diff --git a/include/input/InputTransport.h b/include/input/InputTransport.h
index 5f9a37d..dbc7bfa 100644
--- a/include/input/InputTransport.h
+++ b/include/input/InputTransport.h
@@ -452,8 +452,11 @@
  */
 class InputConsumer {
 public:
-    /* Creates a consumer associated with an input channel. */
+    /* Create a consumer associated with an input channel. */
     explicit InputConsumer(const std::shared_ptr<InputChannel>& channel);
+    /* Create a consumer associated with an input channel, override resampling system property */
+    explicit InputConsumer(const std::shared_ptr<InputChannel>& channel,
+                           bool enableTouchResampling);
 
     /* Destroys the consumer and releases its input channel. */
     ~InputConsumer();
diff --git a/libs/gui/LayerState.cpp b/libs/gui/LayerState.cpp
index 0d3b412..7f0f638 100644
--- a/libs/gui/LayerState.cpp
+++ b/libs/gui/LayerState.cpp
@@ -102,6 +102,11 @@
     SAFE_PARCEL(output.writeUint32, transform);
     SAFE_PARCEL(output.writeBool, transformToDisplayInverse);
     SAFE_PARCEL(output.writeBool, borderEnabled);
+    SAFE_PARCEL(output.writeFloat, borderWidth);
+    SAFE_PARCEL(output.writeFloat, borderColor.r);
+    SAFE_PARCEL(output.writeFloat, borderColor.g);
+    SAFE_PARCEL(output.writeFloat, borderColor.b);
+    SAFE_PARCEL(output.writeFloat, borderColor.a);
     SAFE_PARCEL(output.writeUint32, static_cast<uint32_t>(dataspace));
     SAFE_PARCEL(output.write, hdrMetadata);
     SAFE_PARCEL(output.write, surfaceDamageRegion);
@@ -202,6 +207,16 @@
     SAFE_PARCEL(input.readUint32, &transform);
     SAFE_PARCEL(input.readBool, &transformToDisplayInverse);
     SAFE_PARCEL(input.readBool, &borderEnabled);
+    SAFE_PARCEL(input.readFloat, &tmpFloat);
+    borderWidth = tmpFloat;
+    SAFE_PARCEL(input.readFloat, &tmpFloat);
+    borderColor.r = tmpFloat;
+    SAFE_PARCEL(input.readFloat, &tmpFloat);
+    borderColor.g = tmpFloat;
+    SAFE_PARCEL(input.readFloat, &tmpFloat);
+    borderColor.b = tmpFloat;
+    SAFE_PARCEL(input.readFloat, &tmpFloat);
+    borderColor.a = tmpFloat;
 
     uint32_t tmpUint32 = 0;
     SAFE_PARCEL(input.readUint32, &tmpUint32);
@@ -555,6 +570,8 @@
     if (other.what & eRenderBorderChanged) {
         what |= eRenderBorderChanged;
         borderEnabled = other.borderEnabled;
+        borderWidth = other.borderWidth;
+        borderColor = other.borderColor;
     }
     if (other.what & eFrameRateSelectionPriority) {
         what |= eFrameRateSelectionPriority;
diff --git a/libs/gui/SurfaceComposerClient.cpp b/libs/gui/SurfaceComposerClient.cpp
index caab506..6fc813b 100644
--- a/libs/gui/SurfaceComposerClient.cpp
+++ b/libs/gui/SurfaceComposerClient.cpp
@@ -1940,7 +1940,7 @@
 }
 
 SurfaceComposerClient::Transaction& SurfaceComposerClient::Transaction::enableBorder(
-        const sp<SurfaceControl>& sc, bool shouldEnable) {
+        const sp<SurfaceControl>& sc, bool shouldEnable, float width, const half4& color) {
     layer_state_t* s = getLayerState(sc);
     if (!s) {
         mStatus = BAD_INDEX;
@@ -1949,6 +1949,8 @@
 
     s->what |= layer_state_t::eRenderBorderChanged;
     s->borderEnabled = shouldEnable;
+    s->borderWidth = width;
+    s->borderColor = color;
 
     registerSurfaceControlForCallback(sc);
     return *this;
diff --git a/libs/gui/include/gui/LayerState.h b/libs/gui/include/gui/LayerState.h
index 4621d2b..37a1595 100644
--- a/libs/gui/include/gui/LayerState.h
+++ b/libs/gui/include/gui/LayerState.h
@@ -295,6 +295,9 @@
 
     // Flag to indicate if border needs to be enabled on the layer
     bool borderEnabled;
+    float borderWidth;
+    half4 borderColor;
+
     // Stretch effect to be applied to this layer
     StretchEffect stretchEffect;
 
diff --git a/libs/gui/include/gui/SurfaceComposerClient.h b/libs/gui/include/gui/SurfaceComposerClient.h
index 2b24c3f..569dbf8 100644
--- a/libs/gui/include/gui/SurfaceComposerClient.h
+++ b/libs/gui/include/gui/SurfaceComposerClient.h
@@ -625,7 +625,8 @@
                                          const Rect& destinationFrame);
         Transaction& setDropInputMode(const sp<SurfaceControl>& sc, gui::DropInputMode mode);
 
-        Transaction& enableBorder(const sp<SurfaceControl>& sc, bool shouldEnable);
+        Transaction& enableBorder(const sp<SurfaceControl>& sc, bool shouldEnable, float width,
+                                  const half4& color);
 
         status_t setDisplaySurface(const sp<IBinder>& token,
                 const sp<IGraphicBufferProducer>& bufferProducer);
diff --git a/libs/input/InputTransport.cpp b/libs/input/InputTransport.cpp
index 6195052..8d8433b 100644
--- a/libs/input/InputTransport.cpp
+++ b/libs/input/InputTransport.cpp
@@ -51,7 +51,7 @@
 
 // Latency added during resampling.  A few milliseconds doesn't hurt much but
 // reduces the impact of mispredicted touch positions.
-static const nsecs_t RESAMPLE_LATENCY = 5 * NANOS_PER_MS;
+const std::chrono::duration RESAMPLE_LATENCY = 5ms;
 
 // Minimum time difference between consecutive samples before attempting to resample.
 static const nsecs_t RESAMPLE_MIN_DELTA = 2 * NANOS_PER_MS;
@@ -721,7 +721,11 @@
 // --- InputConsumer ---
 
 InputConsumer::InputConsumer(const std::shared_ptr<InputChannel>& channel)
-      : mResampleTouch(isTouchResamplingEnabled()), mChannel(channel), mMsgDeferred(false) {}
+      : InputConsumer(channel, isTouchResamplingEnabled()) {}
+
+InputConsumer::InputConsumer(const std::shared_ptr<InputChannel>& channel,
+                             bool enableTouchResampling)
+      : mResampleTouch(enableTouchResampling), mChannel(channel), mMsgDeferred(false) {}
 
 InputConsumer::~InputConsumer() {
 }
@@ -751,7 +755,10 @@
             // Receive a fresh message.
             status_t result = mChannel->receiveMessage(&mMsg);
             if (result == OK) {
-                mConsumeTimes.emplace(mMsg.header.seq, systemTime(SYSTEM_TIME_MONOTONIC));
+                const auto [_, inserted] =
+                        mConsumeTimes.emplace(mMsg.header.seq, systemTime(SYSTEM_TIME_MONOTONIC));
+                LOG_ALWAYS_FATAL_IF(!inserted, "Already have a consume time for seq=%" PRIu32,
+                                    mMsg.header.seq);
             }
             if (result) {
                 // Consume the next batched event unless batches are being held for later.
@@ -918,7 +925,7 @@
 
         nsecs_t sampleTime = frameTime;
         if (mResampleTouch) {
-            sampleTime -= RESAMPLE_LATENCY;
+            sampleTime -= std::chrono::nanoseconds(RESAMPLE_LATENCY).count();
         }
         ssize_t split = findSampleNoLaterThan(batch, sampleTime);
         if (split < 0) {
@@ -1166,6 +1173,11 @@
         return;
     }
 
+    if (current->eventTime == sampleTime) {
+        // Prevents having 2 events with identical times and coordinates.
+        return;
+    }
+
     // Resample touch coordinates.
     History oldLastResample;
     oldLastResample.initializeFrom(touchState.lastResample);
diff --git a/libs/input/tests/Android.bp b/libs/input/tests/Android.bp
index 6ffe851..54f7586 100644
--- a/libs/input/tests/Android.bp
+++ b/libs/input/tests/Android.bp
@@ -16,6 +16,7 @@
         "InputDevice_test.cpp",
         "InputEvent_test.cpp",
         "InputPublisherAndConsumer_test.cpp",
+        "TouchResampling_test.cpp",
         "TouchVideoFrame_test.cpp",
         "VelocityTracker_test.cpp",
         "VerifiedInputEvent_test.cpp",
diff --git a/libs/input/tests/InputPublisherAndConsumer_test.cpp b/libs/input/tests/InputPublisherAndConsumer_test.cpp
index 05bc0bc..70e4fda 100644
--- a/libs/input/tests/InputPublisherAndConsumer_test.cpp
+++ b/libs/input/tests/InputPublisherAndConsumer_test.cpp
@@ -16,17 +16,10 @@
 
 #include "TestHelpers.h"
 
-#include <unistd.h>
-#include <sys/mman.h>
-#include <time.h>
-
 #include <attestation/HmacKeyManager.h>
-#include <cutils/ashmem.h>
 #include <gtest/gtest.h>
 #include <gui/constants.h>
 #include <input/InputTransport.h>
-#include <utils/StopWatch.h>
-#include <utils/Timers.h>
 
 using android::base::Result;
 
diff --git a/libs/input/tests/TouchResampling_test.cpp b/libs/input/tests/TouchResampling_test.cpp
new file mode 100644
index 0000000..c09a8e9
--- /dev/null
+++ b/libs/input/tests/TouchResampling_test.cpp
@@ -0,0 +1,562 @@
+/*
+ * Copyright (C) 2022 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "TestHelpers.h"
+
+#include <chrono>
+#include <vector>
+
+#include <attestation/HmacKeyManager.h>
+#include <gtest/gtest.h>
+#include <input/InputTransport.h>
+
+using namespace std::chrono_literals;
+
+namespace android {
+
+struct Pointer {
+    int32_t id;
+    float x;
+    float y;
+};
+
+struct InputEventEntry {
+    std::chrono::nanoseconds eventTime;
+    std::vector<Pointer> pointers;
+    int32_t action;
+};
+
+class TouchResamplingTest : public testing::Test {
+protected:
+    std::unique_ptr<InputPublisher> mPublisher;
+    std::unique_ptr<InputConsumer> mConsumer;
+    PreallocatedInputEventFactory mEventFactory;
+
+    uint32_t mSeq = 1;
+
+    void SetUp() override {
+        std::unique_ptr<InputChannel> serverChannel, clientChannel;
+        status_t result =
+                InputChannel::openInputChannelPair("channel name", serverChannel, clientChannel);
+        ASSERT_EQ(OK, result);
+
+        mPublisher = std::make_unique<InputPublisher>(std::move(serverChannel));
+        mConsumer = std::make_unique<InputConsumer>(std::move(clientChannel),
+                                                    true /* enableTouchResampling */);
+    }
+
+    status_t publishSimpleMotionEventWithCoords(int32_t action, nsecs_t eventTime,
+                                                const std::vector<PointerProperties>& properties,
+                                                const std::vector<PointerCoords>& coords);
+    void publishSimpleMotionEvent(int32_t action, nsecs_t eventTime,
+                                  const std::vector<Pointer>& pointers);
+    void publishInputEventEntries(const std::vector<InputEventEntry>& entries);
+    void consumeInputEventEntries(const std::vector<InputEventEntry>& entries,
+                                  std::chrono::nanoseconds frameTime);
+    void receiveResponseUntilSequence(uint32_t seq);
+};
+
+status_t TouchResamplingTest::publishSimpleMotionEventWithCoords(
+        int32_t action, nsecs_t eventTime, const std::vector<PointerProperties>& properties,
+        const std::vector<PointerCoords>& coords) {
+    const ui::Transform identityTransform;
+    const nsecs_t downTime = 0;
+
+    if (action == AMOTION_EVENT_ACTION_DOWN && eventTime != 0) {
+        ADD_FAILURE() << "Downtime should be equal to 0 (hardcoded for convenience)";
+    }
+    return mPublisher->publishMotionEvent(mSeq++, InputEvent::nextId(), 1 /*deviceId*/,
+                                          AINPUT_SOURCE_TOUCHSCREEN, 0 /*displayId*/, INVALID_HMAC,
+                                          action, 0 /*actionButton*/, 0 /*flags*/, 0 /*edgeFlags*/,
+                                          AMETA_NONE, 0 /*buttonState*/, MotionClassification::NONE,
+                                          identityTransform, 0 /*xPrecision*/, 0 /*yPrecision*/,
+                                          AMOTION_EVENT_INVALID_CURSOR_POSITION,
+                                          AMOTION_EVENT_INVALID_CURSOR_POSITION, identityTransform,
+                                          downTime, eventTime, properties.size(), properties.data(),
+                                          coords.data());
+}
+
+void TouchResamplingTest::publishSimpleMotionEvent(int32_t action, nsecs_t eventTime,
+                                                   const std::vector<Pointer>& pointers) {
+    std::vector<PointerProperties> properties;
+    std::vector<PointerCoords> coords;
+
+    for (const Pointer& pointer : pointers) {
+        properties.push_back({});
+        properties.back().clear();
+        properties.back().id = pointer.id;
+        properties.back().toolType = AMOTION_EVENT_TOOL_TYPE_FINGER;
+
+        coords.push_back({});
+        coords.back().clear();
+        coords.back().setAxisValue(AMOTION_EVENT_AXIS_X, pointer.x);
+        coords.back().setAxisValue(AMOTION_EVENT_AXIS_Y, pointer.y);
+    }
+
+    status_t result = publishSimpleMotionEventWithCoords(action, eventTime, properties, coords);
+    ASSERT_EQ(OK, result);
+}
+
+/**
+ * Each entry is published separately, one entry at a time. As a result, action is used here
+ * on a per-entry basis.
+ */
+void TouchResamplingTest::publishInputEventEntries(const std::vector<InputEventEntry>& entries) {
+    for (const InputEventEntry& entry : entries) {
+        publishSimpleMotionEvent(entry.action, entry.eventTime.count(), entry.pointers);
+    }
+}
+
+/**
+ * Inside the publisher, read responses repeatedly until the desired sequence number is returned.
+ *
+ * Sometimes, when you call 'sendFinishedSignal', you would be finishing a batch which is comprised
+ * of several input events. As a result, consumer will generate multiple 'finish' signals on your
+ * behalf.
+ *
+ * In this function, we call 'receiveConsumerResponse' in a loop until the desired sequence number
+ * is returned.
+ */
+void TouchResamplingTest::receiveResponseUntilSequence(uint32_t seq) {
+    size_t consumedEvents = 0;
+    while (consumedEvents < 100) {
+        android::base::Result<InputPublisher::ConsumerResponse> response =
+                mPublisher->receiveConsumerResponse();
+        ASSERT_TRUE(response.ok());
+        ASSERT_TRUE(std::holds_alternative<InputPublisher::Finished>(*response));
+        const InputPublisher::Finished& finish = std::get<InputPublisher::Finished>(*response);
+        ASSERT_TRUE(finish.handled)
+                << "publisher receiveFinishedSignal should have set handled to consumer's reply";
+        if (finish.seq == seq) {
+            return;
+        }
+        consumedEvents++;
+    }
+    FAIL() << "Got " << consumedEvents << "events, but still no event with seq=" << seq;
+}
+
+/**
+ * All entries are compared against a single MotionEvent, but the same data structure
+ * InputEventEntry is used here for simpler code. As a result, the entire array of InputEventEntry
+ * must contain identical values for the action field.
+ */
+void TouchResamplingTest::consumeInputEventEntries(const std::vector<InputEventEntry>& entries,
+                                                   std::chrono::nanoseconds frameTime) {
+    ASSERT_GE(entries.size(), 1U) << "Must have at least 1 InputEventEntry to compare against";
+
+    uint32_t consumeSeq;
+    InputEvent* event;
+
+    status_t status = mConsumer->consume(&mEventFactory, true /*consumeBatches*/, frameTime.count(),
+                                         &consumeSeq, &event);
+    ASSERT_EQ(OK, status);
+    MotionEvent* motionEvent = static_cast<MotionEvent*>(event);
+
+    ASSERT_EQ(entries.size() - 1, motionEvent->getHistorySize());
+    for (size_t i = 0; i < entries.size(); i++) { // most recent sample is last
+        SCOPED_TRACE(i);
+        const InputEventEntry& entry = entries[i];
+        ASSERT_EQ(entry.action, motionEvent->getAction());
+        ASSERT_EQ(entry.eventTime.count(), motionEvent->getHistoricalEventTime(i));
+        ASSERT_EQ(entry.pointers.size(), motionEvent->getPointerCount());
+
+        for (size_t p = 0; p < motionEvent->getPointerCount(); p++) {
+            SCOPED_TRACE(p);
+            // The pointers can be in any order, both in MotionEvent as well as InputEventEntry
+            ssize_t motionEventPointerIndex = motionEvent->findPointerIndex(entry.pointers[p].id);
+            ASSERT_GE(motionEventPointerIndex, 0) << "Pointer must be present in MotionEvent";
+            ASSERT_EQ(entry.pointers[p].x,
+                      motionEvent->getHistoricalAxisValue(AMOTION_EVENT_AXIS_X,
+                                                          motionEventPointerIndex, i));
+            ASSERT_EQ(entry.pointers[p].x,
+                      motionEvent->getHistoricalRawAxisValue(AMOTION_EVENT_AXIS_X,
+                                                             motionEventPointerIndex, i));
+            ASSERT_EQ(entry.pointers[p].y,
+                      motionEvent->getHistoricalAxisValue(AMOTION_EVENT_AXIS_Y,
+                                                          motionEventPointerIndex, i));
+            ASSERT_EQ(entry.pointers[p].y,
+                      motionEvent->getHistoricalRawAxisValue(AMOTION_EVENT_AXIS_Y,
+                                                             motionEventPointerIndex, i));
+        }
+    }
+
+    status = mConsumer->sendFinishedSignal(consumeSeq, true);
+    ASSERT_EQ(OK, status);
+
+    receiveResponseUntilSequence(consumeSeq);
+}
+
+/**
+ * Timeline
+ * ---------+------------------+------------------+--------+-----------------+----------------------
+ *          0 ms               10 ms              20 ms    25 ms            35 ms
+ *          ACTION_DOWN       ACTION_MOVE      ACTION_MOVE  ^                ^
+ *                                                          |                |
+ *                                                         resampled value   |
+ *                                                                          frameTime
+ * Typically, the prediction is made for time frameTime - RESAMPLE_LATENCY, or 30 ms in this case
+ * However, that would be 10 ms later than the last real sample (which came in at 20 ms).
+ * Therefore, the resampling should happen at 20 ms + RESAMPLE_MAX_PREDICTION = 28 ms.
+ * In this situation, though, resample time is further limited by taking half of the difference
+ * between the last two real events, which would put this time at:
+ * 20 ms + (20 ms - 10 ms) / 2 = 25 ms.
+ */
+TEST_F(TouchResamplingTest, EventIsResampled) {
+    std::chrono::nanoseconds frameTime;
+    std::vector<InputEventEntry> entries, expectedEntries;
+
+    // Initial ACTION_DOWN should be separate, because the first consume event will only return
+    // InputEvent with a single action.
+    entries = {
+            //      id  x   y
+            {0ms, {{0, 10, 20}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 5ms;
+    expectedEntries = {
+            //      id  x   y
+            {0ms, {{0, 10, 20}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // Two ACTION_MOVE events 10 ms apart that move in X direction and stay still in Y
+    entries = {
+            //      id  x   y
+            {10ms, {{0, 20, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {20ms, {{0, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 35ms;
+    expectedEntries = {
+            //      id  x   y
+            {10ms, {{0, 20, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {20ms, {{0, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {25ms, {{0, 35, 30}}, AMOTION_EVENT_ACTION_MOVE}, // resampled value
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+}
+
+/**
+ * Same as above test, but use pointer id=1 instead of 0 to make sure that system does not
+ * have these hardcoded.
+ */
+TEST_F(TouchResamplingTest, EventIsResampledWithDifferentId) {
+    std::chrono::nanoseconds frameTime;
+    std::vector<InputEventEntry> entries, expectedEntries;
+
+    // Initial ACTION_DOWN should be separate, because the first consume event will only return
+    // InputEvent with a single action.
+    entries = {
+            //      id  x   y
+            {0ms, {{1, 10, 20}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 5ms;
+    expectedEntries = {
+            //      id  x   y
+            {0ms, {{1, 10, 20}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // Two ACTION_MOVE events 10 ms apart that move in X direction and stay still in Y
+    entries = {
+            //      id  x   y
+            {10ms, {{1, 20, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {20ms, {{1, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 35ms;
+    expectedEntries = {
+            //      id  x   y
+            {10ms, {{1, 20, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {20ms, {{1, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {25ms, {{1, 35, 30}}, AMOTION_EVENT_ACTION_MOVE}, // resampled value
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+}
+
+/**
+ * Event should not be resampled when sample time is equal to event time.
+ */
+TEST_F(TouchResamplingTest, SampleTimeEqualsEventTime) {
+    std::chrono::nanoseconds frameTime;
+    std::vector<InputEventEntry> entries, expectedEntries;
+
+    // Initial ACTION_DOWN should be separate, because the first consume event will only return
+    // InputEvent with a single action.
+    entries = {
+            //      id  x   y
+            {0ms, {{0, 10, 20}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 5ms;
+    expectedEntries = {
+            //      id  x   y
+            {0ms, {{0, 10, 20}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // Two ACTION_MOVE events 10 ms apart that move in X direction and stay still in Y
+    entries = {
+            //      id  x   y
+            {10ms, {{0, 20, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {20ms, {{0, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 20ms + 5ms /*RESAMPLE_LATENCY*/;
+    expectedEntries = {
+            //      id  x   y
+            {10ms, {{0, 20, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {20ms, {{0, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            // no resampled event because the time of resample falls exactly on the existing event
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+}
+
+/**
+ * Once we send a resampled value to the app, we should continue to "lie" if the pointer
+ * does not move. So, if the pointer keeps the same coordinates, resampled value should continue
+ * to be used.
+ */
+TEST_F(TouchResamplingTest, ResampledValueIsUsedForIdenticalCoordinates) {
+    std::chrono::nanoseconds frameTime;
+    std::vector<InputEventEntry> entries, expectedEntries;
+
+    // Initial ACTION_DOWN should be separate, because the first consume event will only return
+    // InputEvent with a single action.
+    entries = {
+            //      id  x   y
+            {0ms, {{0, 10, 20}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 5ms;
+    expectedEntries = {
+            //      id  x   y
+            {0ms, {{0, 10, 20}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // Two ACTION_MOVE events 10 ms apart that move in X direction and stay still in Y
+    entries = {
+            //      id  x   y
+            {10ms, {{0, 20, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {20ms, {{0, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 35ms;
+    expectedEntries = {
+            //      id  x   y
+            {10ms, {{0, 20, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {20ms, {{0, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {25ms, {{0, 35, 30}}, AMOTION_EVENT_ACTION_MOVE}, // resampled value
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // Coordinate value 30 has been resampled to 35. When a new event comes in with value 30 again,
+    // the system should still report 35.
+    entries = {
+            //      id  x   y
+            {40ms, {{0, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 45ms + 5ms /*RESAMPLE_LATENCY*/;
+    expectedEntries = {
+            //      id  x   y
+            {40ms, {{0, 35, 30}}, AMOTION_EVENT_ACTION_MOVE}, // original event, rewritten
+            {45ms, {{0, 35, 30}}, AMOTION_EVENT_ACTION_MOVE}, // resampled event, rewritten
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+}
+
+TEST_F(TouchResamplingTest, OldEventReceivedAfterResampleOccurs) {
+    std::chrono::nanoseconds frameTime;
+    std::vector<InputEventEntry> entries, expectedEntries;
+
+    // Initial ACTION_DOWN should be separate, because the first consume event will only return
+    // InputEvent with a single action.
+    entries = {
+            //      id  x   y
+            {0ms, {{0, 10, 20}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 5ms;
+    expectedEntries = {
+            //      id  x   y
+            {0ms, {{0, 10, 20}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // Two ACTION_MOVE events 10 ms apart that move in X direction and stay still in Y
+    entries = {
+            //      id  x   y
+            {10ms, {{0, 20, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {20ms, {{0, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 35ms;
+    expectedEntries = {
+            //      id  x   y
+            {10ms, {{0, 20, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {20ms, {{0, 30, 30}}, AMOTION_EVENT_ACTION_MOVE},
+            {25ms, {{0, 35, 30}}, AMOTION_EVENT_ACTION_MOVE}, // resampled value
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+    // Above, the resampled event is at 25ms rather than at 30 ms = 35ms - RESAMPLE_LATENCY
+    // because we are further bound by how far we can extrapolate by the "last time delta".
+    // That's 50% of (20 ms - 10ms) => 5ms. So we can't predict more than 5 ms into the future
+    // from the event at 20ms, which is why the resampled event is at t = 25 ms.
+
+    // We resampled the event to 25 ms. Now, an older 'real' event comes in.
+    entries = {
+            //      id  x   y
+            {24ms, {{0, 40, 30}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 50ms;
+    expectedEntries = {
+            //      id  x   y
+            {24ms, {{0, 35, 30}}, AMOTION_EVENT_ACTION_MOVE}, // original event, rewritten
+            {26ms, {{0, 45, 30}}, AMOTION_EVENT_ACTION_MOVE}, // resampled event, rewritten
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+}
+
+TEST_F(TouchResamplingTest, TwoPointersAreResampledIndependently) {
+    std::chrono::nanoseconds frameTime;
+    std::vector<InputEventEntry> entries, expectedEntries;
+
+    // full action for when a pointer with id=1 appears (some other pointer must already be present)
+    constexpr int32_t actionPointer1Down =
+            AMOTION_EVENT_ACTION_POINTER_DOWN + (1 << AMOTION_EVENT_ACTION_POINTER_INDEX_SHIFT);
+
+    // full action for when a pointer with id=0 disappears (some other pointer must still remain)
+    constexpr int32_t actionPointer0Up =
+            AMOTION_EVENT_ACTION_POINTER_UP + (0 << AMOTION_EVENT_ACTION_POINTER_INDEX_SHIFT);
+
+    // Initial ACTION_DOWN should be separate, because the first consume event will only return
+    // InputEvent with a single action.
+    entries = {
+            //      id  x   y
+            {0ms, {{0, 100, 100}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 5ms;
+    expectedEntries = {
+            //      id  x   y
+            {0ms, {{0, 100, 100}}, AMOTION_EVENT_ACTION_DOWN},
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    entries = {
+            //       id  x   y
+            {10ms, {{0, 100, 100}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 10ms + 5ms /*RESAMPLE_LATENCY*/;
+    expectedEntries = {
+            //       id  x   y
+            {10ms, {{0, 100, 100}}, AMOTION_EVENT_ACTION_MOVE},
+            // no resampled value because frameTime - RESAMPLE_LATENCY == eventTime
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // Second pointer id=1 appears
+    entries = {
+            //      id  x    y
+            {15ms, {{0, 100, 100}, {1, 500, 500}}, actionPointer1Down},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 20ms + 5ms /*RESAMPLE_LATENCY*/;
+    expectedEntries = {
+            //      id  x    y
+            {15ms, {{0, 100, 100}, {1, 500, 500}}, actionPointer1Down},
+            // no resampled value because frameTime - RESAMPLE_LATENCY == eventTime
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // Both pointers move
+    entries = {
+            //      id  x    y
+            {30ms, {{0, 100, 100}, {1, 500, 500}}, AMOTION_EVENT_ACTION_MOVE},
+            {40ms, {{0, 120, 120}, {1, 600, 600}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 45ms + 5ms /*RESAMPLE_LATENCY*/;
+    expectedEntries = {
+            //      id  x    y
+            {30ms, {{0, 100, 100}, {1, 500, 500}}, AMOTION_EVENT_ACTION_MOVE},
+            {40ms, {{0, 120, 120}, {1, 600, 600}}, AMOTION_EVENT_ACTION_MOVE},
+            {45ms, {{0, 130, 130}, {1, 650, 650}}, AMOTION_EVENT_ACTION_MOVE}, // resampled value
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // Both pointers move again
+    entries = {
+            //      id  x    y
+            {60ms, {{0, 120, 120}, {1, 600, 600}}, AMOTION_EVENT_ACTION_MOVE},
+            {70ms, {{0, 130, 130}, {1, 700, 700}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 75ms + 5ms /*RESAMPLE_LATENCY*/;
+    /**
+     * The sample at t = 60, pointer id 0 is not equal to 120, because this value of 120 was
+     * received twice, and resampled to 130. So if we already reported it as "130", we continue
+     * to report it as such. Similar with pointer id 1.
+     */
+    expectedEntries = {
+            {60ms,
+             {{0, 130, 130}, // not 120! because it matches previous real event
+              {1, 650, 650}},
+             AMOTION_EVENT_ACTION_MOVE},
+            {70ms, {{0, 130, 130}, {1, 700, 700}}, AMOTION_EVENT_ACTION_MOVE},
+            {75ms, {{0, 135, 135}, {1, 750, 750}}, AMOTION_EVENT_ACTION_MOVE}, // resampled value
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // First pointer id=0 leaves the screen
+    entries = {
+            //      id  x    y
+            {80ms, {{1, 600, 600}}, actionPointer0Up},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 90ms;
+    expectedEntries = {
+            //      id  x    y
+            {80ms, {{1, 600, 600}}, actionPointer0Up},
+            // no resampled event for ACTION_POINTER_UP
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+
+    // Remaining pointer id=1 is still present, but doesn't move
+    entries = {
+            //      id  x    y
+            {90ms, {{1, 600, 600}}, AMOTION_EVENT_ACTION_MOVE},
+    };
+    publishInputEventEntries(entries);
+    frameTime = 100ms;
+    expectedEntries = {
+            //      id  x    y
+            {90ms, {{1, 600, 600}}, AMOTION_EVENT_ACTION_MOVE},
+            /**
+             * The latest event with ACTION_MOVE was at t = 70, coord = 700.
+             * Use that value for resampling here: (600 - 700) / (90 - 70) * 5 + 600
+             */
+            {95ms, {{1, 575, 575}}, AMOTION_EVENT_ACTION_MOVE}, // resampled value
+    };
+    consumeInputEventEntries(expectedEntries, frameTime);
+}
+
+} // namespace android
diff --git a/libs/renderengine/include/renderengine/BorderRenderInfo.h b/libs/renderengine/include/renderengine/BorderRenderInfo.h
index 85d55fc..0ee6661 100644
--- a/libs/renderengine/include/renderengine/BorderRenderInfo.h
+++ b/libs/renderengine/include/renderengine/BorderRenderInfo.h
@@ -22,10 +22,13 @@
 namespace renderengine {
 
 struct BorderRenderInfo {
+    float width = 0;
+    half4 color;
     Region combinedRegion;
 
     bool operator==(const BorderRenderInfo& rhs) const {
-        return (combinedRegion.hasSameRects(rhs.combinedRegion));
+        return (width == rhs.width && color == rhs.color &&
+                combinedRegion.hasSameRects(rhs.combinedRegion));
     }
 };
 
diff --git a/libs/renderengine/skia/SkiaGLRenderEngine.cpp b/libs/renderengine/skia/SkiaGLRenderEngine.cpp
index ec9ad54..c9f9ec3 100644
--- a/libs/renderengine/skia/SkiaGLRenderEngine.cpp
+++ b/libs/renderengine/skia/SkiaGLRenderEngine.cpp
@@ -1247,15 +1247,11 @@
     }
     for (const auto& borderRenderInfo : display.borderInfoList) {
         SkPaint p;
-        // TODO (b/225977175): Use specified color
-        p.setColor(SkColor4f{.fR = 255 / 255.0f,
-                             .fG = 128 / 255.0f,
-                             .fB = 0 / 255.0f,
-                             .fA = 255 / 255.0f});
+        p.setColor(SkColor4f{borderRenderInfo.color.r, borderRenderInfo.color.g,
+                             borderRenderInfo.color.b, borderRenderInfo.color.a});
         p.setAntiAlias(true);
         p.setStyle(SkPaint::kStroke_Style);
-        // TODO (b/225977175): Use specified width
-        p.setStrokeWidth(20);
+        p.setStrokeWidth(borderRenderInfo.width);
         SkRegion sk_region;
         SkPath path;
 
diff --git a/libs/renderengine/tests/RenderEngineTest.cpp b/libs/renderengine/tests/RenderEngineTest.cpp
index df1b985..61af698 100644
--- a/libs/renderengine/tests/RenderEngineTest.cpp
+++ b/libs/renderengine/tests/RenderEngineTest.cpp
@@ -2433,6 +2433,8 @@
     display.borderInfoList.clear();
     renderengine::BorderRenderInfo info;
     info.combinedRegion = Region(Rect(99, 99, 199, 199));
+    info.width = 20.0f;
+    info.color = half4{1.0f, 128.0f / 255.0f, 0.0f, 1.0f};
     display.borderInfoList.emplace_back(info);
 
     const auto greenBuffer = allocateAndFillSourceBuffer(1, 1, ubyte4(0, 255, 0, 255));
diff --git a/services/sensorservice/BatteryService.cpp b/services/sensorservice/BatteryService.cpp
index 14f9a12..94de55c 100644
--- a/services/sensorservice/BatteryService.cpp
+++ b/services/sensorservice/BatteryService.cpp
@@ -74,23 +74,6 @@
     }
 }
 
-void BatteryService::cleanupImpl(uid_t uid) {
-    if (checkService()) {
-        Mutex::Autolock _l(mActivationsLock);
-        int64_t identity = IPCThreadState::self()->clearCallingIdentity();
-        for (size_t i=0 ; i<mActivations.size() ; ) {
-            const Info& info(mActivations[i]);
-            if (info.uid == uid) {
-                mBatteryStatService->noteStopSensor(info.uid, info.handle);
-                mActivations.removeAt(i);
-            } else {
-              i++;
-            }
-        }
-        IPCThreadState::self()->restoreCallingIdentity(identity);
-    }
-}
-
 bool BatteryService::checkService() {
     if (mBatteryStatService == nullptr) {
         const sp<IServiceManager> sm(defaultServiceManager());
diff --git a/services/sensorservice/BatteryService.h b/services/sensorservice/BatteryService.h
index 09eb2c1..13fc58a 100644
--- a/services/sensorservice/BatteryService.h
+++ b/services/sensorservice/BatteryService.h
@@ -32,7 +32,6 @@
 
     void enableSensorImpl(uid_t uid, int handle);
     void disableSensorImpl(uid_t uid, int handle);
-    void cleanupImpl(uid_t uid);
 
     struct Info {
         uid_t uid;
@@ -58,9 +57,6 @@
     static void disableSensor(uid_t uid, int handle) {
         BatteryService::getInstance().disableSensorImpl(uid, handle);
     }
-    static void cleanup(uid_t uid) {
-        BatteryService::getInstance().cleanupImpl(uid);
-    }
 };
 
 // ---------------------------------------------------------------------------
diff --git a/services/sensorservice/SensorDevice.cpp b/services/sensorservice/SensorDevice.cpp
index 53a3025..de050e0 100644
--- a/services/sensorservice/SensorDevice.cpp
+++ b/services/sensorservice/SensorDevice.cpp
@@ -612,8 +612,10 @@
     return SENSORS_DEVICE_API_VERSION_1_4;
 }
 
-status_t SensorDevice::flush(void* /*ident*/, int handle) {
+status_t SensorDevice::flush(void* ident, int handle) {
     if (mHalWrapper == nullptr) return NO_INIT;
+    if (isClientDisabled(ident)) return INVALID_OPERATION;
+    ALOGD_IF(DEBUG_CONNECTIONS, "\t>>> actuating h/w flush %d", handle);
     return mHalWrapper->flush(handle);
 }
 
@@ -754,6 +756,13 @@
 
 status_t SensorDevice::injectSensorData(const sensors_event_t* injected_sensor_event) {
     if (mHalWrapper == nullptr) return NO_INIT;
+    ALOGD_IF(DEBUG_CONNECTIONS,
+             "sensor_event handle=%d ts=%" PRId64 " data=%.2f, %.2f, %.2f %.2f %.2f %.2f",
+             injected_sensor_event->sensor, injected_sensor_event->timestamp,
+             injected_sensor_event->data[0], injected_sensor_event->data[1],
+             injected_sensor_event->data[2], injected_sensor_event->data[3],
+             injected_sensor_event->data[4], injected_sensor_event->data[5]);
+
     return mHalWrapper->injectSensorData(injected_sensor_event);
 }
 
diff --git a/services/sensorservice/SensorService.cpp b/services/sensorservice/SensorService.cpp
index 8b81d48..88cf5ab 100644
--- a/services/sensorservice/SensorService.cpp
+++ b/services/sensorservice/SensorService.cpp
@@ -1611,7 +1611,9 @@
             } else {
                 ALOGE("sensor interface of handle=0x%08x is null!", handle);
             }
-            c->removeSensor(handle);
+            if (c->removeSensor(handle)) {
+                BatteryService::disableSensor(c->getUid(), handle);
+            }
         }
         SensorRecord* rec = mActiveSensors.valueAt(i);
         ALOGE_IF(!rec, "mActiveSensors[%zu] is null (handle=0x%08x)!", i, handle);
@@ -1631,7 +1633,6 @@
     }
     c->updateLooperRegistration(mLooper);
     mConnectionHolder.removeEventConnection(connection);
-    BatteryService::cleanup(c->getUid());
     if (c->needsWakeLock()) {
         checkWakeLockStateLocked(&connLock);
     }
diff --git a/services/surfaceflinger/CompositionEngine/include/compositionengine/CompositionRefreshArgs.h b/services/surfaceflinger/CompositionEngine/include/compositionengine/CompositionRefreshArgs.h
index 8039bba..f861fc9 100644
--- a/services/surfaceflinger/CompositionEngine/include/compositionengine/CompositionRefreshArgs.h
+++ b/services/surfaceflinger/CompositionEngine/include/compositionengine/CompositionRefreshArgs.h
@@ -33,6 +33,8 @@
 using Outputs = std::vector<std::shared_ptr<compositionengine::Output>>;
 
 struct BorderRenderInfo {
+    float width = 0;
+    half4 color;
     std::vector<int32_t> layerIds;
 };
 /**
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 3ad6e52..fc8dd8b 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 430d673..feca71c 100644
--- a/services/surfaceflinger/CompositionEngine/src/Output.cpp
+++ b/services/surfaceflinger/CompositionEngine/src/Output.cpp
@@ -753,8 +753,13 @@
         for (const auto& id : borderInfo.layerIds) {
             info.combinedRegion.orSelf(*(layerVisibleRegionMap[id]));
         }
-        outputCompositionState.borderInfoList.emplace_back(std::move(info));
-        clientComposeTopLayer |= !info.combinedRegion.isEmpty();
+
+        if (!info.combinedRegion.isEmpty()) {
+            info.width = borderInfo.width;
+            info.color = borderInfo.color;
+            outputCompositionState.borderInfoList.emplace_back(std::move(info));
+            clientComposeTopLayer = true;
+        }
     }
 
     // In this situation we must client compose the top layer instead of using hwc
@@ -1120,6 +1125,10 @@
         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));
 }
@@ -1218,6 +1227,8 @@
     clientCompositionDisplay.colorTransform = outputState.colorTransformMatrix;
     for (auto& info : outputState.borderInfoList) {
         renderengine::BorderRenderInfo borderInfo;
+        borderInfo.width = info.width;
+        borderInfo.color = info.color;
         borderInfo.combinedRegion = info.combinedRegion;
         clientCompositionDisplay.borderInfoList.emplace_back(std::move(borderInfo));
     }
@@ -1428,6 +1439,14 @@
     // 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 ff2aa15..9b2b91c 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 862ab1d..ef6487c 100644
--- a/services/surfaceflinger/CompositionEngine/tests/OutputTest.cpp
+++ b/services/surfaceflinger/CompositionEngine/tests/OutputTest.cpp
@@ -3304,6 +3304,9 @@
         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/DisplayHardware/HWComposer.cpp b/services/surfaceflinger/DisplayHardware/HWComposer.cpp
index 79e4c75..18678e6 100644
--- a/services/surfaceflinger/DisplayHardware/HWComposer.cpp
+++ b/services/surfaceflinger/DisplayHardware/HWComposer.cpp
@@ -739,6 +739,13 @@
             });
 }
 
+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 7dc10ea..4ff4731 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 659efd8..20e38d0 100644
--- a/services/surfaceflinger/DisplayHardware/PowerAdvisor.cpp
+++ b/services/surfaceflinger/DisplayHardware/PowerAdvisor.cpp
@@ -174,7 +174,7 @@
     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;
@@ -183,26 +183,44 @@
         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;
 }
@@ -222,6 +240,282 @@
     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)) {}
@@ -303,6 +597,10 @@
     }
 
     mSupportsPowerHint = checkPowerHintSessionSupported();
+
+    mAllowedActualDeviation =
+            base::GetIntProperty<nsecs_t>("debug.sf.allowed_actual_deviation",
+                                          std::chrono::nanoseconds(250us).count());
 }
 
 AidlPowerHalWrapper::~AidlPowerHalWrapper() {
@@ -310,7 +608,7 @@
         mPowerHintSession->close();
         mPowerHintSession = nullptr;
     }
-};
+}
 
 std::unique_ptr<PowerAdvisor::HalWrapper> AidlPowerHalWrapper::connect() {
     // This only waits if the service is actually declared
@@ -348,7 +646,7 @@
     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;
 }
@@ -402,30 +700,14 @@
     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());
@@ -434,8 +716,8 @@
     }
 }
 
-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;
@@ -444,65 +726,42 @@
     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);
@@ -512,8 +771,8 @@
             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;
     }
 }
 
@@ -529,12 +788,13 @@
     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;
@@ -543,7 +803,7 @@
         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;
 
@@ -560,7 +820,7 @@
 
     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;
         }
@@ -568,7 +828,7 @@
         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
@@ -579,13 +839,12 @@
         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 61bb32b..19a3b88 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 @@
     const bool mUseScreenUpdateTimer;
     std::atomic_bool mSendUpdateImminent = true;
     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/Layer.cpp b/services/surfaceflinger/Layer.cpp
index d47e423..2d3b237 100644
--- a/services/surfaceflinger/Layer.cpp
+++ b/services/surfaceflinger/Layer.cpp
@@ -1157,11 +1157,13 @@
     return StretchEffect{};
 }
 
-bool Layer::enableBorder(bool shouldEnable) {
-    if (mBorderEnabled == shouldEnable) {
+bool Layer::enableBorder(bool shouldEnable, float width, const half4& color) {
+    if (mBorderEnabled == shouldEnable && mBorderWidth == width && mBorderColor == color) {
         return false;
     }
     mBorderEnabled = shouldEnable;
+    mBorderWidth = width;
+    mBorderColor = color;
     return true;
 }
 
@@ -1169,6 +1171,14 @@
     return mBorderEnabled;
 }
 
+float Layer::getBorderWidth() {
+    return mBorderWidth;
+}
+
+const half4& Layer::getBorderColor() {
+    return mBorderColor;
+}
+
 bool Layer::propagateFrameRateForLayerTree(FrameRate parentFrameRate, bool* transactionNeeded) {
     // The frame rate for layer tree is this layer's frame rate if present, or the parent frame rate
     const auto frameRate = [&] {
diff --git a/services/surfaceflinger/Layer.h b/services/surfaceflinger/Layer.h
index 85187e1..60c97f9 100644
--- a/services/surfaceflinger/Layer.h
+++ b/services/surfaceflinger/Layer.h
@@ -895,8 +895,10 @@
 
     bool setStretchEffect(const StretchEffect& effect);
     StretchEffect getStretchEffect() const;
-    bool enableBorder(bool shouldEnable);
+    bool enableBorder(bool shouldEnable, float width, const half4& color);
     bool isBorderEnabled();
+    float getBorderWidth();
+    const half4& getBorderColor();
 
     virtual bool setBufferCrop(const Rect& /* bufferCrop */) { return false; }
     virtual bool setDestinationFrame(const Rect& /* destinationFrame */) { return false; }
@@ -1149,6 +1151,8 @@
     bool findInHierarchy(const sp<Layer>&);
 
     bool mBorderEnabled = false;
+    float mBorderWidth;
+    half4 mBorderColor;
 };
 
 std::ostream& operator<<(std::ostream& stream, const Layer::FrameRate& rate);
diff --git a/services/surfaceflinger/SurfaceFlinger.cpp b/services/surfaceflinger/SurfaceFlinger.cpp
index 1d5d353..b6c247d 100644
--- a/services/surfaceflinger/SurfaceFlinger.cpp
+++ b/services/surfaceflinger/SurfaceFlinger.cpp
@@ -453,6 +453,11 @@
     }
 
     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() {
@@ -1985,12 +1990,6 @@
 
 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.
@@ -2004,10 +2003,6 @@
     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;
@@ -2083,6 +2078,30 @@
         }
     }
 
+    // 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;
@@ -2159,16 +2178,15 @@
         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);
@@ -2179,6 +2197,8 @@
         mDrawingState.traverse([&refreshArgs](Layer* layer) {
             if (layer->isBorderEnabled()) {
                 compositionengine::BorderRenderInfo info;
+                info.width = layer->getBorderWidth();
+                info.color = layer->getBorderColor();
                 layer->traverse(LayerVector::StateSet::Drawing, [&info](Layer* ilayer) {
                     info.layerIds.push_back(ilayer->getSequence());
                 });
@@ -2230,12 +2250,15 @@
 
     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);
     }
@@ -2280,11 +2303,8 @@
         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());
     }
 }
 
@@ -4462,7 +4482,7 @@
         if (layer->setBlurRegions(s.blurRegions)) flags |= eTraversalNeeded;
     }
     if (what & layer_state_t::eRenderBorderChanged) {
-        if (layer->enableBorder(s.borderEnabled)) {
+        if (layer->enableBorder(s.borderEnabled, s.borderWidth, s.borderColor)) {
             flags |= eTraversalNeeded;
         }
     }
diff --git a/services/surfaceflinger/SurfaceFlinger.h b/services/surfaceflinger/SurfaceFlinger.h
index b21799e..91de4a6 100644
--- a/services/surfaceflinger/SurfaceFlinger.h
+++ b/services/surfaceflinger/SurfaceFlinger.h
@@ -1430,12 +1430,12 @@
         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/LayerBorder_test.cpp b/services/surfaceflinger/tests/LayerBorder_test.cpp
index 4b38214..f80c705 100644
--- a/services/surfaceflinger/tests/LayerBorder_test.cpp
+++ b/services/surfaceflinger/tests/LayerBorder_test.cpp
@@ -36,7 +36,7 @@
 
         const auto display = SurfaceComposerClient::getInternalDisplayToken();
         ASSERT_FALSE(display == nullptr);
-
+        mColorOrange = toHalf4({255, 140, 0, 255});
         mParentLayer = createColorLayer("Parent layer", Color::RED);
 
         mContainerLayer = mClient->createSurface(String8("Container Layer"), 0 /* width */,
@@ -82,6 +82,7 @@
     std::function<half3(Color)> toHalf3;
     std::function<half4(Color)> toHalf4;
     sp<SurfaceControl> mParentLayer, mContainerLayer, mEffectLayer1, mEffectLayer2;
+    half4 mColorOrange;
 };
 
 TEST_F(LayerBorderTest, OverlappingVisibleRegions) {
@@ -89,7 +90,7 @@
         t.setCrop(mEffectLayer1, Rect(0, 0, 400, 400));
         t.setCrop(mEffectLayer2, Rect(200, 200, 600, 600));
 
-        t.enableBorder(mContainerLayer, true);
+        t.enableBorder(mContainerLayer, true, 20, mColorOrange);
         t.show(mEffectLayer1);
         t.show(mEffectLayer2);
         t.show(mContainerLayer);
@@ -101,7 +102,7 @@
         t.setCrop(mEffectLayer1, Rect(0, 0, 400, 400));
         t.setCrop(mEffectLayer2, Rect(200, 200, 600, 600));
 
-        t.enableBorder(mEffectLayer1, true);
+        t.enableBorder(mEffectLayer1, true, 20, mColorOrange);
         t.show(mEffectLayer1);
         t.show(mEffectLayer2);
         t.show(mContainerLayer);
@@ -113,7 +114,7 @@
         t.setCrop(mEffectLayer1, Rect(0, 0, 200, 200));
         t.setCrop(mEffectLayer2, Rect(400, 400, 600, 600));
 
-        t.enableBorder(mContainerLayer, true);
+        t.enableBorder(mContainerLayer, true, 20, mColorOrange);
         t.show(mEffectLayer1);
         t.show(mEffectLayer2);
         t.show(mContainerLayer);
@@ -125,7 +126,7 @@
         t.setCrop(mEffectLayer1, Rect(200, 200, 400, 400));
         t.setCrop(mEffectLayer2, Rect(0, 0, 600, 600));
 
-        t.enableBorder(mEffectLayer1, true);
+        t.enableBorder(mEffectLayer1, true, 20, mColorOrange);
         t.show(mEffectLayer1);
         t.show(mEffectLayer2);
         t.show(mContainerLayer);
@@ -140,7 +141,7 @@
         t.setLayer(mEffectLayer1, 30);
         t.setLayer(mEffectLayer2, 20);
 
-        t.enableBorder(mEffectLayer1, true);
+        t.enableBorder(mEffectLayer1, true, 20, mColorOrange);
         t.show(mEffectLayer1);
         t.show(mEffectLayer2);
         t.show(mContainerLayer);
@@ -169,7 +170,7 @@
         t.setCrop(effectLayer3, Rect(400, 400, 800, 800));
         t.setColor(effectLayer3, toHalf3(Color::BLUE));
 
-        t.enableBorder(mContainerLayer, true);
+        t.enableBorder(mContainerLayer, true, 20, mColorOrange);
         t.show(mEffectLayer1);
         t.show(mEffectLayer2);
         t.show(effectLayer3);
@@ -183,7 +184,7 @@
         t.setCrop(mEffectLayer2, Rect(200, 200, 600, 600));
         t.setAlpha(mEffectLayer1, 0.0f);
 
-        t.enableBorder(mEffectLayer1, true);
+        t.enableBorder(mContainerLayer, true, 20, mColorOrange);
         t.show(mEffectLayer1);
         t.show(mEffectLayer2);
         t.show(mContainerLayer);
@@ -196,7 +197,7 @@
         t.setCrop(mEffectLayer2, Rect(200, 200, 600, 600));
         t.setAlpha(mEffectLayer2, 0.5f);
 
-        t.enableBorder(mEffectLayer2, true);
+        t.enableBorder(mEffectLayer2, true, 20, mColorOrange);
         t.show(mEffectLayer1);
         t.show(mEffectLayer2);
         t.show(mContainerLayer);
@@ -208,7 +209,7 @@
         t.setCrop(mEffectLayer1, Rect(0, 0, 400, 400));
         t.setCrop(mEffectLayer2, Rect(200, 200, 600, 600));
 
-        t.enableBorder(mContainerLayer, true);
+        t.enableBorder(mContainerLayer, true, 20, mColorOrange);
         t.hide(mEffectLayer2);
         t.show(mContainerLayer);
     });
@@ -237,7 +238,43 @@
         t.setBuffer(bufferStateLayer, buffer);
         t.setPosition(bufferStateLayer, 100, 100);
         t.show(bufferStateLayer);
-        t.enableBorder(mContainerLayer, true);
+        t.enableBorder(mContainerLayer, true, 20, mColorOrange);
+    });
+}
+
+TEST_F(LayerBorderTest, CustomWidth) {
+    asTransaction([&](Transaction& t) {
+        t.setCrop(mEffectLayer1, Rect(0, 0, 400, 400));
+        t.setCrop(mEffectLayer2, Rect(200, 200, 600, 600));
+
+        t.enableBorder(mContainerLayer, true, 50, mColorOrange);
+        t.show(mEffectLayer1);
+        t.show(mEffectLayer2);
+        t.show(mContainerLayer);
+    });
+}
+
+TEST_F(LayerBorderTest, CustomColor) {
+    asTransaction([&](Transaction& t) {
+        t.setCrop(mEffectLayer1, Rect(0, 0, 400, 400));
+        t.setCrop(mEffectLayer2, Rect(200, 200, 600, 600));
+
+        t.enableBorder(mContainerLayer, true, 20, toHalf4({255, 0, 255, 255}));
+        t.show(mEffectLayer1);
+        t.show(mEffectLayer2);
+        t.show(mContainerLayer);
+    });
+}
+
+TEST_F(LayerBorderTest, CustomWidthAndColorAndOpacity) {
+    asTransaction([&](Transaction& t) {
+        t.setCrop(mEffectLayer1, Rect(0, 0, 200, 200));
+        t.setCrop(mEffectLayer2, Rect(400, 400, 600, 600));
+
+        t.enableBorder(mContainerLayer, true, 40, toHalf4({255, 255, 0, 128}));
+        t.show(mEffectLayer1);
+        t.show(mEffectLayer2);
+        t.show(mContainerLayer);
     });
 }
 
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/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 f74ffa3..a4164c0 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/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