diff --git a/libs/gui/BufferQueue.cpp b/libs/gui/BufferQueue.cpp
index 66fc16d78..d76168023 100644
--- a/libs/gui/BufferQueue.cpp
+++ b/libs/gui/BufferQueue.cpp
@@ -16,6 +16,7 @@
 
 #define LOG_TAG "BufferQueue"
 //#define LOG_NDEBUG 0
+#define ATRACE_TAG ATRACE_TAG_GRAPHICS
 
 #define GL_GLEXT_PROTOTYPES
 #define EGL_EGLEXT_PROTOTYPES
@@ -29,6 +30,7 @@
 
 #include <utils/Log.h>
 #include <gui/SurfaceTexture.h>
+#include <utils/Trace.h>
 
 // This compile option causes SurfaceTexture to return the buffer that is currently
 // attached to the GL texture from dequeueBuffer when no other buffers are
@@ -191,6 +193,7 @@ status_t BufferQueue::setBufferCount(int bufferCount) {
 
 int BufferQueue::query(int what, int* outValue)
 {
+    ATRACE_CALL();
     Mutex::Autolock lock(mMutex);
 
     if (mAbandoned) {
@@ -221,6 +224,7 @@ int BufferQueue::query(int what, int* outValue)
 }
 
 status_t BufferQueue::requestBuffer(int slot, sp<GraphicBuffer>* buf) {
+    ATRACE_CALL();
     ST_LOGV("requestBuffer: slot=%d", slot);
     Mutex::Autolock lock(mMutex);
     if (mAbandoned) {
@@ -239,6 +243,7 @@ status_t BufferQueue::requestBuffer(int slot, sp<GraphicBuffer>* buf) {
 
 status_t BufferQueue::dequeueBuffer(int *outBuf, uint32_t w, uint32_t h,
         uint32_t format, uint32_t usage) {
+    ATRACE_CALL();
     ST_LOGV("dequeueBuffer: w=%d h=%d fmt=%#x usage=%#x", w, h, format, usage);
 
     if ((w && !h) || (!w && h)) {
@@ -458,6 +463,7 @@ status_t BufferQueue::dequeueBuffer(int *outBuf, uint32_t w, uint32_t h,
 }
 
 status_t BufferQueue::setSynchronousMode(bool enabled) {
+    ATRACE_CALL();
     ST_LOGV("setSynchronousMode: enabled=%d", enabled);
     Mutex::Autolock lock(mMutex);
 
@@ -490,6 +496,7 @@ status_t BufferQueue::setSynchronousMode(bool enabled) {
 
 status_t BufferQueue::queueBuffer(int buf, int64_t timestamp,
         uint32_t* outWidth, uint32_t* outHeight, uint32_t* outTransform) {
+    ATRACE_CALL();
     ST_LOGV("queueBuffer: slot=%d time=%lld", buf, timestamp);
 
     sp<FrameAvailableListener> listener;
@@ -553,6 +560,8 @@ status_t BufferQueue::queueBuffer(int buf, int64_t timestamp,
         *outWidth = mDefaultWidth;
         *outHeight = mDefaultHeight;
         *outTransform = 0;
+
+        ATRACE_INT(mConsumerName.string(), mQueue.size());
     } // scope for the lock
 
     // call back without lock held
@@ -563,6 +572,7 @@ status_t BufferQueue::queueBuffer(int buf, int64_t timestamp,
 }
 
 void BufferQueue::cancelBuffer(int buf) {
+    ATRACE_CALL();
     ST_LOGV("cancelBuffer: slot=%d", buf);
     Mutex::Autolock lock(mMutex);
 
@@ -586,6 +596,7 @@ void BufferQueue::cancelBuffer(int buf) {
 }
 
 status_t BufferQueue::setCrop(const Rect& crop) {
+    ATRACE_CALL();
     ST_LOGV("setCrop: crop=[%d,%d,%d,%d]", crop.left, crop.top, crop.right,
             crop.bottom);
 
@@ -599,6 +610,7 @@ status_t BufferQueue::setCrop(const Rect& crop) {
 }
 
 status_t BufferQueue::setTransform(uint32_t transform) {
+    ATRACE_CALL();
     ST_LOGV("setTransform: xform=%#x", transform);
     Mutex::Autolock lock(mMutex);
     if (mAbandoned) {
@@ -610,6 +622,7 @@ status_t BufferQueue::setTransform(uint32_t transform) {
 }
 
 status_t BufferQueue::setScalingMode(int mode) {
+    ATRACE_CALL();
     ST_LOGV("setScalingMode: mode=%d", mode);
 
     switch (mode) {
@@ -628,6 +641,7 @@ status_t BufferQueue::setScalingMode(int mode) {
 
 status_t BufferQueue::connect(int api,
         uint32_t* outWidth, uint32_t* outHeight, uint32_t* outTransform) {
+    ATRACE_CALL();
     ST_LOGV("connect: api=%d", api);
     Mutex::Autolock lock(mMutex);
 
@@ -664,6 +678,7 @@ status_t BufferQueue::connect(int api,
 }
 
 status_t BufferQueue::disconnect(int api) {
+    ATRACE_CALL();
     ST_LOGV("disconnect: api=%d", api);
     Mutex::Autolock lock(mMutex);
 
@@ -818,6 +833,8 @@ status_t BufferQueue::acquire(BufferItem *buffer) {
 
         mSlots[buf].mBufferState = BufferSlot::ACQUIRED;
         mQueue.erase(front);
+
+        ATRACE_INT(mConsumerName.string(), mQueue.size());
     }
     else {
         return -EINVAL; //should be a better return code
@@ -875,6 +892,7 @@ status_t BufferQueue::setDefaultBufferSize(uint32_t w, uint32_t h)
 }
 
 status_t BufferQueue::setBufferCountServer(int bufferCount) {
+    ATRACE_CALL();
     Mutex::Autolock lock(mMutex);
     return setBufferCountServerLocked(bufferCount);
 }
diff --git a/libs/gui/SurfaceTexture.cpp b/libs/gui/SurfaceTexture.cpp
index a23a32801..b42aa34d7 100644
--- a/libs/gui/SurfaceTexture.cpp
+++ b/libs/gui/SurfaceTexture.cpp
@@ -15,6 +15,7 @@
  */
 
 #define LOG_TAG "SurfaceTexture"
+#define ATRACE_TAG ATRACE_TAG_GRAPHICS
 //#define LOG_NDEBUG 0
 
 #define GL_GLEXT_PROTOTYPES
@@ -36,6 +37,7 @@
 
 #include <utils/Log.h>
 #include <utils/String8.h>
+#include <utils/Trace.h>
 
 // This compile option makes SurfaceTexture use the EGL_KHR_fence_sync extension
 // to synchronize access to the buffers.  It will cause dequeueBuffer to stall,
@@ -143,6 +145,7 @@ status_t SurfaceTexture::setDefaultBufferSize(uint32_t w, uint32_t h)
 }
 
 status_t SurfaceTexture::updateTexImage() {
+    ATRACE_CALL();
     ST_LOGV("updateTexImage");
     Mutex::Autolock lock(mMutex);
 
diff --git a/libs/gui/SurfaceTextureClient.cpp b/libs/gui/SurfaceTextureClient.cpp
index dac54a8b1..f88dcaf02 100644
--- a/libs/gui/SurfaceTextureClient.cpp
+++ b/libs/gui/SurfaceTextureClient.cpp
@@ -15,9 +15,11 @@
  */
 
 #define LOG_TAG "SurfaceTextureClient"
+#define ATRACE_TAG ATRACE_TAG_GRAPHICS
 //#define LOG_NDEBUG 0
 
 #include <utils/Log.h>
+#include <utils/Trace.h>
 
 #include <gui/ISurfaceComposer.h>
 #include <gui/SurfaceComposerClient.h>
@@ -121,6 +123,7 @@ int SurfaceTextureClient::hook_perform(ANativeWindow* window, int operation, ...
 }
 
 int SurfaceTextureClient::setSwapInterval(int interval) {
+    ATRACE_CALL();
     // EGL specification states:
     //  interval is silently clamped to minimum and maximum implementation
     //  dependent values before being stored.
@@ -138,6 +141,7 @@ int SurfaceTextureClient::setSwapInterval(int interval) {
 }
 
 int SurfaceTextureClient::dequeueBuffer(android_native_buffer_t** buffer) {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::dequeueBuffer");
     Mutex::Autolock lock(mMutex);
     int buf = -1;
@@ -167,6 +171,7 @@ int SurfaceTextureClient::dequeueBuffer(android_native_buffer_t** buffer) {
 }
 
 int SurfaceTextureClient::cancelBuffer(android_native_buffer_t* buffer) {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::cancelBuffer");
     Mutex::Autolock lock(mMutex);
     int i = getSlotFromBufferLocked(buffer);
@@ -213,6 +218,7 @@ int SurfaceTextureClient::lockBuffer(android_native_buffer_t* buffer) {
 }
 
 int SurfaceTextureClient::queueBuffer(android_native_buffer_t* buffer) {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::queueBuffer");
     Mutex::Autolock lock(mMutex);
     int64_t timestamp;
@@ -236,6 +242,7 @@ int SurfaceTextureClient::queueBuffer(android_native_buffer_t* buffer) {
 }
 
 int SurfaceTextureClient::query(int what, int* value) const {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::query");
     { // scope for the lock
         Mutex::Autolock lock(mMutex);
@@ -404,6 +411,7 @@ int SurfaceTextureClient::dispatchUnlockAndPost(va_list args) {
 
 
 int SurfaceTextureClient::connect(int api) {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::connect");
     Mutex::Autolock lock(mMutex);
     int err = mSurfaceTexture->connect(api,
@@ -415,6 +423,7 @@ int SurfaceTextureClient::connect(int api) {
 }
 
 int SurfaceTextureClient::disconnect(int api) {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::disconnect");
     Mutex::Autolock lock(mMutex);
     freeAllBuffers();
@@ -441,6 +450,7 @@ int SurfaceTextureClient::setUsage(uint32_t reqUsage)
 
 int SurfaceTextureClient::setCrop(Rect const* rect)
 {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::setCrop");
     Mutex::Autolock lock(mMutex);
 
@@ -459,6 +469,7 @@ int SurfaceTextureClient::setCrop(Rect const* rect)
 
 int SurfaceTextureClient::setBufferCount(int bufferCount)
 {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::setBufferCount");
     Mutex::Autolock lock(mMutex);
 
@@ -475,6 +486,7 @@ int SurfaceTextureClient::setBufferCount(int bufferCount)
 
 int SurfaceTextureClient::setBuffersDimensions(int w, int h)
 {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::setBuffersDimensions");
     Mutex::Autolock lock(mMutex);
 
@@ -508,6 +520,7 @@ int SurfaceTextureClient::setBuffersFormat(int format)
 
 int SurfaceTextureClient::setScalingMode(int mode)
 {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::setScalingMode(%d)", mode);
     Mutex::Autolock lock(mMutex);
     // mode is validated on the server
@@ -520,6 +533,7 @@ int SurfaceTextureClient::setScalingMode(int mode)
 
 int SurfaceTextureClient::setBuffersTransform(int transform)
 {
+    ATRACE_CALL();
     ALOGV("SurfaceTextureClient::setBuffersTransform");
     Mutex::Autolock lock(mMutex);
     status_t err = mSurfaceTexture->setTransform(transform);
diff --git a/opengl/libs/EGL/eglApi.cpp b/opengl/libs/EGL/eglApi.cpp
index a5dc832bf..a1bd82dff 100644
--- a/opengl/libs/EGL/eglApi.cpp
+++ b/opengl/libs/EGL/eglApi.cpp
@@ -14,6 +14,8 @@
  ** limitations under the License.
  */
 
+#define ATRACE_TAG ATRACE_TAG_GRAPHICS
+
 #include <ctype.h>
 #include <stdlib.h>
 #include <string.h>
@@ -34,6 +36,7 @@
 #include <utils/KeyedVector.h>
 #include <utils/SortedVector.h>
 #include <utils/String8.h>
+#include <utils/Trace.h>
 
 #include "egl_impl.h"
 #include "egl_tls.h"
@@ -348,6 +351,7 @@ EGLBoolean eglQuerySurface( EGLDisplay dpy, EGLSurface surface,
 }
 
 void EGLAPI eglBeginFrame(EGLDisplay dpy, EGLSurface surface) {
+    ATRACE_CALL();
     clearError();
 
     egl_display_t const * const dp = validate_display(dpy);
@@ -712,6 +716,7 @@ __eglMustCastToProperFunctionPointerType eglGetProcAddress(const char *procname)
 
 EGLBoolean eglSwapBuffers(EGLDisplay dpy, EGLSurface draw)
 {
+    ATRACE_CALL();
     clearError();
 
     egl_display_t const * const dp = validate_display(dpy);
diff --git a/services/surfaceflinger/Layer.cpp b/services/surfaceflinger/Layer.cpp
index c15fdbc1e..43acce823 100644
--- a/services/surfaceflinger/Layer.cpp
+++ b/services/surfaceflinger/Layer.cpp
@@ -14,6 +14,8 @@
  * limitations under the License.
  */
 
+#define ATRACE_TAG ATRACE_TAG_GRAPHICS
+
 #include <stdlib.h>
 #include <stdint.h>
 #include <sys/types.h>
@@ -26,6 +28,7 @@
 #include <utils/Errors.h>
 #include <utils/Log.h>
 #include <utils/StopWatch.h>
+#include <utils/Trace.h>
 
 #include <ui/GraphicBuffer.h>
 #include <ui/PixelFormat.h>
@@ -267,6 +270,8 @@ void Layer::setPerFrameData(hwc_layer_t* hwcl) {
 
 void Layer::onDraw(const Region& clip) const
 {
+    ATRACE_CALL();
+
     if (CC_UNLIKELY(mActiveBuffer == 0)) {
         // the texture has not been created yet, this Layer has
         // in fact never been drawn into. This happens frequently with
@@ -365,6 +370,8 @@ bool Layer::isProtected() const
 
 uint32_t Layer::doTransaction(uint32_t flags)
 {
+    ATRACE_CALL();
+
     const Layer::State& front(drawingState());
     const Layer::State& temp(currentState());
 
@@ -418,6 +425,8 @@ bool Layer::onPreComposition() {
 
 void Layer::lockPageFlip(bool& recomputeVisibleRegions)
 {
+    ATRACE_CALL();
+
     if (mQueuedFrames > 0) {
 
         // if we've already called updateTexImage() without going through
@@ -540,6 +549,8 @@ void Layer::lockPageFlip(bool& recomputeVisibleRegions)
 void Layer::unlockPageFlip(
         const Transform& planeTransform, Region& outDirtyRegion)
 {
+    ATRACE_CALL();
+
     Region postedRegion(mPostedDirtyRegion);
     if (!postedRegion.isEmpty()) {
         mPostedDirtyRegion.clear();
diff --git a/services/surfaceflinger/SurfaceFlinger.cpp b/services/surfaceflinger/SurfaceFlinger.cpp
index 62461bb26..42ed4fa91 100644
--- a/services/surfaceflinger/SurfaceFlinger.cpp
+++ b/services/surfaceflinger/SurfaceFlinger.cpp
@@ -14,6 +14,8 @@
  * limitations under the License.
  */
 
+#define ATRACE_TAG ATRACE_TAG_GRAPHICS
+
 #include <stdlib.h>
 #include <stdio.h>
 #include <stdint.h>
@@ -39,6 +41,7 @@
 #include <utils/String8.h>
 #include <utils/String16.h>
 #include <utils/StopWatch.h>
+#include <utils/Trace.h>
 
 #include <ui/GraphicBufferAllocator.h>
 #include <ui/PixelFormat.h>
@@ -402,6 +405,7 @@ bool SurfaceFlinger::threadLoop()
 
 void SurfaceFlinger::onMessageReceived(int32_t what)
 {
+    ATRACE_CALL();
     switch (what) {
         case MessageQueue::REFRESH: {
 //        case MessageQueue::INVALIDATE: {
@@ -737,6 +741,7 @@ void SurfaceFlinger::commitTransaction()
 
 void SurfaceFlinger::handlePageFlip()
 {
+    ATRACE_CALL();
     const DisplayHardware& hw = graphicPlane(0).displayHardware();
     const Region screenRegion(hw.bounds());