Thanks to @asLody for optimizing this function. This raised the focus that this function should be optimized more.
The current table assumes that the host GPU is able to invert for free, so only AND,OR,XOR are accumulated in the performance metrik.
Performance results:
Instructions
0: 8
1: 30
2: 114
3: 80
4: 24
Latency
0: 8
1: 30
2: 194
3: 24
This falls back to the old approach of using a virtual buffer.
Windows is untested, but this build should fix support for Windows < 10 v1803. However without fastmem support at all.
* Wrong alignment in u64 LOG_DEBUG -> memcpy.
* Huge shift exponent in stride calculation for linear buffer, unused result -> skipped.
* Large shift in buffer cache if word = 0, skip checking for set bits.
Non of those were critical, so this should not change any behavior.
At least with the assumption, that the last one used masking behavior, which always yield continuous_bits = 0.
We just create one memory subsystem. This is a constant all the time.
So there is no need to call the non-inlined parent.Memory() helper on every callback.
This code was used to switch the CPU ID on thread switches.
However since "hle: kernel: multicore: Replace n-JITs impl. with 4 JITs.", the CPU ID is not a constant.
This has been dead code since this rewrite, and dropped in dynarmic as well. So there is no need to keep it.
We had used conan for opus before, but there was a bug in the AVX detection.
However we still had the Findopus.cmake file within the repository, but not used.
This patch reenables the Findopus helper and prefer the system wide installation of opus.
Else the fence might get submited out-of-order into the queue, which makes testing them pointless.
Overhead should be tiny as the mutex is just moved from the queue to the writing code.
This was implicitly done by `is_powered_on = false`, however the explicit method allows us to block until the GPU is actually gone.
This should fix a race condition while removing the other subsystems while the GPU is still active.
It shall block until there is something to consume in the queue.
And use it for the GPU emulation instead of the spin loop.
This is only in booting the emulator, however in BOTW this is the case for about 1 second.
Advantage: Altering the handler does not need a full recompilation.
Disadvantage: noreturn is droped, so the caller is a bit slower.
We quite often run yuzu with a YOLO assertion handler. In fact, only very few
games run at all with asserts. This patch allows developers to patch the handler
without recompiling everything. The overhead of the missing "noreturn" attribute
shoul be negletable.
Neither core nor web_services use OpenSSL nor LibreSSL.
However they need to link them as it's a requirement of httplib.
So let's declare this within httplib instead of core and web_services.
module._memory was already moved over to a new shared_ptr.
So code_memory_size was not increased at all.
This lowers the heap space and so saves a bit of memory, usually between 50 to 100 MB.
This fixes a regression of c0a01f3adc
This function is called rarely and blocks quite often for a long time.
So don't waste power and let the CPU sleep.
This might also increase the performance as the other cores might be allowed to clock higher.
As fetching command list headers and and the list of command headers is a fixed 1:1 relation now, they can be implemented within a single call.
This cleans up the Step() logic quite a bit.
Fetching every u32 from memory leads to a big overhead. So let's fetch all of them as a block if possible.
This reduces the Memory::* calls by the dma_pusher by a factor of 10.
On invalidating the streaming buffer, we need to reupload all vertex buffers.
But we don't need to reconfigure the vertex format.
This was a (silly) misstake in #1723.
Thanks at Rodrigo for discovering the issue.
Fun fact, as configuring the vertex format also invalidate the vertex buffer,
this misstake had no affect on the behavior.
They were missed, and Copy is very high in profile here. It doesn't block the GPU,
but it stalls the driver thread. So with our bad GL instructions, this might block quite a while.
This was created with the unfinished resampling PR in mind.
As the resampling is now on the audio thread, we don't need to care about this here any more.
Those implementations are quite costly, so there is no need to inline them to the caller.
Ressource deletion is often a performance bug, so in this way, we support to add breakpoints to them.
We uploaded the wrong data before. So the offset on the host GPU pointer may work for the first vertices, the last ones run out bounds.
Let's just offset the upload instead.
It allows us to use texture views and it reduces the overhead within the GPU driver.
But it disallows us to reallocate the texture, but we don't do so anyways.
In the end, it is the new way to allocate textures, so there is no need to use the old way.
This virtual function is called in a very hot spot, and it does nothing.
If this kind of feature is required, please be more specific and add callbacks
in the switch statement within Maxwell3D::WriteReg. There is no point in having
another switch statement within the rasterizer.
This is called ~3k times per frame in SMO ingame.
My laptop spends ~3ms per frame on allocating and freeing this string.
Let's just stop printing this kind of redundant information.
This patch caches VAO objects instead of re-emiting all pointers per draw call.
Configuring this pointers is known as a fast task, but it yields too many GL
calls. So for better performance, just bind the VAO instead of 16 pointers.
The idea of this cache is to avoid redundant uploads. So we are going
to cache the uploaded buffers within the stream_buffer and just reuse
the old pointers.
The next step is to implement a VBO cache on GPU memory, but for now,
I want to check the overhead of the cache management. Fetching the
buffer over PCI-E should be quite fast.
The std::string generation with its malloc and free requirement
was a noticeable overhead. Also switch to an ordered_map to
avoid the std::hash call. As those maps usually have a size of
two elements, the lookup time shall not matter.
This OpenGL call synchronize the worker thread of the nvidia blob.
It can be verified on linux with the __GL_THREADED_OPTIMIZATIONS=1 environment variable.
Those errors should not happen on tested drivers.
It was used as a workaround for https://bugs.freedesktop.org/show_bug.cgi?id=94148
Markus Wick