Suggested by encounter, it fixes this error:
[FATAL | aurora::gpu] WebGPU error 2: Error while parsing WGSL: :446:5 error: cannot assign to value of type 'swizzle<function, vec3<f32>, read_write, 4, 3>'
prev.rgb = tevreg0.rgb;
* handle_draw optimization
1. cache vtxSize
2. remove heap allocation from draw call merging index buffer
3. move things that aren't on the "draw call merging hot path" out to other functions, to reduce stack frame size of handle_draw (it was previously using __chkstk)
* Make ByteBuffer expand exponentially
Fix half a second of startup time being just that.
* Babysit the compiler's inlining decisions
* Merged draw call idxBuf now a global static
* GXFlush() doesn't write NOPs to FIFO
I'm sure this made sense on actual hardware
* Don't bind unused textures/samplers
* Remove C++ RAII types from build_bind_groups
Most of this function's code was fucking with WebGPU AddRef and ReleaseRef. Seriously.
* More refs in common.hpp
* Index buffer in handle_draw_unmerged now not dynamically allocated for small draws
* Fix gx_test_stubs.cpp
* Some cleanup & renaming
---------
Co-authored-by: Luke Street <luke@street.dev>
* Sprinkle Tracy all over the project
* Move Tracy FetchContent to extern file, disable Tracy by default
* Use older CMake syntax for setting cache variables
* Fix test compile
* Rename ALIGN macro to AURORA_ALIGN
Needed because it conflicts with a macro in Apple's SDK headers on my macbook, and Tracy is now pulling in that header.
* Set up dawn cache
Massively speeds up pipeline compilation on the second go around, especially for D3D12.
Currently, uses SQLite for the cache. SQLite is committed in as the "amalgam" blob, but I'm open for better ways to fetch it.
* Compress cache contents with zstd
More significantly for Vulkan, less so for D3D12.
* Store cache keys as XXH128 instead
This reduces the size of the D3D12 cache by a ridiculous amount
* Stop using exceptions for error handling
* FetchContent for sqlite; cleanup & make zstd optional
---------
Co-authored-by: Luke Street <luke@street.dev>
* Add functionality to lock GX framebuffer aspect ratio
This leaves the imgui (so including OS) framebuffer alone.
AuroraWindowSize now has new fields for the native framebuffer, the aspect ratio controls modify the existing fields.
* Fix from merge
* fix gx_test_stubs.cpp
* add ability to set path of card image before CARDInit
* add missing func to osx/linux variant of ResolveDolphinCardPath
* remove format specifiers in formatting funcs
* missed some oops
* refactor card loading to allow loading slot A and B if present
if neither slots have card images, a slot A image will be created at the save directory
this commit also changes card image names to "MemoryCardA/B.USA.raw" to fit with dolphins naming conventions (region code will always be USA unless changed in the macro, which could eventually be changed to be something in aurora config)
* only remove file name from CARDSetBasePath if not a directory