diff --git a/libs/gui/Surface.cpp b/libs/gui/Surface.cpp
index cabc1773e0..563c9a90f0 100644
--- a/libs/gui/Surface.cpp
+++ b/libs/gui/Surface.cpp
@@ -26,6 +26,9 @@
 #include <deque>
 #include <mutex>
 #include <thread>
+#include <sys/ioctl.h>
+#include <fcntl.h>
+#include <unistd.h>
 
 #include <inttypes.h>
 
@@ -36,8 +39,10 @@
 #include <gui/LayerState.h>
 #include <private/gui/ComposerService.h>
 #include <private/gui/ComposerServiceAIDL.h>
+#include <binder/IPCThreadState.h>
 #endif
 
+#include <android-base/properties.h>
 #include <android/native_window.h>
 
 #include <android-base/strings.h>
@@ -103,6 +108,37 @@ bool isInterceptorRegistrationOp(int op) {
 }
 
 } // namespace
+  //
+#define GED_MAGIC 'g'
+#define GED_BRIDGE_COMMAND_GPU_TIMESTAMP      103
+#define GED_IOWR(INDEX)  _IOWR(GED_MAGIC, INDEX, GED_BRIDGE_PACKAGE)
+#define GED_BRIDGE_IO_GPU_TIMESTAMP \
+    GED_IOWR(GED_BRIDGE_COMMAND_GPU_TIMESTAMP)
+typedef struct _GED_BRIDGE_PACKAGE {
+    unsigned int ui32FunctionID;
+    int i32Size;
+    void *pvParamIn;
+    int i32InBufferSize;
+    void *pvParamOut;
+    int i32OutBufferSize;
+} GED_BRIDGE_PACKAGE;
+
+struct GED_BRIDGE_IN_GPU_TIMESTAMP {
+    int pid;
+    uint64_t ullWnd;
+    int32_t i32FrameID;
+    int fence_fd;
+    int QedBuffer_length;
+    int isSF;
+};
+
+struct GED_BRIDGE_OUT_GPU_TIMESTAMP {
+    int eError;
+    int is_ged_kpi_enabled;
+};
+
+static int doMtkGedKpi = 1;
+static int ged_fd = -1;
 
 Surface::ProducerDeathListenerProxy::ProducerDeathListenerProxy(wp<SurfaceListener> surfaceListener)
       : mSurfaceListener(surfaceListener) {}
@@ -175,6 +211,43 @@ Surface::Surface(const sp<IGraphicBufferProducer>& bufferProducer, bool controll
     mSwapIntervalZero = false;
     mMaxBufferCount = NUM_BUFFER_SLOTS;
     mSurfaceControlHandle = surfaceControlHandle;
+
+    if (ged_fd == -1 && doMtkGedKpi == 1) {
+        ALOGE("Opening ged");
+        ged_fd = open("/proc/ged", O_RDONLY);
+        ALOGE("Opening ged ret = %d", ged_fd);
+        {
+        struct GED_BRIDGE_IN_GPU_TIMESTAMP in = {
+            .pid = 0,
+            //.ullWnd = (uint64_t)(intptr_t)this,
+                .ullWnd = 0,
+                .i32FrameID = 0,
+            .fence_fd = 0,
+            .isSF = 0,
+            .QedBuffer_length = 0,
+        };
+        struct GED_BRIDGE_OUT_GPU_TIMESTAMP out;
+        memset(&in, 0, sizeof(in));
+        GED_BRIDGE_PACKAGE package = {
+            .ui32FunctionID = GED_BRIDGE_IO_GPU_TIMESTAMP,
+            .i32Size = sizeof(GED_BRIDGE_PACKAGE),
+            .pvParamIn = &in,
+            .i32InBufferSize = sizeof(in),
+            .pvParamOut = &out,
+            .i32OutBufferSize = sizeof(out),
+        };
+        if (ged_fd >= 0) {
+            int ret = ioctl(ged_fd, GED_BRIDGE_IO_GPU_TIMESTAMP, &package);
+            ALOGE("First null timestamp ioctl returned %d %d %d", ret, out.eError, out.is_ged_kpi_enabled);
+            if (out.is_ged_kpi_enabled != 1) {
+                ALOGE("is_ged_kpi_enabled reported disabled");
+                doMtkGedKpi = 0;
+            }
+        } else {
+            ALOGE("No /proc/ged");
+        }
+        }
+    }
 }
 
 Surface::~Surface() {
@@ -754,6 +827,36 @@ int Surface::dequeueBuffer(sp<GraphicBuffer>* buffer, int* fenceFd) {
         }
     }
 
+    if (mGraphicBufferProducer != nullptr && ged_fd >= 0) {
+        uint64_t uniqueId;
+        mGraphicBufferProducer->getUniqueId(&uniqueId);
+
+        const int32_t dupFenceFd = fence->isValid() ? fence->dup() : -1;
+
+        struct GED_BRIDGE_IN_GPU_TIMESTAMP in = {
+            .pid = mPid,
+            .ullWnd = uniqueId,
+            .i32FrameID = static_cast<int32_t>(reinterpret_cast<intptr_t>(gbuf->handle)) & 0x3fffffff,
+            .fence_fd = dupFenceFd,
+            .isSF = mIsSurfaceFlinger ? 1 : 0,
+            .QedBuffer_length = -2,
+        };
+        struct GED_BRIDGE_OUT_GPU_TIMESTAMP out;
+        memset(&out, 0, sizeof(out));
+        GED_BRIDGE_PACKAGE package = {
+            .ui32FunctionID = GED_BRIDGE_IO_GPU_TIMESTAMP,
+            .i32Size = sizeof(GED_BRIDGE_PACKAGE),
+            .pvParamIn = &in,
+            .i32InBufferSize = sizeof(in),
+            .pvParamOut = &out,
+            .i32OutBufferSize = sizeof(out),
+        };
+
+        int ret = ioctl(ged_fd, GED_BRIDGE_IO_GPU_TIMESTAMP, &package);
+        ALOGV("GPU timestamp ioctl returned %d %d %d %d", ret, out.eError, out.is_ged_kpi_enabled, in.i32FrameID);
+
+        close(dupFenceFd);
+    }
     if (fence->isValid()) {
         *fenceFd = fence->dup();
         if (*fenceFd == -1) {
@@ -1322,6 +1425,60 @@ void Surface::onBufferQueuedLocked(int slot, sp<Fence> fence,
     }
 
     mQueueBufferCondition.broadcast();
+    if (mGraphicBufferProducer != nullptr && ged_fd >= 0) {
+        sp<GraphicBuffer>& gbuf(mSlots[slot].buffer);
+        uint64_t uniqueId;
+        mGraphicBufferProducer->getUniqueId(&uniqueId);
+
+        const int32_t dupFenceFd = fence->isValid() ? fence->dup() : -1;
+        // onQueue
+        {
+            struct GED_BRIDGE_IN_GPU_TIMESTAMP in = {
+                .pid = mPid,
+                .ullWnd = uniqueId,
+                .i32FrameID = static_cast<int32_t>(reinterpret_cast<intptr_t>(gbuf->handle)) & 0x3fffffff,
+                .fence_fd = dupFenceFd,
+                .isSF = mIsSurfaceFlinger ? 1 : 0,
+                .QedBuffer_length = static_cast<int>(output.numPendingBuffers),
+            };
+            struct GED_BRIDGE_OUT_GPU_TIMESTAMP out;
+            memset(&out, 0, sizeof(out));
+            GED_BRIDGE_PACKAGE package = {
+                .ui32FunctionID = GED_BRIDGE_IO_GPU_TIMESTAMP,
+                .i32Size = sizeof(GED_BRIDGE_PACKAGE),
+                .pvParamIn = &in,
+                .i32InBufferSize = sizeof(in),
+                .pvParamOut = &out,
+                .i32OutBufferSize = sizeof(out),
+            };
+            int ret = ioctl(ged_fd, GED_BRIDGE_IO_GPU_TIMESTAMP, &package);
+            ALOGV("GPU timestamp ioctl returned %d %d %d", ret, out.eError, in.i32FrameID);
+        }
+        // acquire
+        {
+            struct GED_BRIDGE_IN_GPU_TIMESTAMP in = {
+                .pid = mPid,
+                .isSF = mIsSurfaceFlinger ? 1 : 0,
+                .ullWnd = uniqueId,
+                .i32FrameID = static_cast<int32_t>(reinterpret_cast<intptr_t>(gbuf->handle)) & 0x3fffffff,
+                .fence_fd = dupFenceFd,
+                .QedBuffer_length = -1,
+            };
+            struct GED_BRIDGE_OUT_GPU_TIMESTAMP out;
+            memset(&out, 0, sizeof(out));
+            GED_BRIDGE_PACKAGE package = {
+                .ui32FunctionID = GED_BRIDGE_IO_GPU_TIMESTAMP,
+                .i32Size = sizeof(GED_BRIDGE_PACKAGE),
+                .pvParamIn = &in,
+                .i32InBufferSize = sizeof(in),
+                .pvParamOut = &in,
+                .i32OutBufferSize = sizeof(out),
+            };
+            int ret = ioctl(ged_fd, GED_BRIDGE_IO_GPU_TIMESTAMP, &package);
+            ALOGV("GPU timestamp ioctl returned %d %d %d", ret, out.eError, in.i32FrameID);
+        }
+        close(dupFenceFd);
+    }
 
     if (CC_UNLIKELY(atrace_is_tag_enabled(ATRACE_TAG_GRAPHICS))) {
         static gui::FenceMonitor gpuCompletionThread("GPU completion");
@@ -2196,6 +2353,46 @@ int Surface::connect(int api, const sp<SurfaceListener>& listener, bool reportBu
         SURF_LOGE_IF(idErr != NO_ERROR, "Unable to get ID from IGBP: %d", idErr);
         mIsConnected = true;
     }
+
+    // For MTK GED KPI, we need to grab the Surface owner's PID
+    // and also know whether that owner is surfaceflinger
+    if (api == NATIVE_WINDOW_API_EGL && ged_fd >= 0) {
+        IPCThreadState *ipc = IPCThreadState::selfOrNull();
+        const sp<IBinder>& token = IInterface::asBinder(mGraphicBufferProducer);
+        mPid =  (token != NULL && NULL != token->localBinder())
+            ? getpid()
+            : (ipc != nullptr)?ipc->getCallingPid():-1;
+
+        // We've got caller PID. Now checking whether it is surfaceflinger
+        char cmdline[128];
+        char path[128];
+        snprintf(path, sizeof(path)-1, "/proc/%d/cmdline", mPid);
+        int fd = open(path, O_RDONLY);
+        read(fd, cmdline, sizeof(cmdline)-1);
+        // Normally cmdline is already \0-separated, but well
+        for(unsigned i=0; i<sizeof(cmdline); i++)
+            if(cmdline[i] == '\n')
+                cmdline[i] = 0;
+        cmdline[sizeof(cmdline)-1] = 0;
+
+        close(fd);
+
+        // Truncate to last / (also called basename)
+        const char *c = strrchr(cmdline, '/');
+        if (c != nullptr) {
+            c = c+1;
+        } else {
+            c = cmdline;
+        }
+        if(strcmp(c, "surfaceflinger") == 0) {
+            ALOGE("is surfaceflinger = 1");
+            mIsSurfaceFlinger = true;
+        } else {
+            ALOGE("is surfaceflinger = 0");
+            mIsSurfaceFlinger = false;
+        }
+    }
+
     if (!err && api == NATIVE_WINDOW_API_CPU) {
         mConnectedToCpu = true;
         // Clear the dirty region in case we're switching from a non-CPU API
diff --git a/libs/gui/include/gui/Surface.h b/libs/gui/include/gui/Surface.h
index 9a6e75c333..d7c51d7d77 100644
--- a/libs/gui/include/gui/Surface.h
+++ b/libs/gui/include/gui/Surface.h
@@ -761,6 +761,9 @@ class Surface
     // Buffers that are successfully dequeued/attached and handed to clients
     std::unordered_set<int> mDequeuedSlots;
 
+    pid_t mPid;
+    bool mIsSurfaceFlinger;
+
     // Indicates whether this surface holds the mouse cursor, and subsequently determines whether
     // the GRALLOC_USAGE_CURSOR usage flag should be set on the buffer created when this surface is
     // locked.