[DO NOT SQUASH] Fast-forward third_party/gfxreconstruct to e30fe1f85

[footballhead]

subtree pull in order to get Improve developer build times by hiding SHA

Conflicts

FileProcessor

FileProcessor was refactored. While this doesn't alter functionality, it breaks some of our modifications:

• ProcessStateMarker() was removed. Its logic now lives in ProcessVisitor, DispatchVisitor, ProcessStateEndMarker(), and ProcessStateBeginMarker().
• ProcessFrameMarker() was removed. Its logic now lives in ProcessVisitor, DispatchVisitor, and ProcessFrameEndMarker()
• ReadBlockHeader() was effectively replaced. Its logic now lives largely in BlockProcessor. There is still a function with the same name but it's functionally different.
• ActiveFileContext now stores a FileInputStreamPtr instead of a filename. All file operations use FileInputStreamPtr
• SeekActiveFile() signature was modified to take a FileInputStreamPtr instead of a std::string

This broke the following features:

• Single frame looping
• PM4 dump
• RenderDoc capture
• Block modification

Single frame looping

We need to react to State End and Frame End markers. Now, we mark FileProcessor::ProcessStateEndMarker() and FileProcessor::ProcessFrameEndMarker() as virtual and override those in DiveFileProcessor. (Although, in the future, this functionality should be moved into DiveVulkanReplayConsumer. This derives from MarkerConsumerBase and provides the same functions for consumers to override but in a more stable way).

Given the changes to FileProcessor, we can drastically simplify the DiveFileProcessor::ProcessFrameEndMarker() implementation. The previous implementation started with a copy of FileProcessor::ProcessFrameMarker() then modified to support looping. Now, the only thing that we need to work around is the assert in FileProcessor::ProcessFrameDelimiter() (which ensures frames are processed in order) by manipulating the capture frame number.

Now that all file operations (seek, tell) require FileInputStreamPtr, we store a weak reference to the GFXR file instead of a std::string in DiveFileProcessor.

FileProcessor::TellFile() was moved to FileInputStream::FileTell(). This class abstracts the file descriptor so it's a logical place to expose a tell API.

Since DiveFileProcessor relies on SeekActiveFile(), ensure the required function is still virtual.

PM4 Dump, RenderDoc Capture

Like single frame looping, this is fixed by making FileProcessor::ProcessStateEndMarker() and FileProcessor:ProcessFrameEndMarker() virtual. (Also like single frame looping, this should be moved into DiveVulkanReplayConsumer.)

Block Modification

FileProcessor::StoreBlockInfo() is used to expose block information to derived classes. I ensured that the call is still in what I believe is the best place to get the block info before any reading occurs.

VulkanReplayConsumerBase

Some queue flag variables were renamed. When applying our patch, the names were updated.

replay_settings.h

Ensured that our flags for GPU time and single frame looping are still present.

android_main

Despite the addition of --capture (a.k.a. "recapture"), ensured that Dive additions persist.

[footballhead]

At 466d6fd:

The following tests FAILED:
	  3 - Regression_StackOverflow_MacOS (Subprocess aborted)
	  4 - DiveUI_Screenshot (Subprocess aborted)
	  5 - DiveUI_Screenshot_NativeStyle (Subprocess aborted)
	 90 - ModifyGfxrSeparateAssetsNoOp (Failed)

[footballhead]

The following tests FAILED:
3 - Regression_StackOverflow_MacOS (Subprocess aborted)
4 - DiveUI_Screenshot (Subprocess aborted)
5 - DiveUI_Screenshot_NativeStyle (Subprocess aborted)
90 - ModifyGfxrSeparateAssetsNoOp (Failed)

StoreBlockInfo wasn't being called. It was originally in ReadBlockHeader, then was shuffled into PeekBlockHeader by LunarG/gfxreconstruct@3131d85; the two functions are very similar so it probably tripped up the auto conflict resolition. But in that same commit, PreloadFileProcessor stopped using PeekBlockHeader. So there are no more callers