- For some reason this has a massive impact on performance above some arbitrary threshold of calls
Shows up under surface_cache::get_merged_memory_region when doing gathers.
- NV GPUs have a tendancy to be off by a very small margin, breaking rendering when greaterThan/lessThan checks are used.
- NOTE: Currently this setting is using the sRGB flag which indicates 8-bit output.
Only one game is currently known to care about this behaviour so this is good enough for now.
* Update asmjit dependency (aarch64 branch)
* Disable USE_DISCORD_RPC by default
* Dump some JIT objects in rpcs3 cache dir
* Add SIGILL handler for all platforms
* Fix resetting zeroing denormals in thread pool
* Refactor most v128:: utils into global gv_** functions
* Refactor PPU interpreter (incomplete), remove "precise"
* - Instruction specializations with multiple accuracy flags
* - Adjust calling convention for speed
* - Removed precise/fast setting, replaced with static
* - Started refactoring interpreters for building at runtime JIT
* (I got tired of poor compiler optimizations)
* - Expose some accuracy settings (SAT, NJ, VNAN, FPCC)
* - Add exec_bytes PPU thread variable (akin to cycle count)
* PPU LLVM: fix VCTUXS+VCTSXS instruction NaN results
* SPU interpreter: remove "precise" for now (extremely non-portable)
* - As with PPU, settings changed to static/dynamic for interpreters.
* - Precise options will be implemented later
* Fix termination after fatal error dialog
- This is just not part of spec, there is no enforcement for multiple of block size for width or height of s3tc compressed images.
- This restriction does indeed exist for ASTC and ETC but we're not using those formats.
Optimize CPU capability tests for arch-tuned builds.
Separate streaming and non-streaming utilities.
Rewritten copy_data_swap_u32(_cmp) with AVX2 path.
- Replace old format-based detection with proper aspect test.
Explicit image aspect has been available for a long time, but older
code was not updated.
- Specifically fixes a corner case where double transforms are required.
Technically this can be made more readable using transformation matrices:
* M1 = transform_virtual_to_physical()
* M2 = transform_image_to_virtual()
* M3 = M1 * M2
* Result = Input * M3
But we don't use a CPU-side matrix library and it is not reasonable to do this on the GPU.
- Just use RGB565 for all blit targets. Avoids really dumb transforms done by GPU hw.
- When X16 is used, all the channels get written to R channel alone. CmdBlit does perform format conversion!
- gl: Force image copy when blit is requested with compatible targets. Avoids format conversion issues.
- Partial clears either in active clear channels or scissor region must get barrier inserts to load previous data.
- Fixes some incorrectly discarded data during clear where data in untouched/uninitialized channels is lost.
- Mainly affects nvidia where x/w * w can sometimes return a value smaller than x.
In such conditions, floor(x) will return x-1 if x is an integer which is horribly wrong and exaggerates minor precision drift to great proportions.
Depending on the dpi settings, the debug overlay was almost unreadable.
I also took the liberty to refactor some redundant client size calls and to add some margin to the left of the debug text.
