/PowerAdvisor.h
+++ b/services/surfaceflinger/DisplayHardware/PowerAdvisor.h
@@ -18,11 +18,15 @@
 
 #include <atomic>
 #include <chrono>
+#include <unordered_map>
 #include <unordered_set>
 
+#include <ui/DisplayId.h>
+#include <ui/FenceTime.h>
 #include <utils/Mutex.h>
 
 #include <android/hardware/power/IPower.h>
+#include <compositionengine/impl/OutputCompositionState.h>
 #include <ui/DisplayIdentification.h>
 #include "../Scheduler/OneShotTimer.h"
 
@@ -44,13 +48,47 @@
     virtual void setExpensiveRenderingExpected(DisplayId displayId, bool expected) = 0;
     virtual bool isUsingExpensiveRendering() = 0;
     virtual void notifyDisplayUpdateImminent() = 0;
+    // Checks both if it supports and if it's enabled
     virtual bool usePowerHintSession() = 0;
     virtual bool supportsPowerHintSession() = 0;
     virtual bool isPowerHintSessionRunning() = 0;
-    virtual void setTargetWorkDuration(int64_t targetDurationNanos) = 0;
-    virtual void sendActualWorkDuration(int64_t actualDurationNanos, nsecs_t timestamp) = 0;
+    // Sends a power hint that updates to the target work duration for the frame
+    virtual void setTargetWorkDuration(nsecs_t targetDuration) = 0;
+    // Sends a power hint for the actual known work duration at the end of the frame
+    virtual void sendActualWorkDuration() = 0;
+    // Sends a power hint for the upcoming frame predicted from previous frame timing
+    virtual void sendPredictedWorkDuration() = 0;
+    // Sets whether the power hint session is enabled
     virtual void enablePowerHint(bool enabled) = 0;
+    // Initializes the power hint session
     virtual bool startPowerHintSession(const std::vector<int32_t>& threadIds) = 0;
+    // Provides PowerAdvisor with a copy of the gpu fence so it can determine the gpu end time
+    virtual void setGpuFenceTime(DisplayId displayId, std::unique_ptr<FenceTime>&& fenceTime) = 0;
+    // Reports the start and end times of a present call this frame for a given display
+    virtual void setValidateTiming(DisplayId displayId, nsecs_t validateStartTime,
+                                   nsecs_t validateEndTime) = 0;
+    // Reports the start and end times of a present call this frame for a given display
+    virtual void setPresentTiming(DisplayId displayId, nsecs_t presentStartTime,
+                                  nsecs_t presentEndTime) = 0;
+    virtual void setExpectedPresentTime(nsecs_t expectedPresentTime) = 0;
+    // Reports whether a display used client composition this frame
+    virtual void setRequiresClientComposition(DisplayId displayId,
+                                              bool requiresClientComposition) = 0;
+    // Reports whether a given display skipped validation this frame
+    virtual void setSkippedValidate(DisplayId displayId, bool skipped) = 0;
+    // Reports how much a given display delayed its present call this frame
+    virtual void setPresentDelayedTime(
+            DisplayId displayId, std::chrono::steady_clock::time_point earliestFrameStartTime) = 0;
+    // Reports the start delay for SurfaceFlinger this frame
+    virtual void setFrameDelay(nsecs_t frameDelayDuration) = 0;
+    // Reports the SurfaceFlinger commit start time this frame
+    virtual void setCommitStart(nsecs_t commitStartTime) = 0;
+    // Reports the SurfaceFlinger composite end time this frame
+    virtual void setCompositeEnd(nsecs_t compositeEndTime) = 0;
+    // Reports the list of the currently active displays
+    virtual void setDisplays(std::vector<DisplayId>& displayIds) = 0;
+    // Sets the target duration for the entire pipeline including the gpu
+    virtual void setTotalFrameTargetWorkDuration(nsecs_t targetDuration) = 0;
 };
 
 namespace impl {
@@ -70,12 +108,11 @@
         virtual void restartPowerHintSession() = 0;
         virtual void setPowerHintSessionThreadIds(const std::vector<int32_t>& threadIds) = 0;
         virtual bool startPowerHintSession() = 0;
-        virtual void setTargetWorkDuration(int64_t targetDurationNanos) = 0;
-        virtual void sendActualWorkDuration(int64_t actualDurationNanos,
-                                            nsecs_t timeStampNanos) = 0;
+        virtual void setTargetWorkDuration(nsecs_t targetDuration) = 0;
+        virtual void sendActualWorkDuration(nsecs_t actualDuration, nsecs_t timestamp) = 0;
         virtual bool shouldReconnectHAL() = 0;
         virtual std::vector<int32_t> getPowerHintSessionThreadIds() = 0;
-        virtual std::optional<int64_t> getTargetWorkDuration() = 0;
+        virtual std::optional<nsecs_t> getTargetWorkDuration() = 0;
     };
 
     PowerAdvisor(SurfaceFlinger& flinger);
@@ -89,10 +126,28 @@
     bool usePowerHintSession() override;
     bool supportsPowerHintSession() override;
     bool isPowerHintSessionRunning() override;
-    void setTargetWorkDuration(int64_t targetDurationNanos) override;
-    void sendActualWorkDuration(int64_t actualDurationNanos, nsecs_t timestamp) override;
+    void setTargetWorkDuration(nsecs_t targetDuration) override;
+    void sendActualWorkDuration() override;
+    void sendPredictedWorkDuration() override;
     void enablePowerHint(bool enabled) override;
     bool startPowerHintSession(const std::vector<int32_t>& threadIds) override;
+    void setGpuFenceTime(DisplayId displayId, std::unique_ptr<FenceTime>&& fenceTime);
+    void setValidateTiming(DisplayId displayId, nsecs_t valiateStartTime,
+                           nsecs_t validateEndTime) override;
+    void setPresentTiming(DisplayId displayId, nsecs_t presentStartTime,
+                          nsecs_t presentEndTime) override;
+    void setSkippedValidate(DisplayId displayId, bool skipped) override;
+    void setRequiresClientComposition(DisplayId displayId, bool requiresClientComposition) override;
+    void setExpectedPresentTime(nsecs_t expectedPresentTime) override;
+    void setPresentDelayedTime(
+            DisplayId displayId,
+            std::chrono::steady_clock::time_point earliestFrameStartTime) override;
+
+    void setFrameDelay(nsecs_t frameDelayDuration) override;
+    void setCommitStart(nsecs_t commitStartTime) override;
+    void setCompositeEnd(nsecs_t compositeEndTime) override;
+    void setDisplays(std::vector<DisplayId>& displayIds) override;
+    void setTotalFrameTargetWorkDuration(nsecs_t targetDuration) override;
 
 private:
     HalWrapper* getPowerHal() REQUIRES(mPowerHalMutex);
@@ -100,15 +155,6 @@
     std::mutex mPowerHalMutex;
 
     std::atomic_bool mBootFinished = false;
-    std::optional<bool> mPowerHintEnabled;
-    std::optional<bool> mSupportsPowerHint;
-    bool mPowerHintSessionRunning = false;
-
-    // An adjustable safety margin which moves the "target" earlier to allow flinger to
-    // go a bit over without dropping a frame, especially since we can't measure
-    // the exact time HWC finishes composition so "actual" durations are measured
-    // from the end of present() instead, which is a bit later.
-    static constexpr const std::chrono::nanoseconds kTargetSafetyMargin = 2ms;
 
     std::unordered_set<DisplayId> mExpensiveDisplays;
     bool mNotifiedExpensiveRendering = false;
@@ -117,6 +163,108 @@
     std::atomic_bool mSendUpdateImminent = true;
     std::atomic<nsecs_t> mLastScreenUpdatedTime = 0;
     std::optional<scheduler::OneShotTimer> mScreenUpdateTimer;
+
+    // Higher-level timing data used for estimation
+    struct DisplayTimeline {
+        nsecs_t prePresentTime = -1;
+        nsecs_t postPresentTime = -1;
+        // Usually equals prePresentTime but can be delayed if we wait for the next valid vsync
+        nsecs_t presentStartTime = -1;
+        // When we think we started waiting for the fence after calling into present and
+        // after potentially waiting for the earliest present time
+        nsecs_t preFenceWaitTime = -1;
+        // How long we ran after we finished waiting for the fence but before present happened
+        nsecs_t postFenceDuration = 0;
+        // Are we likely to have waited for the present fence during composition
+        bool probablyWaitsForFence = false;
+        // Estimate one frame's timeline from that of a previous frame
+        DisplayTimeline estimateTimelineFromReference(nsecs_t fenceTime, nsecs_t displayStartTime);
+    };
+
+    struct GpuTimeline {
+        nsecs_t duration = 0;
+        nsecs_t startTime = -1;
+    };
+
+    // Power hint session data recorded from the pipeline
+    struct DisplayTimingData {
+        std::unique_ptr<FenceTime> gpuEndFenceTime;
+        std::optional<nsecs_t> gpuStartTime;
+        std::optional<nsecs_t> lastValidGpuEndTime;
+        std::optional<nsecs_t> lastValidGpuStartTime;
+        std::optional<nsecs_t> presentStartTime;
+        std::optional<nsecs_t> presentEndTime;
+        std::optional<nsecs_t> validateStartTime;
+        std::optional<nsecs_t> validateEndTime;
+        std::optional<nsecs_t> presentDelayedTime;
+        bool usedClientComposition = false;
+        bool skippedValidate = false;
+        // Calculate high-level timing milestones from more granular display timing data
+        DisplayTimeline calculateDisplayTimeline(nsecs_t fenceTime);
+        // Estimate the gpu duration for a given display from previous gpu timing data
+        std::optional<GpuTimeline> estimateGpuTiming(std::optional<nsecs_t> previousEnd);
+    };
+
+    template <class T, size_t N>
+    class RingBuffer {
+        std::array<T, N> elements = {};
+        size_t mIndex = 0;
+        size_t numElements = 0;
+
+    public:
+        void append(T item) {
+            mIndex = (mIndex + 1) % N;
+            numElements = std::min(N, numElements + 1);
+            elements[mIndex] = item;
+        }
+        bool isFull() const { return numElements == N; }
+        // Allows access like [0] == current, [-1] = previous, etc..
+        T& operator[](int offset) {
+            size_t positiveOffset =
+                    static_cast<size_t>((offset % static_cast<int>(N)) + static_cast<int>(N));
+            return elements[(mIndex + positiveOffset) % N];
+        }
+    };
+
+    // Filter and sort the display ids by a given property
+    std::vector<DisplayId> getOrderedDisplayIds(std::optional<nsecs_t> DisplayTimingData::*sortBy);
+    // Estimates a frame's total work duration including gpu time.
+    // Runs either at the beginning or end of a frame, using the most recent data available
+    std::optional<nsecs_t> estimateWorkDuration(bool earlyHint);
+    // There are two different targets and actual work durations we care about,
+    // this normalizes them together and takes the max of the two
+    nsecs_t combineTimingEstimates(nsecs_t totalDuration, nsecs_t flingerDuration);
+
+    std::unordered_map<DisplayId, DisplayTimingData> mDisplayTimingData;
+
+    // Current frame's delay
+    nsecs_t mFrameDelayDuration = 0;
+    // Last frame's composite end time
+    nsecs_t mLastCompositeEndTime = -1;
+    // Last frame's post-composition duration
+    nsecs_t mLastPostcompDuration = 0;
+    // Buffer of recent commit start times
+    RingBuffer<nsecs_t, 2> mCommitStartTimes;
+    // Buffer of recent expected present times
+    RingBuffer<nsecs_t, 3> mExpectedPresentTimes;
+    // Target for the entire pipeline including gpu
+    std::optional<nsecs_t> mTotalFrameTargetDuration;
+    // Updated list of display IDs
+    std::vector<DisplayId> mDisplayIds;
+
+    std::optional<bool> mPowerHintEnabled;
+    std::optional<bool> mSupportsPowerHint;
+    bool mPowerHintSessionRunning = false;
+
+    // An adjustable safety margin which moves the "target" earlier to allow flinger to
+    // go a bit over without dropping a frame, especially since we can't measure
+    // the exact time HWC finishes composition so "actual" durations are measured
+    // from the end of present() instead, which is a bit later.
+    static constexpr const std::chrono::nanoseconds kTargetSafetyMargin = 1ms;
+
+    // How long we expect hwc to run after the present call until it waits for the fence
+    static constexpr const std::chrono::nanoseconds kPrefenceDelayValidated = 150us;
+    static constexpr const std::chrono::nanoseconds kPrefenceDelaySkippedValidate = 250us;
 };
 
 class AidlPowerHalWrapper : public PowerAdvisor::HalWrapper {
@@ -133,50 +281,50 @@
     void restartPowerHintSession() override;
     void setPowerHintSessionThreadIds(const std::vector<int32_t>& threadIds) override;
     bool startPowerHintSession() override;
-    void setTargetWorkDuration(int64_t targetDurationNanos) override;
-    void sendActualWorkDuration(int64_t actualDurationNanos, nsecs_t timeStampNanos) override;
+    void setTargetWorkDuration(nsecs_t targetDuration) override;
+    void sendActualWorkDuration(nsecs_t actualDuration, nsecs_t timestamp) override;
     bool shouldReconnectHAL() override;
     std::vector<int32_t> getPowerHintSessionThreadIds() override;
-    std::optional<int64_t> getTargetWorkDuration() override;
+    std::optional<nsecs_t> getTargetWorkDuration() override;
 
 private:
+    friend class AidlPowerHalWrapperTest;
+
     bool checkPowerHintSessionSupported();
     void closePowerHintSession();
-    bool shouldReportActualDurationsNow();
-    bool shouldSetTargetDuration(int64_t targetDurationNanos);
+    bool shouldReportActualDurations();
+
+    // Used for testing
+    void setAllowedActualDeviation(nsecs_t);
 
     const sp<hardware::power::IPower> mPowerHal = nullptr;
     bool mHasExpensiveRendering = false;
     bool mHasDisplayUpdateImminent = false;
     // Used to indicate an error state and need for reconstruction
     bool mShouldReconnectHal = false;
-    // This is not thread safe, but is currently protected by mPowerHalMutex so it needs no lock
+
+    // Power hint session data
+
+    // Concurrent access for this is protected by mPowerHalMutex
     sp<hardware::power::IPowerHintSession> mPowerHintSession = nullptr;
     // Queue of actual durations saved to report
     std::vector<hardware::power::WorkDuration> mPowerHintQueue;
-    // The latest un-normalized values we have received for target and actual
-    int64_t mTargetDuration = kDefaultTarget.count();
-    std::optional<int64_t> mActualDuration;
+    // The latest values we have received for target and actual
+    nsecs_t mTargetDuration = kDefaultTarget.count();
+    std::optional<nsecs_t> mActualDuration;
     // The list of thread ids, stored so we can restart the session from this class if needed
     std::vector<int32_t> mPowerHintThreadIds;
-    bool mSupportsPowerHint;
+    bool mSupportsPowerHint = false;
     // Keep track of the last messages sent for rate limiter change detection
-    std::optional<int64_t> mLastActualDurationSent;
-    // timestamp of the last report we sent, used to avoid stale sessions
-    int64_t mLastActualReportTimestamp = 0;
-    int64_t mLastTargetDurationSent = kDefaultTarget.count();
-    // Whether to normalize all the actual values as error terms relative to a constant target
-    // This saves a binder call by not setting the target, and should not affect the pid values
-    static const bool sNormalizeTarget;
+    std::optional<nsecs_t> mLastActualDurationSent;
+    // Timestamp of the last report we sent, used to avoid stale sessions
+    nsecs_t mLastActualReportTimestamp = 0;
+    nsecs_t mLastTargetDurationSent = kDefaultTarget.count();
+    // Max amount the error term can vary without causing an actual value report
+    nsecs_t mAllowedActualDeviation = -1;
     // Whether we should emit ATRACE_INT data for hint sessions
     static const bool sTraceHintSessionData;
-
-    // Max percent the actual duration can vary without causing a report (eg: 0.1 = 10%)
-    static constexpr double kAllowedActualDeviationPercent = 0.1;
-    // Max percent the target duration can vary without causing a report (eg: 0.1 = 10%)
-    static constexpr double kAllowedTargetDeviationPercent = 0.1;
-    // Target used for init and normalization, the actual value does not really matter
-    static constexpr const std::chrono::nanoseconds kDefaultTarget = 50ms;
+    static constexpr const std::chrono::nanoseconds kDefaultTarget = 16ms;
     // Amount of time after the last message was sent before the session goes stale
     // actually 100ms but we use 80 here to ideally avoid going stale
     static constexpr const std::chrono::nanoseconds kStaleTimeout = 80ms;
diff --git a/services/surfaceflinger/Scheduler/Scheduler.cpp b/services/surfaceflinger/Scheduler/Scheduler.cpp
index 37f0fec..08a1ede 100644
--- a/services/surfaceflinger/Scheduler/Scheduler.cpp
+++ b/services/surfaceflinger/Scheduler/Scheduler.cpp
@@ -554,11 +554,12 @@
     }
 }
 
-void Scheduler::setDisplayPowerState(bool normal) {
+void Scheduler::setDisplayPowerMode(hal::PowerMode powerMode) {
     {
         std::lock_guard<std::mutex> lock(mPolicyLock);
-        mPolicy.isDisplayPowerStateNormal = normal;
+        mPolicy.displayPowerMode = powerMode;
     }
+    mVsyncSchedule->getController().setDisplayPowerMode(powerMode);
 
     if (mDisplayPowerTimer) {
         mDisplayPowerTimer->reset();
@@ -706,7 +707,8 @@
     // If Display Power is not in normal operation we want to be in performance mode. When coming
     // back to normal mode, a grace period is given with DisplayPowerTimer.
     if (mDisplayPowerTimer &&
-        (!mPolicy.isDisplayPowerStateNormal || mPolicy.displayPowerTimer == TimerState::Reset)) {
+        (mPolicy.displayPowerMode != hal::PowerMode::ON ||
+         mPolicy.displayPowerTimer == TimerState::Reset)) {
         constexpr GlobalSignals kNoSignals;
         return {configs->getMaxRefreshRateByPolicy(), kNoSignals};
     }
diff --git a/services/surfaceflinger/Scheduler/Scheduler.h b/services/surfaceflinger/Scheduler/Scheduler.h
index 0c72124..a8043bf 100644
--- a/services/surfaceflinger/Scheduler/Scheduler.h
+++ b/services/surfaceflinger/Scheduler/Scheduler.h
@@ -186,7 +186,7 @@
     // Indicates that touch interaction is taking place.
     void onTouchHint();
 
-    void setDisplayPowerState(bool normal);
+    void setDisplayPowerMode(hal::PowerMode powerMode);
 
     VSyncDispatch& getVsyncDispatch() { return mVsyncSchedule->getDispatch(); }
 
@@ -325,7 +325,7 @@
         TimerState idleTimer = TimerState::Reset;
         TouchState touch = TouchState::Inactive;
         TimerState displayPowerTimer = TimerState::Expired;
-        bool isDisplayPowerStateNormal = true;
+        hal::PowerMode displayPowerMode = hal::PowerMode::ON;
 
         // Chosen display mode.
         DisplayModePtr mode;
diff --git a/services/surfaceflinger/Scheduler/VSyncReactor.cpp b/services/surfaceflinger/Scheduler/VSyncReactor.cpp
index bdcab51..13cd304 100644
--- a/services/surfaceflinger/Scheduler/VSyncReactor.cpp
+++ b/services/surfaceflinger/Scheduler/VSyncReactor.cpp
@@ -146,6 +146,11 @@
         return false;
     }
 
+    if (mDisplayPowerMode == hal::PowerMode::DOZE ||
+        mDisplayPowerMode == hal::PowerMode::DOZE_SUSPEND) {
+        return true;
+    }
+
     if (!mLastHwVsync && !HwcVsyncPeriod) {
         return false;
     }
@@ -206,6 +211,11 @@
     return mMoreSamplesNeeded;
 }
 
+void VSyncReactor::setDisplayPowerMode(hal::PowerMode powerMode) {
+    std::scoped_lock lock(mMutex);
+    mDisplayPowerMode = powerMode;
+}
+
 void VSyncReactor::dump(std::string& result) const {
     std::lock_guard lock(mMutex);
     StringAppendF(&result, "VsyncReactor in use\n");
diff --git a/services/surfaceflinger/Scheduler/VSyncReactor.h b/services/surfaceflinger/Scheduler/VSyncReactor.h
index 6a1950a..4501487 100644
--- a/services/surfaceflinger/Scheduler/VSyncReactor.h
+++ b/services/surfaceflinger/Scheduler/VSyncReactor.h
@@ -49,6 +49,8 @@
     bool addHwVsyncTimestamp(nsecs_t timestamp, std::optional<nsecs_t> hwcVsyncPeriod,
                              bool* periodFlushed) final;
 
+    void setDisplayPowerMode(hal::PowerMode powerMode) final;
+
     void dump(std::string& result) const final;
 
 private:
@@ -73,6 +75,8 @@
     std::optional<nsecs_t> mPeriodTransitioningTo GUARDED_BY(mMutex);
     std::optional<nsecs_t> mLastHwVsync GUARDED_BY(mMutex);
 
+    hal::PowerMode mDisplayPowerMode GUARDED_BY(mMutex) = hal::PowerMode::ON;
+
     const bool mSupportKernelIdleTimer = false;
 };
 
diff --git a/services/surfaceflinger/Scheduler/VsyncController.h b/services/surfaceflinger/Scheduler/VsyncController.h
index 59f6537..726a420 100644
--- a/services/surfaceflinger/Scheduler/VsyncController.h
+++ b/services/surfaceflinger/Scheduler/VsyncController.h
@@ -18,7 +18,10 @@
 
 #include <cstddef>
 #include <memory>
+#include <mutex>
 
+#include <DisplayHardware/HWComposer.h>
+#include <DisplayHardware/Hal.h>
 #include <ui/FenceTime.h>
 #include <utils/Mutex.h>
 #include <utils/RefBase.h>
@@ -70,6 +73,13 @@
      */
     virtual void setIgnorePresentFences(bool ignore) = 0;
 
+    /*
+     * Sets the primary display power mode to the controller.
+     *
+     * \param [in] powerMode
+     */
+    virtual void setDisplayPowerMode(hal::PowerMode powerMode) = 0;
+
     virtual void dump(std::string& result) const = 0;
 
 protected:
diff --git a/services/surfaceflinger/SurfaceFlinger.cpp b/services/surfaceflinger/SurfaceFlinger.cpp
index 6a17cd8..b7d6968 100644
--- a/services/surfaceflinger/SurfaceFlinger.cpp
+++ b/services/surfaceflinger/SurfaceFlinger.cpp
@@ -444,6 +444,11 @@
     }
 
     mIgnoreHdrCameraLayers = ignore_hdr_camera_layers(false);
+
+    // Power hint session mode, representing which hint(s) to send: early, late, or both)
+    mPowerHintSessionMode =
+            {.late = base::GetBoolProperty("debug.sf.send_late_power_session_hint"s, true),
+             .early = base::GetBoolProperty("debug.sf.send_early_power_session_hint"s, true)};
 }
 
 LatchUnsignaledConfig SurfaceFlinger::getLatchUnsignaledConfig() {
@@ -1991,12 +1996,6 @@
 
 bool SurfaceFlinger::commit(nsecs_t frameTime, int64_t vsyncId, nsecs_t expectedVsyncTime)
         FTL_FAKE_GUARD(kMainThreadContext) {
-    // we set this once at the beginning of commit to ensure consistency throughout the whole frame
-    mPowerHintSessionData.sessionEnabled = mPowerAdvisor->usePowerHintSession();
-    if (mPowerHintSessionData.sessionEnabled) {
-        mPowerHintSessionData.commitStart = systemTime();
-    }
-
     // calculate the expected present time once and use the cached
     // value throughout this frame to make sure all layers are
     // seeing this same value.
@@ -2010,10 +2009,6 @@
     const nsecs_t lastScheduledPresentTime = mScheduledPresentTime;
     mScheduledPresentTime = expectedVsyncTime;
 
-    if (mPowerHintSessionData.sessionEnabled) {
-        mPowerAdvisor->setTargetWorkDuration(mExpectedPresentTime -
-                                             mPowerHintSessionData.commitStart);
-    }
     const auto vsyncIn = [&] {
         if (!ATRACE_ENABLED()) return 0.f;
         return (mExpectedPresentTime - systemTime()) / 1e6f;
@@ -2089,6 +2084,30 @@
         }
     }
 
+    // Save this once per commit + composite to ensure consistency
+    mPowerHintSessionEnabled = mPowerAdvisor->usePowerHintSession();
+    if (mPowerHintSessionEnabled) {
+        nsecs_t vsyncPeriod;
+        {
+            Mutex::Autolock lock(mStateLock);
+            vsyncPeriod = getVsyncPeriodFromHWC();
+        }
+        mPowerAdvisor->setCommitStart(frameTime);
+        mPowerAdvisor->setExpectedPresentTime(mExpectedPresentTime);
+        const nsecs_t idealSfWorkDuration =
+                mVsyncModulator->getVsyncConfig().sfWorkDuration.count();
+        // Frame delay is how long we should have minus how long we actually have
+        mPowerAdvisor->setFrameDelay(idealSfWorkDuration - (mExpectedPresentTime - frameTime));
+        mPowerAdvisor->setTotalFrameTargetWorkDuration(idealSfWorkDuration);
+        mPowerAdvisor->setTargetWorkDuration(vsyncPeriod);
+
+        // Send early hint here to make sure there's not another frame pending
+        if (mPowerHintSessionMode.early) {
+            // Send a rough prediction for this frame based on last frame's timing info
+            mPowerAdvisor->sendPredictedWorkDuration();
+        }
+    }
+
     if (mTracingEnabledChanged) {
         mLayerTracingEnabled = mLayerTracing.isEnabled();
         mTracingEnabledChanged = false;
@@ -2165,16 +2184,15 @@
         FTL_FAKE_GUARD(kMainThreadContext) {
     ATRACE_FORMAT("%s %" PRId64, __func__, vsyncId);
 
-    if (mPowerHintSessionData.sessionEnabled) {
-        mPowerHintSessionData.compositeStart = systemTime();
-    }
-
     compositionengine::CompositionRefreshArgs refreshArgs;
     const auto& displays = FTL_FAKE_GUARD(mStateLock, mDisplays);
     refreshArgs.outputs.reserve(displays.size());
+    std::vector<DisplayId> displayIds;
     for (const auto& [_, display] : displays) {
         refreshArgs.outputs.push_back(display->getCompositionDisplay());
+        displayIds.push_back(display->getId());
     }
+    mPowerAdvisor->setDisplays(displayIds);
     mDrawingState.traverseInZOrder([&refreshArgs](Layer* layer) {
         if (auto layerFE = layer->getCompositionEngineLayerFE())
             refreshArgs.layers.push_back(layerFE);
@@ -2222,12 +2240,15 @@
 
     mCompositionEngine->present(refreshArgs);
 
-    if (mPowerHintSessionData.sessionEnabled) {
-        mPowerHintSessionData.presentEnd = systemTime();
-    }
-
     mTimeStats->recordFrameDuration(frameTime, systemTime());
 
+    // Send a power hint hint after presentation is finished
+    if (mPowerHintSessionEnabled) {
+        if (mPowerHintSessionMode.late) {
+            mPowerAdvisor->sendActualWorkDuration();
+        }
+    }
+
     if (mScheduler->onPostComposition(presentTime)) {
         scheduleComposite(FrameHint::kNone);
     }
@@ -2272,11 +2293,8 @@
         scheduleCommit(FrameHint::kNone);
     }
 
-    // calculate total render time for performance hinting if adpf cpu hint is enabled,
-    if (mPowerHintSessionData.sessionEnabled) {
-        const nsecs_t flingerDuration =
-                (mPowerHintSessionData.presentEnd - mPowerHintSessionData.commitStart);
-        mPowerAdvisor->sendActualWorkDuration(flingerDuration, mPowerHintSessionData.presentEnd);
+    if (mPowerHintSessionEnabled) {
+        mPowerAdvisor->setCompositeEnd(systemTime());
     }
 }
 
@@ -4929,7 +4947,7 @@
     if (isDisplayActiveLocked(display)) {
         mTimeStats->setPowerMode(mode);
         mRefreshRateStats->setPowerMode(mode);
-        mScheduler->setDisplayPowerState(mode == hal::PowerMode::ON);
+        mScheduler->setDisplayPowerMode(mode);
     }
 
     ALOGD("Finished setting power mode %d on display %s", mode, to_string(displayId).c_str());
@@ -6895,9 +6913,7 @@
         return NO_ERROR;
     }
 
-    status_t setPolicyResult = overridePolicy
-            ? display->refreshRateConfigs().setOverridePolicy(policy)
-            : display->refreshRateConfigs().setDisplayManagerPolicy(*policy);
+    const status_t setPolicyResult = display->setRefreshRatePolicy(policy, overridePolicy);
     if (setPolicyResult < 0) {
         return BAD_VALUE;
     }
diff --git a/services/surfaceflinger/SurfaceFlinger.h b/services/surfaceflinger/SurfaceFlinger.h
index f14c755..83134a2 100644
--- a/services/surfaceflinger/SurfaceFlinger.h
+++ b/services/surfaceflinger/SurfaceFlinger.h
@@ -1435,12 +1435,12 @@
         return mScheduler->getLayerFramerate(now, id);
     }
 
+    bool mPowerHintSessionEnabled;
+
     struct {
-        bool sessionEnabled = false;
-        nsecs_t commitStart;
-        nsecs_t compositeStart;
-        nsecs_t presentEnd;
-    } mPowerHintSessionData GUARDED_BY(kMainThreadContext);
+        bool late = false;
+        bool early = false;
+    } mPowerHintSessionMode;
 
     nsecs_t mAnimationTransactionTimeout = s2ns(5);
 
diff --git a/services/surfaceflinger/tests/unittests/AidlPowerHalWrapperTest.cpp b/services/surfaceflinger/tests/unittests/AidlPowerHalWrapperTest.cpp
index 9ab35d7..53de4a6 100644
--- a/services/surfaceflinger/tests/unittests/AidlPowerHalWrapperTest.cpp
+++ b/services/surfaceflinger/tests/unittests/AidlPowerHalWrapperTest.cpp
@@ -52,6 +52,8 @@
     void verifyAndClearExpectations();
     void sendActualWorkDurationGroup(std::vector<WorkDuration> durations,
                                      std::chrono::nanoseconds sleepBeforeLastSend);
+    std::chrono::nanoseconds mAllowedDeviation;
+    std::chrono::nanoseconds mStaleTimeout;
 };
 
 void AidlPowerHalWrapperTest::SetUp() {
@@ -59,6 +61,9 @@
     mMockSession = new NiceMock<MockIPowerHintSession>();
     ON_CALL(*mMockHal.get(), getHintSessionPreferredRate(_)).WillByDefault(Return(Status::ok()));
     mWrapper = std::make_unique<AidlPowerHalWrapper>(mMockHal);
+    mWrapper->setAllowedActualDeviation(std::chrono::nanoseconds{10ms}.count());
+    mAllowedDeviation = std::chrono::nanoseconds{mWrapper->mAllowedActualDeviation};
+    mStaleTimeout = AidlPowerHalWrapper::kStaleTimeout;
 }
 
 void AidlPowerHalWrapperTest::verifyAndClearExpectations() {
@@ -76,6 +81,7 @@
         mWrapper->sendActualWorkDuration(duration.durationNanos, duration.timeStampNanos);
     }
 }
+
 WorkDuration toWorkDuration(std::chrono::nanoseconds durationNanos, int64_t timeStampNanos) {
     WorkDuration duration;
     duration.durationNanos = durationNanos.count();
@@ -83,6 +89,10 @@
     return duration;
 }
 
+WorkDuration toWorkDuration(std::pair<std::chrono::nanoseconds, nsecs_t> timePair) {
+    return toWorkDuration(timePair.first, timePair.second);
+}
+
 std::string printWorkDurations(const ::std::vector<WorkDuration>& durations) {
     std::ostringstream os;
     for (auto duration : durations) {
@@ -112,7 +122,7 @@
     EXPECT_FALSE(mWrapper->startPowerHintSession());
 }
 
-TEST_F(AidlPowerHalWrapperTest, restartNewPoserHintSessionWithNewThreadIds) {
+TEST_F(AidlPowerHalWrapperTest, restartNewPowerHintSessionWithNewThreadIds) {
     ASSERT_TRUE(mWrapper->supportsPowerHintSession());
 
     std::vector<int32_t> threadIds = {1, 2};
@@ -149,12 +159,8 @@
 
     std::chrono::nanoseconds base = 100ms;
     // test cases with target work duration and whether it should update hint against baseline 100ms
-    const std::vector<std::pair<std::chrono::nanoseconds, bool>> testCases = {{0ms, false},
-                                                                              {-1ms, false},
-                                                                              {200ms, true},
-                                                                              {2ms, true},
-                                                                              {91ms, false},
-                                                                              {109ms, false}};
+    const std::vector<std::pair<std::chrono::nanoseconds, bool>> testCases =
+            {{0ms, true}, {-1ms, true}, {200ms, true}, {2ms, true}, {100ms, false}, {109ms, true}};
 
     for (const auto& test : testCases) {
         // reset to 100ms baseline
@@ -200,21 +206,21 @@
     // 100ms
     const std::vector<std::pair<std::vector<std::pair<std::chrono::nanoseconds, nsecs_t>>, bool>>
             testCases = {{{{-1ms, 100}}, false},
-                         {{{91ms, 100}}, false},
-                         {{{109ms, 100}}, false},
+                         {{{100ms - (mAllowedDeviation / 2), 100}}, false},
+                         {{{100ms + (mAllowedDeviation / 2), 100}}, false},
+                         {{{100ms + (mAllowedDeviation + 1ms), 100}}, true},
+                         {{{100ms - (mAllowedDeviation + 1ms), 100}}, true},
                          {{{100ms, 100}, {200ms, 200}}, true},
                          {{{100ms, 500}, {100ms, 600}, {3ms, 600}}, true}};
 
     for (const auto& test : testCases) {
         // reset actual duration
-        sendActualWorkDurationGroup({base}, 80ms);
+        sendActualWorkDurationGroup({base}, mStaleTimeout);
 
         auto raw = test.first;
         std::vector<WorkDuration> durations(raw.size());
         std::transform(raw.begin(), raw.end(), durations.begin(),
-                       [](std::pair<std::chrono::nanoseconds, nsecs_t> d) {
-                           return toWorkDuration(d.first, d.second);
-                       });
+                       [](auto d) { return toWorkDuration(d); });
         EXPECT_CALL(*mMockSession.get(), reportActualWorkDuration(durations))
                 .Times(test.second ? 1 : 0);
         sendActualWorkDurationGroup(durations, 0ms);
@@ -222,40 +228,6 @@
     }
 }
 
-TEST_F(AidlPowerHalWrapperTest, sendAdjustedActualWorkDuration) {
-    ASSERT_TRUE(mWrapper->supportsPowerHintSession());
-
-    std::vector<int32_t> threadIds = {1, 2};
-    mWrapper->setPowerHintSessionThreadIds(threadIds);
-    EXPECT_CALL(*mMockHal.get(), createHintSession(_, _, threadIds, _, _))
-            .WillOnce(DoAll(SetArgPointee<4>(mMockSession), Return(Status::ok())));
-    ASSERT_TRUE(mWrapper->startPowerHintSession());
-    verifyAndClearExpectations();
-
-    std::chrono::nanoseconds lastTarget = 100ms;
-    EXPECT_CALL(*mMockSession.get(), updateTargetWorkDuration(lastTarget.count())).Times(1);
-    mWrapper->setTargetWorkDuration(lastTarget.count());
-    std::chrono::nanoseconds newTarget = 105ms;
-    mWrapper->setTargetWorkDuration(newTarget.count());
-    EXPECT_CALL(*mMockSession.get(), updateTargetWorkDuration(newTarget.count())).Times(0);
-    std::chrono::nanoseconds actual = 21ms;
-    // 100 / 105 * 21ms = 20ms
-    std::chrono::nanoseconds expectedActualSent = 20ms;
-    std::vector<WorkDuration> expectedDurations = {toWorkDuration(expectedActualSent, 1)};
-
-    EXPECT_CALL(*mMockSession.get(), reportActualWorkDuration(_))
-            .WillOnce(DoAll(
-                    [expectedDurations](const ::std::vector<WorkDuration>& durationsSent) {
-                        EXPECT_EQ(expectedDurations, durationsSent)
-                                << base::StringPrintf("actual sent: %s vs expected: %s",
-                                                      printWorkDurations(durationsSent).c_str(),
-                                                      printWorkDurations(expectedDurations)
-                                                              .c_str());
-                    },
-                    Return(Status::ok())));
-    mWrapper->sendActualWorkDuration(actual.count(), 1);
-}
-
 TEST_F(AidlPowerHalWrapperTest, sendActualWorkDuration_exceedsStaleTime) {
     ASSERT_TRUE(mWrapper->supportsPowerHintSession());
 
@@ -269,22 +241,23 @@
     auto base = toWorkDuration(100ms, 0);
     // test cases with actual work durations and whether it should update hint against baseline
     // 100ms
-    const std::vector<std::pair<std::vector<std::pair<std::chrono::nanoseconds, nsecs_t>>, bool>>
-            testCases = {{{{91ms, 100}}, true}, {{{109ms, 100}}, true}};
+    const std::vector<std::tuple<std::vector<std::pair<std::chrono::nanoseconds, nsecs_t>>,
+                                 std::chrono::nanoseconds, bool>>
+            testCases = {{{{100ms, 100}}, mStaleTimeout, true},
+                         {{{100ms + (mAllowedDeviation / 2), 100}}, mStaleTimeout, true},
+                         {{{100ms, 100}}, mStaleTimeout / 2, false}};
 
     for (const auto& test : testCases) {
         // reset actual duration
-        sendActualWorkDurationGroup({base}, 80ms);
+        sendActualWorkDurationGroup({base}, mStaleTimeout);
 
-        auto raw = test.first;
+        auto raw = std::get<0>(test);
         std::vector<WorkDuration> durations(raw.size());
         std::transform(raw.begin(), raw.end(), durations.begin(),
-                       [](std::pair<std::chrono::nanoseconds, nsecs_t> d) {
-                           return toWorkDuration(d.first, d.second);
-                       });
+                       [](auto d) { return toWorkDuration(d); });
         EXPECT_CALL(*mMockSession.get(), reportActualWorkDuration(durations))
-                .Times(test.second ? 1 : 0);
-        sendActualWorkDurationGroup(durations, 80ms);
+                .Times(std::get<2>(test) ? 1 : 0);
+        sendActualWorkDurationGroup(durations, std::get<1>(test));
         verifyAndClearExpectations();
     }
 }
diff --git a/services/surfaceflinger/tests/unittests/SchedulerTest.cpp b/services/surfaceflinger/tests/unittests/SchedulerTest.cpp
index aab2795..93c809e 100644
--- a/services/surfaceflinger/tests/unittests/SchedulerTest.cpp
+++ b/services/surfaceflinger/tests/unittests/SchedulerTest.cpp
@@ -159,8 +159,8 @@
     mScheduler->recordLayerHistory(layer.get(), 0, LayerHistory::LayerUpdateType::Buffer);
     ASSERT_EQ(0u, mScheduler->getNumActiveLayers());
 
-    constexpr bool kPowerStateNormal = true;
-    mScheduler->setDisplayPowerState(kPowerStateNormal);
+    constexpr hal::PowerMode kPowerModeOn = hal::PowerMode::ON;
+    mScheduler->setDisplayPowerMode(kPowerModeOn);
 
     constexpr uint32_t kDisplayArea = 999'999;
     mScheduler->onActiveDisplayAreaChanged(kDisplayArea);
@@ -226,8 +226,8 @@
 
     mScheduler->recordLayerHistory(layer.get(), 0, LayerHistory::LayerUpdateType::Buffer);
 
-    constexpr bool kPowerStateNormal = true;
-    mScheduler->setDisplayPowerState(kPowerStateNormal);
+    constexpr hal::PowerMode kPowerModeOn = hal::PowerMode::ON;
+    mScheduler->setDisplayPowerMode(kPowerModeOn);
 
     constexpr uint32_t kDisplayArea = 999'999;
     mScheduler->onActiveDisplayAreaChanged(kDisplayArea);
diff --git a/services/surfaceflinger/tests/unittests/SurfaceFlinger_PowerHintTest.cpp b/services/surfaceflinger/tests/unittests/SurfaceFlinger_PowerHintTest.cpp
index 8de9e4b..c2d87f2 100644
--- a/services/surfaceflinger/tests/unittests/SurfaceFlinger_PowerHintTest.cpp
+++ b/services/surfaceflinger/tests/unittests/SurfaceFlinger_PowerHintTest.cpp
@@ -74,6 +74,7 @@
     mFlinger.setupRenderEngine(std::unique_ptr<renderengine::RenderEngine>(mRenderEngine));
     mFlinger.setupTimeStats(std::shared_ptr<TimeStats>(mTimeStats));
     mFlinger.setupComposer(std::unique_ptr<Hwc2::Composer>(mComposer));
+    mFlinger.setPowerHintSessionMode(true, true);
     mFlinger.setupPowerAdvisor(std::unique_ptr<Hwc2::PowerAdvisor>(mPowerAdvisor));
     static constexpr bool kIsPrimary = true;
     FakeHwcDisplayInjector(DEFAULT_DISPLAY_ID, hal::DisplayType::PHYSICAL, kIsPrimary)
@@ -142,10 +143,7 @@
                 std::this_thread::sleep_for(mockHwcRunTime);
                 return hardware::graphics::composer::V2_1::Error::NONE;
             });
-    EXPECT_CALL(*mPowerAdvisor,
-                sendActualWorkDuration(Gt(mockHwcRunTime.count()),
-                                       Gt(now + mockHwcRunTime.count())))
-            .Times(1);
+    EXPECT_CALL(*mPowerAdvisor, sendActualWorkDuration()).Times(1);
     static constexpr bool kVsyncId = 123; // arbitrary
     mFlinger.commitAndComposite(now, kVsyncId, now + mockVsyncPeriod.count());
 }
diff --git a/services/surfaceflinger/tests/unittests/TestableSurfaceFlinger.h b/services/surfaceflinger/tests/unittests/TestableSurfaceFlinger.h
index f1a69fb..b70fdcd 100644
--- a/services/surfaceflinger/tests/unittests/TestableSurfaceFlinger.h
+++ b/services/surfaceflinger/tests/unittests/TestableSurfaceFlinger.h
@@ -330,6 +330,10 @@
         layer->mDrawingParent = drawingParent;
     }
 
+    void setPowerHintSessionMode(bool early, bool late) {
+        mFlinger->mPowerHintSessionMode = {.late = late, .early = early};
+    }
+
     /* ------------------------------------------------------------------------
      * Forwarding for functions being tested
      */
diff --git a/services/surfaceflinger/tests/unittests/VSyncReactorTest.cpp b/services/surfaceflinger/tests/unittests/VSyncReactorTest.cpp
index 4eb9055..30a3f9a 100644
--- a/services/surfaceflinger/tests/unittests/VSyncReactorTest.cpp
+++ b/services/surfaceflinger/tests/unittests/VSyncReactorTest.cpp
@@ -349,6 +349,23 @@
     }
 }
 
+TEST_F(VSyncReactorTest, addHwVsyncTimestampDozePreempt) {
+    bool periodFlushed = false;
+    nsecs_t const newPeriod = 4000;
+
+    mReactor.startPeriodTransition(newPeriod);
+
+    auto time = 0;
+    // If the power mode is not DOZE or DOZE_SUSPEND, it is still collecting timestamps.
+    EXPECT_TRUE(mReactor.addHwVsyncTimestamp(time, std::nullopt, &periodFlushed));
+    EXPECT_FALSE(periodFlushed);
+
+    // Set power mode to DOZE to trigger period flushing.
+    mReactor.setDisplayPowerMode(hal::PowerMode::DOZE);
+    EXPECT_FALSE(mReactor.addHwVsyncTimestamp(time, std::nullopt, &periodFlushed));
+    EXPECT_TRUE(periodFlushed);
+}
+
 TEST_F(VSyncReactorTest, addPresentFenceWhileAwaitingPeriodConfirmationRequestsHwVsync) {
     auto time = 0;
     bool periodFlushed = false;
diff --git a/services/surfaceflinger/tests/unittests/mock/DisplayHardware/MockPowerAdvisor.h b/services/surfaceflinger/tests/unittests/mock/DisplayHardware/MockPowerAdvisor.h
index c598cbc..05cc544 100644
--- a/services/surfaceflinger/tests/unittests/mock/DisplayHardware/MockPowerAdvisor.h
+++ b/services/surfaceflinger/tests/unittests/mock/DisplayHardware/MockPowerAdvisor.h
@@ -36,11 +36,31 @@
     MOCK_METHOD(bool, usePowerHintSession, (), (override));
     MOCK_METHOD(bool, supportsPowerHintSession, (), (override));
     MOCK_METHOD(bool, isPowerHintSessionRunning, (), (override));
-    MOCK_METHOD(void, setTargetWorkDuration, (int64_t targetDurationNanos), (override));
-    MOCK_METHOD(void, sendActualWorkDuration, (int64_t actualDurationNanos, nsecs_t timestamp),
-                (override));
+    MOCK_METHOD(void, setTargetWorkDuration, (int64_t targetDuration), (override));
+    MOCK_METHOD(void, sendActualWorkDuration, (), (override));
+    MOCK_METHOD(void, sendPredictedWorkDuration, (), (override));
     MOCK_METHOD(void, enablePowerHint, (bool enabled), (override));
     MOCK_METHOD(bool, startPowerHintSession, (const std::vector<int32_t>& threadIds), (override));
+    MOCK_METHOD(void, setGpuFenceTime,
+                (DisplayId displayId, std::unique_ptr<FenceTime>&& fenceTime), (override));
+    MOCK_METHOD(void, setValidateTiming,
+                (DisplayId displayId, nsecs_t valiateStartTime, nsecs_t validateEndTime),
+                (override));
+    MOCK_METHOD(void, setPresentTiming,
+                (DisplayId displayId, nsecs_t presentStartTime, nsecs_t presentEndTime),
+                (override));
+    MOCK_METHOD(void, setSkippedValidate, (DisplayId displayId, bool skipped), (override));
+    MOCK_METHOD(void, setRequiresClientComposition,
+                (DisplayId displayId, bool requiresClientComposition), (override));
+    MOCK_METHOD(void, setExpectedPresentTime, (nsecs_t expectedPresentTime), (override));
+    MOCK_METHOD(void, setPresentDelayedTime,
+                (DisplayId displayId,
+                 std::chrono::steady_clock::time_point earliestFrameStartTime));
+    MOCK_METHOD(void, setFrameDelay, (nsecs_t frameDelayDuration), (override));
+    MOCK_METHOD(void, setCommitStart, (nsecs_t commitStartTime), (override));
+    MOCK_METHOD(void, setCompositeEnd, (nsecs_t compositeEndtime), (override));
+    MOCK_METHOD(void, setDisplays, (std::vector<DisplayId> & displayIds), (override));
+    MOCK_METHOD(void, setTotalFrameTargetWorkDuration, (int64_t targetDuration), (override));
 };
 
 } // namespace android::Hwc2::mock
diff --git a/services/surfaceflinger/tests/unittests/mock/MockVsyncController.h b/services/surfaceflinger/tests/unittests/mock/MockVsyncController.h
index 314f681..4ef91da 100644
--- a/services/surfaceflinger/tests/unittests/mock/MockVsyncController.h
+++ b/services/surfaceflinger/tests/unittests/mock/MockVsyncController.h
@@ -31,6 +31,7 @@
     MOCK_METHOD3(addHwVsyncTimestamp, bool(nsecs_t, std::optional<nsecs_t>, bool*));
     MOCK_METHOD1(startPeriodTransition, void(nsecs_t));
     MOCK_METHOD1(setIgnorePresentFences, void(bool));
+    MOCK_METHOD(void, setDisplayPowerMode, (hal::PowerMode), (override));
 
     MOCK_CONST_METHOD1(dump, void(std::string&));
 };