diff --git a/vulkan/libvulkan/api.cpp b/vulkan/libvulkan/api.cpp
index 673a066..71048db 100644
--- a/vulkan/libvulkan/api.cpp
+++ b/vulkan/libvulkan/api.cpp
@@ -21,6 +21,8 @@
 // There are a few of them requiring manual code for things such as layer
 // discovery or chaining.  They call into functions defined in this file.
 
+#define ATRACE_TAG ATRACE_TAG_GRAPHICS
+
 #include <stdlib.h>
 #include <string.h>
 
@@ -32,6 +34,7 @@
 #include <android-base/strings.h>
 #include <cutils/properties.h>
 #include <log/log.h>
+#include <utils/Trace.h>
 
 #include <vulkan/vk_layer_interface.h>
 #include <graphicsenv/GraphicsEnv.h>
@@ -1176,6 +1179,8 @@
 VkResult CreateInstance(const VkInstanceCreateInfo* pCreateInfo,
                         const VkAllocationCallbacks* pAllocator,
                         VkInstance* pInstance) {
+    ATRACE_CALL();
+
     if (!EnsureInitialized())
         return VK_ERROR_INITIALIZATION_FAILED;
 
@@ -1184,6 +1189,8 @@
 
 void DestroyInstance(VkInstance instance,
                      const VkAllocationCallbacks* pAllocator) {
+    ATRACE_CALL();
+
     if (instance != VK_NULL_HANDLE)
         LayerChain::DestroyInstance(instance, pAllocator);
 }
@@ -1192,17 +1199,23 @@
                       const VkDeviceCreateInfo* pCreateInfo,
                       const VkAllocationCallbacks* pAllocator,
                       VkDevice* pDevice) {
+    ATRACE_CALL();
+
     return LayerChain::CreateDevice(physicalDevice, pCreateInfo, pAllocator,
                                     pDevice);
 }
 
 void DestroyDevice(VkDevice device, const VkAllocationCallbacks* pAllocator) {
+    ATRACE_CALL();
+
     if (device != VK_NULL_HANDLE)
         LayerChain::DestroyDevice(device, pAllocator);
 }
 
 VkResult EnumerateInstanceLayerProperties(uint32_t* pPropertyCount,
                                           VkLayerProperties* pProperties) {
+    ATRACE_CALL();
+
     if (!EnsureInitialized())
         return VK_ERROR_INITIALIZATION_FAILED;
 
@@ -1225,6 +1238,8 @@
     const char* pLayerName,
     uint32_t* pPropertyCount,
     VkExtensionProperties* pProperties) {
+    ATRACE_CALL();
+
     if (!EnsureInitialized())
         return VK_ERROR_INITIALIZATION_FAILED;
 
@@ -1253,6 +1268,8 @@
 VkResult EnumerateDeviceLayerProperties(VkPhysicalDevice physicalDevice,
                                         uint32_t* pPropertyCount,
                                         VkLayerProperties* pProperties) {
+    ATRACE_CALL();
+
     uint32_t count;
     const LayerChain::ActiveLayer* layers =
         LayerChain::GetActiveLayers(physicalDevice, count);
@@ -1275,6 +1292,8 @@
     const char* pLayerName,
     uint32_t* pPropertyCount,
     VkExtensionProperties* pProperties) {
+    ATRACE_CALL();
+
     if (pLayerName) {
         // EnumerateDeviceLayerProperties enumerates active layers for
         // backward compatibility.  The extension query here should work for
@@ -1302,6 +1321,8 @@
 }
 
 VkResult EnumerateInstanceVersion(uint32_t* pApiVersion) {
+    ATRACE_CALL();
+
     *pApiVersion = VK_API_VERSION_1_1;
     return VK_SUCCESS;
 }
diff --git a/vulkan/libvulkan/driver.cpp b/vulkan/libvulkan/driver.cpp
index a607a5d..421f727 100644
--- a/vulkan/libvulkan/driver.cpp
+++ b/vulkan/libvulkan/driver.cpp
@@ -14,6 +14,8 @@
  * limitations under the License.
  */
 
+#define ATRACE_TAG ATRACE_TAG_GRAPHICS
+
 #include <malloc.h>
 #include <stdlib.h>
 #include <string.h>
@@ -31,6 +33,7 @@
 #include <configstore/Utils.h>
 #include <cutils/properties.h>
 #include <graphicsenv/GraphicsEnv.h>
+#include <utils/Trace.h>
 #include <utils/Vector.h>
 
 #include "android-base/properties.h"
@@ -150,6 +153,8 @@
 void* LoadLibrary(const android_dlextinfo& dlextinfo,
                   const char* subname,
                   int subname_len) {
+    ATRACE_CALL();
+
     const char kLibFormat[] = "vulkan.%*s.so";
     char* name = static_cast<char*>(
         alloca(sizeof(kLibFormat) + static_cast<size_t>(subname_len)));
@@ -164,6 +169,8 @@
 
 int LoadDriver(android_namespace_t* library_namespace,
                const hwvulkan_module_t** module) {
+    ATRACE_CALL();
+
     const android_dlextinfo dlextinfo = {
         .flags = ANDROID_DLEXT_USE_NAMESPACE,
         .library_namespace = library_namespace,
@@ -198,6 +205,8 @@
 }
 
 int LoadBuiltinDriver(const hwvulkan_module_t** module) {
+    ATRACE_CALL();
+
     auto ns = android_get_exported_namespace("sphal");
     if (!ns)
         return -ENOENT;
@@ -205,6 +214,8 @@
 }
 
 int LoadUpdatedDriver(const hwvulkan_module_t** module) {
+    ATRACE_CALL();
+
     auto ns = android::GraphicsEnv::getInstance().getDriverNamespace();
     if (!ns)
         return -ENOENT;
@@ -212,6 +223,8 @@
 }
 
 bool Hal::Open() {
+    ATRACE_CALL();
+
     ALOG_ASSERT(!hal_.dev_, "OpenHAL called more than once");
 
     // Use a stub device unless we successfully open a real HAL device.
@@ -242,9 +255,11 @@
     }
 
     hwvulkan_device_t* device;
+    ATRACE_BEGIN("hwvulkan module open");
     result =
         module->common.methods->open(&module->common, HWVULKAN_DEVICE_0,
                                      reinterpret_cast<hw_device_t**>(&device));
+    ATRACE_END();
     if (result != 0) {
         // Any device with a Vulkan HAL should be able to open the device.
         ALOGE("failed to open Vulkan HAL device: %s (%d)", strerror(-result),
@@ -260,6 +275,8 @@
 }
 
 bool Hal::InitDebugReportIndex() {
+    ATRACE_CALL();
+
     uint32_t count;
     if (dev_->EnumerateInstanceExtensionProperties(nullptr, &count, nullptr) !=
         VK_SUCCESS) {
@@ -821,8 +838,10 @@
         }
     }
 
+    ATRACE_BEGIN("driver.EnumerateInstanceExtensionProperties");
     VkResult result = Hal::Device().EnumerateInstanceExtensionProperties(
         pLayerName, pPropertyCount, pProperties);
+    ATRACE_END();
 
     if (!pLayerName && (result == VK_SUCCESS || result == VK_INCOMPLETE)) {
         int idx = Hal::Get().GetDebugReportIndex();
@@ -931,8 +950,10 @@
         *pPropertyCount -= count;
     }
 
+    ATRACE_BEGIN("driver.EnumerateDeviceExtensionProperties");
     VkResult result = data.driver.EnumerateDeviceExtensionProperties(
         physicalDevice, pLayerName, pPropertyCount, pProperties);
+    ATRACE_END();
 
     if (pProperties) {
         // map VK_ANDROID_native_buffer to VK_KHR_swapchain
@@ -968,12 +989,15 @@
     if (result != VK_SUCCESS)
         return result;
 
+    ATRACE_BEGIN("AllocateInstanceData");
     InstanceData* data = AllocateInstanceData(data_allocator);
+    ATRACE_END();
     if (!data)
         return VK_ERROR_OUT_OF_HOST_MEMORY;
 
     data->hook_extensions |= wrapper.GetHookExtensions();
 
+    ATRACE_BEGIN("autoDowngradeApiVersion");
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wold-style-cast"
     uint32_t api_version = ((pCreateInfo->pApplicationInfo)
@@ -989,7 +1013,9 @@
     if (!pfn_enumerate_instance_version) {
         icd_api_version = VK_API_VERSION_1_0;
     } else {
+        ATRACE_BEGIN("pfn_enumerate_instance_version");
         result = (*pfn_enumerate_instance_version)(&icd_api_version);
+        ATRACE_END();
     }
     uint32_t icd_api_major_version = VK_VERSION_MAJOR(icd_api_version);
     uint32_t icd_api_minor_version = VK_VERSION_MINOR(icd_api_version);
@@ -1000,12 +1026,15 @@
         wrapper.DowngradeApiVersion();
     }
 #pragma clang diagnostic pop
+    ATRACE_END();
 
     // call into the driver
     VkInstance instance;
+    ATRACE_BEGIN("driver.CreateInstance");
     result = Hal::Device().CreateInstance(
         static_cast<const VkInstanceCreateInfo*>(wrapper), pAllocator,
         &instance);
+    ATRACE_END();
     if (result != VK_SUCCESS) {
         FreeInstanceData(data, data_allocator);
         return result;
@@ -1066,8 +1095,10 @@
     if (result != VK_SUCCESS)
         return result;
 
+    ATRACE_BEGIN("AllocateDeviceData");
     DeviceData* data = AllocateDeviceData(data_allocator,
                                           instance_data.debug_report_callbacks);
+    ATRACE_END();
     if (!data)
         return VK_ERROR_OUT_OF_HOST_MEMORY;
 
@@ -1075,9 +1106,11 @@
 
     // call into the driver
     VkDevice dev;
+    ATRACE_BEGIN("driver.CreateDevice");
     result = instance_data.driver.CreateDevice(
         physicalDevice, static_cast<const VkDeviceCreateInfo*>(wrapper),
         pAllocator, &dev);
+    ATRACE_END();
     if (result != VK_SUCCESS) {
         FreeDeviceData(data, data_allocator);
         return result;
@@ -1114,8 +1147,10 @@
     }
 
     VkPhysicalDeviceProperties properties;
+    ATRACE_BEGIN("driver.GetPhysicalDeviceProperties");
     instance_data.driver.GetPhysicalDeviceProperties(physicalDevice,
                                                      &properties);
+    ATRACE_END();
 
     data->driver_device = dev;
     data->driver_version = properties.driverVersion;
@@ -1141,6 +1176,8 @@
 VkResult EnumeratePhysicalDevices(VkInstance instance,
                                   uint32_t* pPhysicalDeviceCount,
                                   VkPhysicalDevice* pPhysicalDevices) {
+    ATRACE_CALL();
+
     const auto& data = GetData(instance);
 
     VkResult result = data.driver.EnumeratePhysicalDevices(
@@ -1157,6 +1194,8 @@
     VkInstance instance,
     uint32_t* pPhysicalDeviceGroupCount,
     VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties) {
+    ATRACE_CALL();
+
     VkResult result = VK_SUCCESS;
     const auto& data = GetData(instance);
 
@@ -1217,6 +1256,8 @@
                     uint32_t queueFamilyIndex,
                     uint32_t queueIndex,
                     VkQueue* pQueue) {
+    ATRACE_CALL();
+
     const auto& data = GetData(device);
 
     data.driver.GetDeviceQueue(device, queueFamilyIndex, queueIndex, pQueue);
@@ -1226,6 +1267,8 @@
 void GetDeviceQueue2(VkDevice device,
                      const VkDeviceQueueInfo2* pQueueInfo,
                      VkQueue* pQueue) {
+    ATRACE_CALL();
+
     const auto& data = GetData(device);
 
     data.driver.GetDeviceQueue2(device, pQueueInfo, pQueue);
@@ -1236,6 +1279,8 @@
 AllocateCommandBuffers(VkDevice device,
                        const VkCommandBufferAllocateInfo* pAllocateInfo,
                        VkCommandBuffer* pCommandBuffers) {
+    ATRACE_CALL();
+
     const auto& data = GetData(device);
 
     VkResult result = data.driver.AllocateCommandBuffers(device, pAllocateInfo,
diff --git a/vulkan/libvulkan/layers_extensions.cpp b/vulkan/libvulkan/layers_extensions.cpp
index ba4cf00..af1adcf 100644
--- a/vulkan/libvulkan/layers_extensions.cpp
+++ b/vulkan/libvulkan/layers_extensions.cpp
@@ -14,6 +14,8 @@
  * limitations under the License.
  */
 
+#define ATRACE_TAG ATRACE_TAG_GRAPHICS
+
 #include "layers_extensions.h"
 
 #include <alloca.h>
@@ -33,6 +35,7 @@
 #include <log/log.h>
 #include <nativebridge/native_bridge.h>
 #include <nativeloader/native_loader.h>
+#include <utils/Trace.h>
 #include <ziparchive/zip_archive.h>
 
 // TODO(jessehall): The whole way we deal with extensions is pretty hokey, and
@@ -428,6 +431,8 @@
 }
 
 void DiscoverLayersInPathList(const std::string& pathstr) {
+    ATRACE_CALL();
+
     std::vector<std::string> paths = android::base::Split(pathstr, ":");
     for (const auto& path : paths) {
         ForEachFileInPath(path, [&](const std::string& filename) {
@@ -477,6 +482,8 @@
 }  // anonymous namespace
 
 void DiscoverLayers() {
+    ATRACE_CALL();
+
     if (property_get_bool("ro.debuggable", false) &&
         prctl(PR_GET_DUMPABLE, 0, 0, 0, 0)) {
         DiscoverLayersInPathList(kSystemLayerLibraryDir);
diff --git a/vulkan/libvulkan/swapchain.cpp b/vulkan/libvulkan/swapchain.cpp
index 8601373..ea1119a 100644
--- a/vulkan/libvulkan/swapchain.cpp
+++ b/vulkan/libvulkan/swapchain.cpp
@@ -14,6 +14,8 @@
  * limitations under the License.
  */
 
+#define ATRACE_TAG ATRACE_TAG_GRAPHICS
+
 #include <algorithm>
 
 #include <grallocusage/GrallocUsageConversion.h>
@@ -21,6 +23,7 @@
 #include <ui/BufferQueueDefs.h>
 #include <sync/sync.h>
 #include <utils/StrongPointer.h>
+#include <utils/Trace.h>
 #include <utils/Vector.h>
 #include <system/window.h>
 #include <android/hardware/graphics/common/1.0/types.h>
@@ -486,6 +489,8 @@
     const VkAndroidSurfaceCreateInfoKHR* pCreateInfo,
     const VkAllocationCallbacks* allocator,
     VkSurfaceKHR* out_surface) {
+    ATRACE_CALL();
+
     if (!allocator)
         allocator = &GetData(instance).allocator;
     void* mem = allocator->pfnAllocation(allocator->pUserData, sizeof(Surface),
@@ -528,6 +533,8 @@
 void DestroySurfaceKHR(VkInstance instance,
                        VkSurfaceKHR surface_handle,
                        const VkAllocationCallbacks* allocator) {
+    ATRACE_CALL();
+
     Surface* surface = SurfaceFromHandle(surface_handle);
     if (!surface)
         return;
@@ -548,6 +555,8 @@
                                             uint32_t /*queue_family*/,
                                             VkSurfaceKHR surface_handle,
                                             VkBool32* supported) {
+    ATRACE_CALL();
+
     const Surface* surface = SurfaceFromHandle(surface_handle);
     if (!surface) {
         return VK_ERROR_SURFACE_LOST_KHR;
@@ -589,6 +598,8 @@
     VkPhysicalDevice /*pdev*/,
     VkSurfaceKHR surface,
     VkSurfaceCapabilitiesKHR* capabilities) {
+    ATRACE_CALL();
+
     int err;
     ANativeWindow* window = SurfaceFromHandle(surface)->window.get();
 
@@ -662,6 +673,8 @@
                                             VkSurfaceKHR surface_handle,
                                             uint32_t* count,
                                             VkSurfaceFormatKHR* formats) {
+    ATRACE_CALL();
+
     const InstanceData& instance_data = GetData(pdev);
 
     // TODO(jessehall): Fill out the set of supported formats. Longer term, add
@@ -734,6 +747,8 @@
     VkPhysicalDevice physicalDevice,
     const VkPhysicalDeviceSurfaceInfo2KHR* pSurfaceInfo,
     VkSurfaceCapabilities2KHR* pSurfaceCapabilities) {
+    ATRACE_CALL();
+
     VkResult result = GetPhysicalDeviceSurfaceCapabilitiesKHR(
         physicalDevice, pSurfaceInfo->surface,
         &pSurfaceCapabilities->surfaceCapabilities);
@@ -769,6 +784,8 @@
     const VkPhysicalDeviceSurfaceInfo2KHR* pSurfaceInfo,
     uint32_t* pSurfaceFormatCount,
     VkSurfaceFormat2KHR* pSurfaceFormats) {
+    ATRACE_CALL();
+
     if (!pSurfaceFormats) {
         return GetPhysicalDeviceSurfaceFormatsKHR(physicalDevice,
                                                   pSurfaceInfo->surface,
@@ -800,6 +817,8 @@
                                                  VkSurfaceKHR surface,
                                                  uint32_t* count,
                                                  VkPresentModeKHR* modes) {
+    ATRACE_CALL();
+
     int err;
     int query_value;
     ANativeWindow* window = SurfaceFromHandle(surface)->window.get();
@@ -851,6 +870,8 @@
 VkResult GetDeviceGroupPresentCapabilitiesKHR(
     VkDevice,
     VkDeviceGroupPresentCapabilitiesKHR* pDeviceGroupPresentCapabilities) {
+    ATRACE_CALL();
+
     ALOGV_IF(pDeviceGroupPresentCapabilities->sType !=
                  VK_STRUCTURE_TYPE_DEVICE_GROUP_PRESENT_CAPABILITIES_KHR,
              "vkGetDeviceGroupPresentCapabilitiesKHR: invalid "
@@ -873,6 +894,8 @@
     VkDevice,
     VkSurfaceKHR,
     VkDeviceGroupPresentModeFlagsKHR* pModes) {
+    ATRACE_CALL();
+
     *pModes = VK_DEVICE_GROUP_PRESENT_MODE_LOCAL_BIT_KHR;
     return VK_SUCCESS;
 }
@@ -882,6 +905,8 @@
                                                VkSurfaceKHR surface,
                                                uint32_t* pRectCount,
                                                VkRect2D* pRects) {
+    ATRACE_CALL();
+
     if (!pRects) {
         *pRectCount = 1;
     } else {
@@ -923,6 +948,8 @@
                             const VkSwapchainCreateInfoKHR* create_info,
                             const VkAllocationCallbacks* allocator,
                             VkSwapchainKHR* swapchain_handle) {
+    ATRACE_CALL();
+
     int err;
     VkResult result = VK_SUCCESS;
 
@@ -1147,9 +1174,11 @@
     int32_t legacy_usage = 0;
     if (dispatch.GetSwapchainGrallocUsage2ANDROID) {
         uint64_t consumer_usage, producer_usage;
+        ATRACE_BEGIN("dispatch.GetSwapchainGrallocUsage2ANDROID");
         result = dispatch.GetSwapchainGrallocUsage2ANDROID(
             device, create_info->imageFormat, create_info->imageUsage,
             swapchain_image_usage, &consumer_usage, &producer_usage);
+        ATRACE_END();
         if (result != VK_SUCCESS) {
             ALOGE("vkGetSwapchainGrallocUsage2ANDROID failed: %d", result);
             return VK_ERROR_SURFACE_LOST_KHR;
@@ -1157,9 +1186,11 @@
         legacy_usage =
             android_convertGralloc1To0Usage(producer_usage, consumer_usage);
     } else if (dispatch.GetSwapchainGrallocUsageANDROID) {
+        ATRACE_BEGIN("dispatch.GetSwapchainGrallocUsageANDROID");
         result = dispatch.GetSwapchainGrallocUsageANDROID(
             device, create_info->imageFormat, create_info->imageUsage,
             &legacy_usage);
+        ATRACE_END();
         if (result != VK_SUCCESS) {
             ALOGE("vkGetSwapchainGrallocUsageANDROID failed: %d", result);
             return VK_ERROR_SURFACE_LOST_KHR;
@@ -1254,8 +1285,10 @@
             &image_native_buffer.usage2.producer,
             &image_native_buffer.usage2.consumer);
 
+        ATRACE_BEGIN("dispatch.CreateImage");
         result =
             dispatch.CreateImage(device, &image_create, nullptr, &img.image);
+        ATRACE_END();
         if (result != VK_SUCCESS) {
             ALOGD("vkCreateImage w/ native buffer failed: %u", result);
             break;
@@ -1279,8 +1312,11 @@
             }
         }
         if (result != VK_SUCCESS) {
-            if (img.image)
+            if (img.image) {
+                ATRACE_BEGIN("dispatch.DestroyImage");
                 dispatch.DestroyImage(device, img.image, nullptr);
+                ATRACE_END();
+            }
         }
     }
 
@@ -1299,6 +1335,8 @@
 void DestroySwapchainKHR(VkDevice device,
                          VkSwapchainKHR swapchain_handle,
                          const VkAllocationCallbacks* allocator) {
+    ATRACE_CALL();
+
     const auto& dispatch = GetData(device).driver;
     Swapchain* swapchain = SwapchainFromHandle(swapchain_handle);
     if (!swapchain)
@@ -1324,6 +1362,8 @@
                                VkSwapchainKHR swapchain_handle,
                                uint32_t* count,
                                VkImage* images) {
+    ATRACE_CALL();
+
     Swapchain& swapchain = *SwapchainFromHandle(swapchain_handle);
     ALOGW_IF(swapchain.surface.swapchain_handle != swapchain_handle,
              "getting images for non-active swapchain 0x%" PRIx64
@@ -1352,6 +1392,8 @@
                              VkSemaphore semaphore,
                              VkFence vk_fence,
                              uint32_t* image_index) {
+    ATRACE_CALL();
+
     Swapchain& swapchain = *SwapchainFromHandle(swapchain_handle);
     ANativeWindow* window = swapchain.surface.window.get();
     VkResult result;
@@ -1432,6 +1474,8 @@
 VkResult AcquireNextImage2KHR(VkDevice device,
                               const VkAcquireNextImageInfoKHR* pAcquireInfo,
                               uint32_t* pImageIndex) {
+    ATRACE_CALL();
+
     // TODO: this should actually be the other way around and this function
     // should handle any additional structures that get passed in
     return AcquireNextImageKHR(device, pAcquireInfo->swapchain,
@@ -1461,6 +1505,8 @@
 
 VKAPI_ATTR
 VkResult QueuePresentKHR(VkQueue queue, const VkPresentInfoKHR* present_info) {
+    ATRACE_CALL();
+
     ALOGV_IF(present_info->sType != VK_STRUCTURE_TYPE_PRESENT_INFO_KHR,
              "vkQueuePresentKHR: invalid VkPresentInfoKHR structure type %d",
              present_info->sType);
@@ -1672,6 +1718,8 @@
     VkDevice,
     VkSwapchainKHR swapchain_handle,
     VkRefreshCycleDurationGOOGLE* pDisplayTimingProperties) {
+    ATRACE_CALL();
+
     Swapchain& swapchain = *SwapchainFromHandle(swapchain_handle);
     VkResult result = VK_SUCCESS;
 
@@ -1687,6 +1735,8 @@
     VkSwapchainKHR swapchain_handle,
     uint32_t* count,
     VkPastPresentationTimingGOOGLE* timings) {
+    ATRACE_CALL();
+
     Swapchain& swapchain = *SwapchainFromHandle(swapchain_handle);
     ANativeWindow* window = swapchain.surface.window.get();
     VkResult result = VK_SUCCESS;
@@ -1711,6 +1761,8 @@
 VkResult GetSwapchainStatusKHR(
     VkDevice,
     VkSwapchainKHR swapchain_handle) {
+    ATRACE_CALL();
+
     Swapchain& swapchain = *SwapchainFromHandle(swapchain_handle);
     VkResult result = VK_SUCCESS;
 
@@ -1728,6 +1780,7 @@
     uint32_t swapchainCount,
     const VkSwapchainKHR* pSwapchains,
     const VkHdrMetadataEXT* pHdrMetadataEXTs) {
+    ATRACE_CALL();
 
     for (uint32_t idx = 0; idx < swapchainCount; idx++) {
         Swapchain* swapchain = SwapchainFromHandle(pSwapchains[idx]);