This extension is currently required to get good performance.
So instead of maybe triggering a GPU hang, let's just disable the shader and show a warning instead.
Sadly our postprocessing framework does not support reporting a warning, so the way to display the warning is a bit hacky.
But this is an ascii-art shader, we already have a font in the shader, so let's use it.
Dolphin had a nice but horrible slow asci art shader.
Beside being slow, it used the font from dolphin's old OSD implementation.
So it was also broken after the VideoCommon redesign.
This patch reintroduces the (almost) identical method for the asci art generation.
However with many improvements:
* Inline definition of the font, so no external dependency any more
* Optimized performance by partial unrolling with CSE'ing some memory loads
* Shader subgroup instructions, so 32 times faster on desktop CPUs on Vulkan + OGL (no D3D support)
* Option to select character size by internal or window resultion
* Dropped the last row of pixels of each character (only modification of the outcome)
While the NV extension is totally fine, the KHR extension should be able to support more hardware.
For NVIDIA, the hardware either supports both or neither, it just needs a driver from the last two years.
For AMD, the drivers from late 2022-12 seems to bring support for the KHR extension.
For Intel, the KHR is also supported for some years.
We must not provide the /Externals directory as global include directory.
Here, this yield a crash because of external minizip header and system library mismatch.
Soundtouch itself recormends to include it with <SoundTouch.h> and -I/usr/include/soundtouch, so this should fit better.
This was a huge speedup with disabled fastmem, but it still requires the fastmem arena.
So let's disable it for now, even if this commit has a huge performance hit with disabled fastmem.
This will help to disable all inter-instruction dependencies.
So android users can check if only a single instruction is broken without compiling dolphin on their own.
psq_l with QUANTIZE_FLOAT does not use the FPU, so it does not trim the precision of the u32 input data.
We already have the helper ConvertToDouble for floating point u32->u64 convertion used in lfs, so let's use it here as well.
The idea of this code was to not unroll loops, but it was completely broken.
So we've unrolled all loops, but only up to the second iteration.
Honestly, a better check would test if we branch to code which is already in the compiling block. But this is out of scope for now.
But testing shows that this unrolling actually improve the performance. So instead of fixing this bug, this check can be dropped.
This register is defined as "optional reserved" within the aarch64 ABI.
Linux doesn't use it, but we must not modify it on ios or windows.
As we have plenty of registers on aarch64, let's just always skip this one.
Coherent mappings have a lower overhead and less GL codes.
So enables coherent mapping by default for all drivers.
Both Qualcomm and ARM performs very bad with explicit flushing, so this change helps them as well.
AFAIK there was one GPU generation which was slower on coherent mapping: nvidia tesla
So Geforce 200 and 300 series should be tested with this PR before merging.
As this was last tested many years ago, this issue might have been fixed as well.
Those GPUs are close to 10 years old and not supported any more by nvidia.
We have to decide if we want to link to
libGL.so -- legacy
libOpenGL.so -- GLVND
As we use xxxGetProcAddress, we don't care about feature
levels of the linked library, so we're very fine with the
new way.
Yes, this commit is only to blame OSX and Mali. Through the former supports unsynchronized mappings, the latter supports *no* way to stream dynamic data at all. Let's try to make bad news, as they ignore friendly feature requests. Maybe we just need to make more noise...
CNTVCT_EL0 is force-enabled on all linux plattforms.
Windows is untested, but as this is the best way to get *any* low
overhead performance counters, they likely use it as well.
Within Cleanup(), it is called at *every* end of the block. This generates bigger code,
but it is the only way to handle blocks with multiple exit nodes.
Seems like I was wrong that ANDI2R doesn't require a temporary register here.
There is *one* case when the mask won't fit in the ARM AND instruction:
mask = 0xFFFFFFFF
But let's just use MOV instead of AND here for this case...
The accuracy doesn't match ppc, and worse, it doesn't set the error flags if the input is zero.
Lets stop to ship broken instructions, so right now, the interpreter is the closest one.
New Super Mario Bros on PAL still renders at 60 fps, but skips every 5th XFB copy.
So our detection of "per frame" fails, and we require twice the amound of texture objects.
But our pool frees unused textures after 3 frames, so half of them needs to be reallocated
every few frames.
This commit removes the lock for render targets. It was introduced to not update a texture
while it is still in use. But render targets aren't updated while rendering, so this
lock isn't needed. Non-rendertarget textures however aren't as dynamic, so the lock should
have no performance update.
degasus