AppVeyor hangs after a build, making me think it's not running bazel in batch mode. Travis was not building with swiftshader as defined in its bazelrc, which is also weird.

As titled

Meta bug for tracking initial Android bring-up:

• xtool setup of Android SDK/NDK
• Travis CI setup of Android SDK/NDK
• target_platform configuration for crosstool/cpu combinations
• portable_cc_test (or, maybe new skylark magic?)
  • adb use cross-platform?
  • adb cc_test/cc_binary data dep upload
  • gtest log writing and retrieval
• xtool build/test to build Android targets (when SDK is present)
• jni_utils ported to xrtl
• javah skylark magic for exposing NDK code to java
• java_cc_wrapper for bringing Java APIs into NDK code

Not great to have input events deferred - would be better when there's no need for ordering to instead execute the handlers inline on the message loop, if possible. Would need to change around PostEvent & co, maybe making it only for X11.

Offsets should be 0,0 for interior control origin. Win32 seems to double pad or something.

file_manifest.cc tries to access envvar("TEST_SRCDIR"), but that's only set when running under bazel. We need to either figure out what this path would be in 'xtool sln' or have a fallback.

heap-checker.h

Thread was written before ArrayView; the Thread::Wait* functions are a bit hard to use because of this. Updating it and other places that take vectors today would be nice.

Would also be nice to add ostream helpers for all threading return codes to enable CHECKs and logging.

Add a MessageLoop::AdoptCurrent() that returned a MessageLoop-compatible instance for posting. This would let us interact with the existing main-thread looper/runloop/etc on platforms where we are not the entry point.

Would be nice to have a clang build on Windows (as clang can do that now!) and it would enable clang-tidy for xtool tidy and xtool presubmit. We could upstream this in bazel, maybe, too as --cpu=x64_windows_clang.

find_changed_source_files is TODO

https://github.com/bazelbuild/buildtools

Numbers show it being not much different than pthreads and it would really help contention around multiple waits. Linux/Android can use epoll/eventfd, and BSD/Darwin can use kqueue.

.

Currently blocked on bazelbuild/bazel#2768. Things run under cygwin, but the paths are all for cygwin's bazel output path, which is not good for Visual Studio.

Add a skylark rule that accepts .java files and spits out a header per file; or one single header (to avoid outs mess). @local_jdk should have javah, but there may be a better bazel way to get it (or from the Android NDK).

The tests pass on MacOS but not on Windows and some Linux variants. Not sure why, as most of this is from TensorFlow.

[ RUN      ] LoggingTest.FailedChecks
C:\users\ben\appdata\local\temp\_bazel_ben\4qe-y0h3\execroot\xrtl\xrtl\base\logging_test.cc(72): error: Death test: if ((!(false))) ::xrtl::internal::LogMessageFatal("C:\\users\\ben\\appdata\\local\\temp\\_bazel_ben\\4qe-y0h3\\execroot\\xrtl\\xrtl\\base\\logging_test.cc", 72) << "Check failed: " "false" " "
    Result: died but not with expected error.
  Expected: false
Actual msg:

::Sleep is terrible. Need to round down and then spin for a bit to get within even the right order of magnitude.

Relevant magic:
bazelbuild/bazel@321aa54

Currently the ES3Shader push constant reflection code and the ES3RenderPassCommandEncoder::UpdatePushConstants code don't handle most supported types (like anything integer/ivec/etc).

Currently we use the exhaustive list of vulkan pixel formats - it'd make the API much easier to work with if we filtered it down to the ones guaranteed available on our target platforms.

glTF has a decent breakdown here: KhronosGroup/glTF#835

the command and output is as below, I am running it on an ubuntu 14.04 machine.

$ ./xtool setup
Setting up the build environment...

• git submodule init / update...
  Cloning into 'third_party/glm'...
  remote: Counting objects: 46915, done.
  remote: Compressing objects: 100% (7/7), done.
  remote: Total 46915 (delta 2), reused 4 (delta 1), pack-reused 46907
  Receiving objects: 100% (46915/46915), 62.37 MiB | 4.36 MiB/s, done.
  error: RPC failed; result=56, HTTP code = 200
  Resolving deltas: 100% (34376/34376), done.
  Clone of 'https://github.com/g-truc/glm.git' into submodule path 'third_party/glm' failed
  Traceback (most recent call last):
  File "./xtool", line 1821, in
  sys.exit(main())
  File "./xtool", line 126, in main
  return_code = command.execute(args, pass_args, os.getcwd())
  File "./xtool", line 662, in execute
  git_submodule_update()
  File "./xtool", line 474, in git_submodule_update
  '--recursive',
  File "./xtool", line 324, in shell_call
  subprocess.check_call(command, shell=False, stdout=stdout_file)
  File "/usr/lib/python2.7/subprocess.py", line 540, in check_call
  raise CalledProcessError(retcode, cmd)
  subprocess.CalledProcessError: Command '['git', 'submodule', 'update', '--init', '--recursive']' returned non-zero exit status 1

Using the material design icon font is probably the easiest thing to do with imgui.

Icon font here:
https://github.com/google/material-design-icons/tree/master/iconfont

Example code to generate headers with defines:
https://github.com/juliettef/IconFontCppHeaders/blob/master/IconsMaterialDesign.h

Our gtest dependency on ourselves (for flags/logging) causes issues on some bazel platforms. Example error here from travis.

Need to find a workaround for this, possibly by just poking into files in //third_party/gtest and not treating it as a separate repo.

This is currently breaking clang-tidy in presubmits.

Only need objects created when mapping, otherwise can be deferred until first use on the queue thread.
Thinking:

• context has a single platform context for resource stuff (this is the 'root' share group context)
• queues have a context of their own
• swap chains have a context of their own

The context could then have a small thread that listens for requests to map/unmap buffers and delete objects, and creations happen delayed on first execution.

Currently call bazel info execution_root and others a bunch. Would be nice to cache per run for each configuration in some helper, as the results should be the same for each config.

https://github.com/google/oss-fuzz

Find minimal shader module definition (exported functions, push constant ranges, binding slots, etc) and add to shader_tool output. To allow the compile-time headers without requiring the use of embedded shaders a new header file (--output_spirv_layout={input}_layout.h) could contain the information.

Currently playing with fire by reusing the same buffers for drawing. Need to write a gfx::RingBuffer to reuse that does the right things here, as well as add QueueFence queries and waiting to properly manage the read/write heads.

Can't have QueueFence be an Event, unfortunately, but could add a Wait() method on it as well as a query.

Can support both wait and query easily on vulkan, webgl2/es3, and d3d. Metal needs some custom work via command submit completion handlers, but nothing bad.