- Move the overlay option strings to XSTR 1940 and 1941. The touch
overlay on master uses 1936 to 1939.
- Give FrameProfiler the main thread ID when it is made. Before, it
used the thread of the first event, which was a race.
- Let mouse motion go to the game while the pointer is over the
overlay, so that mouse flight does not stop there.
- Find the dynamic bindings of the PerDraw key with
dynamic_offset_slot(), as the other call sites do.
- Make disable_inside_open_scope count the GPU timestamp queries, so
that it fails if end() does not close the GPU event.
- Make the copy test copy the category, not bind a reference.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- Use structured bindings and std::scoped_lock in the category registry.
- Return size_t from getCount() and update its callers and tests.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- Updated `trace_event` to store `category_id` instead of a pointer to prevent dangling references to category objects that may already be destroyed.
- Modified all references to category names to use `Category::getNameById()` instead of dereferencing pointers.
- Expanded memory statistics in the profiler overlay with new sound, Lua, and OS-level memory usage tracking.
- Added new tests to verify proper use of category IDs in frame profiling and event tracing.
- Improved memory safety and stability across tracing, profiling, and overlay systems.
The category registry kept a Category pointer for each id. The comment said that all categories
are global statics, but that is not correct. The Lua API tracing_category constructs a Category
and copies it into a script-owned object. The registry then kept a pointer to the dead temporary.
build_overlay_snapshot called getName() on that pointer, made an SCP_string from the garbage
char*, and the process stopped with SIGABRT.
The registry now keeps the name by value and never keeps a Category pointer. getById() becomes
getNameById(), which is all that build_overlay_snapshot needs. A copy of a Category keeps the id
of the original, which stays correct because the id only selects a name.
Add two tests: one shows that the name stays available after the Category object is dead, and one
drives the full path that failed, from processEvent to the overlay snapshot.
Co-Authored-By: Claude Opus 5 <[email protected]>
- Consolidated dynamic descriptor offset logic into `DescriptorWriter::bindSets`, replacing repetitive `vkCmdBindDescriptorSets` patterns across Vulkan modules.
- Eliminated redundant descriptor binding operations for Material and PerDraw sets by centralizing logic in `bindSets`.
- Simplified descriptors by directly splitting offset changes from descriptor contents, avoiding unnecessary updates.
- Updated comments and documentation to align with the new approach while maintaining reference consistency.
- Applied optimizations across Vulkan modules (`VulkanPostProcessing`, `VulkanDescriptorManager`, etc.) to reduce per-draw overhead and improve readability.
The Material and PerDraw descriptor sets were reallocated and fully rewritten on
essentially every draw call. model_draw_list::render_buffer rebinds ModelData with
a fresh uniform buffer offset for each queued draw, and that offset was part of the
Material set's memoization key, so the cache in VulkanDrawManager::applyMaterial
missed every time - taking a frame-pool allocation, a full template write, the
16-entry material texture resolution and a vkUpdateDescriptorSets with it.
GenericData/Matrices did the same to the PerDraw set. renderShadowDraw was worse
still: all three sets allocated and rewritten per shadow draw, with no memoization
at all.
A large asteroid field (Solaris m21) measured 2566 descriptor sets and 19107
descriptor writes per frame at 1606 model draws, and grew an extra descriptor pool
chunk every frame.
Declare the bindings whose offset moves per draw - MaterialBinding::ModelData,
MaterialBinding::ShadowMapData, PerDrawBinding::GenericData and
PerDrawBinding::Matrices - as eUniformBufferDynamic. DescriptorWriter::setBuffer
now splits a dynamic binding's offset out of the descriptor, writing
{buffer, 0, range} and stashing the offset for vkCmdBindDescriptorSets'
pDynamicOffsets, so an offset-only change leaves the set's contents identical and
the memoization caches start hitting. Shader sources are unchanged.
VulkanStateTracker::bindDescriptorSet now compares the dynamic offsets as well as
the set handle. This is required, not an optimization: the common case is now the
same set rebound with a new offset, and the old handle-only redundancy check would
have skipped that bind and fed every draw the previous draw's uniforms.
renderShadowDraw memoizes its three sets the way applyMaterial does. It does not
invalidate applyMaterial's caches - those record what a set contains, the shadow
pass allocates its own sets rather than overwriting theirs, and it binds through
the state tracker, so the next material cache hit still rebinds correctly.
Also switch allocateFrameSet to the non-allocating allocateDescriptorSets overload;
the vector-returning one heap-allocated once per set, thousands of times a frame.
MAX_SETS_PER_POOL stays at 1024 - the per-frame pool growth it used to log was a
symptom of the churn, and one chunk now covers the scene outright.
Measured on that mission in a Debug build, A/B on the same tree with and without
this change:
descriptor sets/frame @ ~1370 draws 2566 -> 250-345
descriptor writes/frame 19107 -> ~1700-2300
median frametime 92.9/83.6 ms -> 72.0/66.5 ms
Render Buffer (profiler overlay) 24.68 ms (30.8%) -> 12.73 ms (18.4%)
Build Shadow Map 5.44 ms -> off the top-5 entirely
Release builds already inlined most of this away and are unchanged, as expected.
Verified with the validation layers (sync validation enabled): no new VUIDs versus
the same build without the change.
Co-Authored-By: Claude Opus 5 <[email protected]>
- Introduced `accumulate_self_times` for single-pass exclusive time computation, improving performance and accuracy.
- Updated `FrameProfiler` and overlay snapshot logic to leverage `Category`'s stable `getId` and `getCount`.
- Replaced pie chart with a stacked bar for frame budget visualization, optimizing rendering cost.
- Cached `get_tid` and `get_pid` with thread-local storage to minimize syscall overhead.
- Added support for unit testing frame profiling functionality and included new test cases.
- Improved code comments and clarified documentation.
Replaces the old text-dump frame profile display (-profile_frame_time)
with a runtime-toggleable ImGui/ImPlot overlay showing average/median
frametime, a scrolling frametime graph, and a pie chart of the top
contributors to frame time. Toggled via a new "Frame Profiler Overlay"
option, which also seeds from -profile_frame_time for backward
compatibility and now drives frame profiling collection at runtime
instead of only at startup.
* use std::move() for local variables that are copied but not reused
Addresses ~140 Coverity "Variable copied when it could be moved"
defects across 74 files. Every std::move() targets a local variable
whose value is consumed by push_back, emplace, assignment, lambda
capture, or a function call and is never read again afterward.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
* clang: use std::make_shared
* clang: use std::make_unique
* address feedback
---------
Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>
Several changes related to the volumetrics model fix:
1. Update the `model_load()` prototype with better parameter names, including a new `ErrorType` enum defined in pstypes.h
2. Update the `MIN`, `MAX`, and `CLAMP` macros, all of which were susceptible to performance issues because they duplicated their tokens. This could cause functions or calculations to be re-run unnecessarily. Since the codebase is now on C++17, `std::clamp()` can be used.
3. Update the rest of the codebase to fix any associated issues.
The old code is not able to handle multi-threaded events so this change
simply filters out all the events that originated from another thread.
This fixes an Assertion caused by an invalid event.
The timers on Windows apparently have a lower resolution than Linux
which causes issues with overlapping events. These changes apply an
offset to all timestamps if the same timestamp is detected while
processing the buffered events.
This is basically the old code but it has been ported to the new tracing
system.
There is still some weirdness with the percentages but I have no idea
where that comes from since I don't know the code.
