Add wait before present option

Force wait on completion of queue operations for all queues before calling Present.
This is needed for accurate acquisition of instrumentation data on some platforms.

Change-Id: Ie5c07df95ad4420ed516f6f59c719d91887fa73a

USAGE_android.md

@@ -833,7 +833,7 @@ usage: gfxrecon.py replay [-h] [--push-file LOCAL_FILE] [--version] [--pause-fra
                           [--measurement-file DEVICE_FILE] [--quit-after-measurement-range]
                           [--flush-measurement-range] [-m MODE]
                           [--swapchain MODE] [--use-captured-swapchain-indices]
-                          [--use-colorspace-fallback]
+                          [--use-colorspace-fallback] [--wait-before-present]
                           [file]
 
 Launch the replay tool.
@@ -973,6 +973,10 @@ optional arguments:
   --sgfr FRAME-RANGES, --skip-get-fence-ranges FRAME-RANGES
                         Frame ranges where --sgfs applies. Default is all frames
                         (forwarded to replay tool)
+  --wait-before-present
+                        Force wait on completion of queue operations for all queues
+                        before calling Present. This is needed for accurate acquisition
+                        of instrumentation data on some platforms.
 ```
 
 The command will force-stop an active replay process before starting the replay

USAGE_desktop_Vulkan.md

@@ -526,14 +526,13 @@ gfxrecon-replay         [-h | --help] [--version] [--gpu <index>]
                         [--surface-index <N>] [--remove-unsupported] [--validate]
                         [-m <mode> | --memory-translation <mode>]
                         [--fwo <x,y> | --force-windowed-origin <x,y>]
-                        [--use-captured-swapchain-indices]
                         [--swapchain MODE] [--use-captured-swapchain-indices]
                         [--mfr|--measurement-frame-range <start-frame>-<end-frame>]
                         [--measurement-file <file>] [--quit-after-measurement-range]
                         [--flush-measurement-range]
                         [--log-level <level>] [--log-file <file>] [--log-debugview]
-                        [--no-debug-popup] <file>
-                        [--use-colorspace-fallback]
+                        [--no-debug-popup] [--use-colorspace-fallback]
+                        [--wait-before-present] <file>
 
 Required arguments:
   <file>                Path to the capture file to replay.
@@ -695,6 +694,10 @@ Optional arguments:
   --sgfr <frame-ranges>
               Frame ranges where --sgfs applies. The format is:
                 <frame-start-1>-<frame-end-1>[,<frame-start-1>-<frame-end-1>]*
+  --wait-before-present
+              Force wait on completion of queue operations for all queues
+              before calling Present. This is needed for accurate acquisition
+              of instrumentation data on some platforms.
 ```
 
 ### Key Controls

android/scripts/gfxrecon.py

@@ -93,6 +93,7 @@ def CreateReplayParser():
     parser.add_argument('--flush-inside-measurement-range', action='store_true', default=False, help='If this is specified the replayer will flush and wait for all current GPU work to finish at end of each frame inside the measurement range. (forwarded to replay tool)')
     parser.add_argument('--sgfs', '--skip-get-fence-status', metavar='STATUS', default=0, help='Specify behaviour to skip calls to vkWaitForFences and vkGetFenceStatus. Default is 0 - No skip (forwarded to replay tool)')
     parser.add_argument('--sgfr', '--skip-get-fence-ranges', metavar='FRAME-RANGES', default='', help='Frame ranges where --sgfs applies. Default is all frames (forwarded to replay tool)')
+    parser.add_argument('--wait-before-present', action='store_true', default=False, help='Force wait on completion of queue operations for all queues before calling Present. This is needed for accurate acquisition of instrumentation data on some platforms.')
     parser.add_argument('-m', '--memory-translation', metavar='MODE', choices=['none', 'remap', 'realign', 'rebind'], help='Enable memory translation for replay on GPUs with memory types that are not compatible with the capture GPU\'s memory types.  Available modes are: none, remap, realign, rebind (forwarded to replay tool)')
     parser.add_argument('--swapchain', metavar='MODE', choices=['virtual', 'captured', 'offscreen'], help='Choose a swapchain mode to replay. Available modes are: virtual, captured, offscreen (forwarded to replay tool)')
     parser.add_argument('--vssb', '--virtual-swapchain-skip-blit', action='store_true', default=False, help='Skip blit to real swapchain to gain performance during replay.')
@@ -217,6 +218,9 @@ def MakeExtrasString(args):
         arg_list.append('-m')
         arg_list.append('{}'.format(args.memory_translation))
 
+    if args.wait_before_present:
+        arg_list.append('--wait-before-present')
+
     if args.file:
         arg_list.append(args.file)
     elif not args.version:

framework/decode/vulkan_replay_consumer_base.cpp

@@ -6143,6 +6143,12 @@ VulkanReplayConsumerBase::OverrideQueuePresentKHR(PFN_vkQueuePresentKHR
         modified_present_info.pImageIndices = modified_image_indices_.data();
     }
 
+    if (options_.wait_before_present)
+    {
+        VkDevice device = MapHandle<DeviceInfo>(queue_info->parent_id, &VulkanObjectInfoTable::GetDeviceInfo);
+        GetDeviceTable(device)->DeviceWaitIdle(device);
+    }
+
     // Only attempt to find imported or shadow semaphores if we know at least one around.
     if ((!have_imported_semaphores_) && (shadow_semaphores_.empty()) && (modified_present_info.swapchainCount != 0))
     {

framework/decode/vulkan_replay_options.h

@@ -71,6 +71,7 @@ struct VulkanReplayOptions : public ReplayOptions
     std::string                  replace_dir;
     SkipGetFenceStatus           skip_get_fence_status{ SkipGetFenceStatus::NoSkip };
     std::vector<util::UintRange> skip_get_fence_ranges;
+    bool                         wait_before_present{ false };
 };
 
 GFXRECON_END_NAMESPACE(decode)

tools/replay/replay_settings.h

@@ -32,7 +32,7 @@ const char kOptions[] =
     "screenshot-all,--onhb|--omit-null-hardware-buffers,--qamr|--quit-after-measurement-range,--fmr|--flush-"
     "measurement-range,--flush-inside-measurement-range,--vssb|--virtual-swapchain-skip-blit,--use-captured-swapchain-"
     "indices,--dcp,--discard-cached-psos,--use-colorspace-fallback,--use-cached-psos,--dx12-override-object-names,--"
-    "offscreen-swapchain-frame-boundary";
+    "offscreen-swapchain-frame-boundary,--wait-before-present";
 const char kArguments[] =
     "--log-level,--log-file,--gpu,--gpu-group,--pause-frame,--wsi,--surface-index,-m|--memory-translation,"
     "--replace-shaders,--screenshots,--denied-messages,--allowed-messages,--screenshot-format,--"
@@ -255,6 +255,10 @@ static void PrintUsage(const char* exe_name)
     GFXRECON_WRITE_CONSOLE("  --sgfr <frame-ranges>");
     GFXRECON_WRITE_CONSOLE("          \t\tFrame ranges where --sgfs applies. The format is:");
     GFXRECON_WRITE_CONSOLE("          \t\t\t<frame-start-1>-<frame-end-1>[,<frame-start-1>-<frame-end-1>]*");
+    GFXRECON_WRITE_CONSOLE("  --wait-before-present");
+    GFXRECON_WRITE_CONSOLE("          \t\tForce wait on completion of queue operations for all queues");
+    GFXRECON_WRITE_CONSOLE("          \t\tbefore calling Present. This is needed for accurate acquisition");
+    GFXRECON_WRITE_CONSOLE("          \t\tof instrumentation data on some platforms.");
 
 #if defined(WIN32)
     GFXRECON_WRITE_CONSOLE("")

tools/tool_settings.h

@@ -119,6 +119,7 @@ const char kExpandFlagsOption[]                   = "--expand-flags";
 const char kFilePerFrameOption[]                  = "--file-per-frame";
 const char kSkipGetFenceStatus[]                  = "--skip-get-fence-status";
 const char kSkipGetFenceRanges[]                  = "--skip-get-fence-ranges";
+const char kWaitBeforePresent[]                   = "--wait-before-present";
 #if defined(WIN32)
 const char kDxTwoPassReplay[]             = "--dx12-two-pass-replay";
 const char kDxOverrideObjectNames[]       = "--dx12-override-object-names";
@@ -957,6 +958,10 @@ GetVulkanReplayOptions(const gfxrecon::util::ArgumentParser&           arg_parse
         replay_options.skip_get_fence_ranges =
             gfxrecon::util::GetUintRanges(skip_get_fence_ranges.c_str(), "skip-get-fence-ranges");
     }
+    if (arg_parser.IsOptionSet(kWaitBeforePresent))
+    {
+        replay_options.wait_before_present = true;
+    }
 
     return replay_options;
 }