Should we ever introduce anything else that has to be done when a command
buffer is executed (e.g. invalidating constants from previous commit), we
don't have to update all the callers.
This could happen because the vertex memory was already committed, if a
uniform buffer allocation failed and caused a command buffer to be
executed, it would be associated with the previous command buffer rather
than the buffer containing the draw that consumed these vertices.
This was happening when a fence wait happened mid-frame. The data written
between the fence being queued and the allocation occuring was incorrectly
assumed to be consumed by the GPU.
This is mainly for potential Android fifoci usage, and thus is not
exposed anywhere in the UI. To enable, set DumpFramesAsImages under
Settings in GFX.ini.
Instead of resetting two command buffers, now we only have to call
vkResetCommandPool once at the start of a frame.
NV's recommends using one pool per frame/thread. May offer a very small
boost in performance on some systems.
anv seems to set this to zero, which is fine according to the spec, but
we were using it as a maximum, which was resulting in a swap chain
without any buffers being created.
There's not a lot of point in passing these around or storing them
(texture cache/state tracker mainly) as there will only ever be a single
instance of the class.
Also adds downcast helpers such as Vulkan::Renderer::GetInstance().
This is an issue because a driver may have to maintain two copies of a
texture if it batches all uploads together at the start of a frame.
In the Vulkan backend, we do something similar to avoid breaking out of a
render pass to copy a texture from the streaming buffer to the destination
image.
This was causing issues in the sms-bubbles fifolog, where an EFB copy to
the same address of a previously-used texture caused the previous texture
to be re-used again for a different image later on in the frame, causing
the original contents to be discarded.
This happened when the source texture was an EFB copy, therefore it had
not been populated prior to the draw command buffer being executed, and
the conversion was occurring in the init command list.
Should fix a possible reference to deleted framebuffers, as well as fixing
the issues with the render area being correct if the game's source area
changes, or auto-scaling is enabled.
This is needed because for some reason the WSI for NV Vulkan drivers
doesn't return VK_ERROR_OUT_OF_DATE_KHR, so there is no other way to know
that a resize has occured apart from polling, which is a poor solution for
X11 (since it is blocking).
Also swaps the byte order from RGBA->BGRA to match GL/D3D12, and what
the read handler is expecting.
Depth reads will now return the minimum depth of all samples, instead of
the average of all samples.
Using glMapBufferRange to read back the contents of the SSBO is extremely
slow on NVIDIA drivers. This is more noticeable at higher internal
resolutions. Using glGetBufferSubData instead does not seem to exhibit
this slowdown.
Sorts out references that cause some modules to be kept around after
backend shutdown.
Should also solve the issue with errors being thrown due to the config
being loaded after device creation, leading to the incorrect device being
used in a multi-adapter system.
Moves render target restoring to RestoreAPIState, this also means no need
to manually restore after allocating in a buffer that caused execution,
because the manager restores it for us.
Remove a method that wasn't used from D3DUtil.cpp, and fixes a few errors
in EFB poke drawing.
This is not optimal, but for those texture packs with extremely large
images, it won't crash. Releasing after the frame completes is an option
too, however, there is the risk of running out of memory by doing this.
Drivers that don't support GL_ARB_shading_language_420pack require that
the storage qualifier be specified even when inside an interface block.
AMD's driver throws a compile error when "centroid in/out" is used within
an interface block.
Our previous behavior was to include the storage qualifier regardless, but
this wasn't working on AMD, therefore we should check for the presence of
the extension and include based on this, instead.
This applies to callers that do not have full knowledge of the command
list state, and thus, cannot restore it should allocations cause command
list execution. Instead we reallocate a new buffer. Should not happen
often enough for this to be a concern, as it's mainly for the utility
classes.
Stenzek