I forgot to finish CSubmitScheduler synchronization in
7c84c23114. Even ++m_FrameID was
forgotten. And it worked at the time. A proper synchronization is
described in renderer/backend/Vulkan/SwapChain.h.
It's better to construct a js-array from a `JS::RootedValueVector`.
Because it is more strongly typed and the index doesn't has to be
specified when appending an element.
Some usages are replaced with `JS::RootedValueArray`.
Fixes: #8702
Currently we always choose the best device. But it's not always
desirable. A more safe approach is to use the default device (with
index 0). The only downside of that is if a user didn't adjust
settings then the game might run on an integrated GPU instead of a
discrete one. In the future it'll be solved by selecting GPU in
options: #8529Fixes#8455
Now it's possible to destroy the old swapchain before creating a new
one. It might make the swapchain creation a bit slower but with a
lower memory peak.
With C++20 the custom container PS:span, which was a backport of
std::span is no longer needed.
Signed-off-by: Ralph Sennhauser <ralph.sennhauser@gmail.com>
Make include-what-you-use happy with some files in source and fix what
needs to be fixed.
Add markers to precompiled.h header includes to avoid
include-what-you-use wanting them to be removed.
Ref: #8086
Signed-off-by: Ralph Sennhauser <ralph.sennhauser@gmail.com>
Make include-what-you-use happy with files in source/renderer and
fix what needs to be fixed.
Ref: #8086
Signed-off-by: Ralph Sennhauser <ralph.sennhauser@gmail.com>
Make include-what-you-use happy with some files in source/renderer and
fix what needs to be fixed.
Ref: #8086
Signed-off-by: Ralph Sennhauser <ralph.sennhauser@gmail.com>
Goal
----
Avoid corrupting the dynamic font-atlas on Vulkan by blocking any
re-uploads until the command buffer that created / last updated the
texture has actually been submitted.
What changed
------------
* **Queue-aware textures**
* Added `queueSubmitAware` flag to `IDevice::CreateTexture*` APIs.
* `Vulkan::CTexture` now stores two booleans:
- `m_QueueSubmitAware` – opt-in per texture.
- `m_PendingQueueSubmit` – set to *true* the moment an upload
is recorded, cleared once the submit scheduler has flushed.
* `CRingCommandContext::ScheduleUpload` marks the texture as
pending (`SetPendingQueueSubmit(true)`).
* **Device-side watcher**
* `Vulkan::CDevice` keeps a
`m_TextureUploadWatcherQueue`. Each frame it checks textures that
were uploaded ≥ `NUMBER_OF_FRAMES_IN_FLIGHT` frames ago and
clears their pending flag.
* New helpers
`ScheduleTextureUploadWatch `,
`ProcessTextureUploadWatchQueue()`.
* **Font code**
* Atlas texture is now created with
`queueSubmitAware = true`.
* `CFont::UploadTextureAtlasToGPU()` early-outs when
`IsPendingQueueSubmit()` returns *true*, instead of tracking a
submit-handle or the manual `m_IsLoadingTextureToGPU` flag
(removed).
Why this is better
------------------
The logic to wait for a flush is localised inside the rendering
backend, so `CFont` only needs to ask *“is my texture busy?”*.
This removes the fragile submit-handle bookkeeping and works even if
the scheduler issues multiple submits per frame in future.
Result
------
Atlas uploads are deferred until the previous submit completes,
eliminating the intermittent glyph corruption on the Vulkan backend
while leaving GL and the dummy backend unchanged.
Make include-what-you-use happy with some files in source/renderer and
fix what needs to be fixed.
Ref: #8086
Signed-off-by: Ralph Sennhauser <ralph.sennhauser@gmail.com>
Some have their editor configured to remove trailing whitespace and
editing such a file would "fix" it, adding an unrelated change.
Fix whitespace violations excluding third party libs and generated files
like glad or patches.
Enable pre-commit hook trailing-whitespace to enforce it in the future.
Fixes: #8016
Signed-off-by: Ralph Sennhauser <ralph.sennhauser@gmail.com>
According to our stats `GL_ARB_fragment_program_shadow` is supported
by all our GL users. Also we're going to remove OpenGL ARB in A29. So
I remove ARBShadersShadow completely.
RPI4 returns `false` for `textureCompressionBC` because it doesn't
support formats above BC3. As the Vulkan specification requires to
support all BC formats to have `true` for `textureCompressionBC`.
According to #6936 some lower devices (especially virtual ones) might
report more memory than regular ones. So we can't use the memory
amount as a score for now.
The idea is similar to the storage images but we need a separate
descriptor set in Vulkan and a program interface to gather used buffer
in GL.
For Vulkan we also need to track buffers to free used descriptor sets.
Vladislav Belov