Closes#17005
Release Notes:
- Improved GPU context management: share a single context with multiple
surfaces.
### High Level
Blade got a proper support for Surface objects in
https://github.com/kvark/blade/pull/203.
That was mainly motivated by Zed needing to draw multiple windows. With
the Surface API, Zed is now able to have the GPU context tied to the
"Platform" instead of "Window". Practically speaking, this means:
- architecture more sound
- faster to open/close windows
- less surprises, more robust
### Concerns
1. Zed has been using a temporary workaround for the platform bug on
some Intel+Nvidia machines that makes us unable to present -
https://github.com/kvark/blade/pull/144 . This workaround is no longer
available with the new architecture. I'm looking for ideas on how to
approach this better.
- we are now picking up the change in
https://github.com/kvark/blade/pull/210, which allows forcing a specific
Device ID. This should allow Zed users to work around the issue. We
could help them to automate it, too.
2. ~~Metal-rs dependency is switched to
https://github.com/kvark/metal-rs/tree/blade, since upstream isn't
responsive in merging changes that are required for Blade. Hopefully,
temporary.~~
- ~~we can also hack around it by just transmuting the texture
references, since we know those are unchanged in the branch. That would
allow Blade to use it's own version of Metal, temporarily, if switching
metal-rs in the workspace is a concern.~~
- merged my metal-rs changes and updated Zed to use the upstream github
reference
---------
Co-authored-by: Mikayla Maki <mikayla@zed.dev>
Co-authored-by: Mikayla Maki <mikayla.c.maki@gmail.com>
Co-authored-by: Conrad Irwin <conrad.irwin@gmail.com>
Release Notes:
- N/A
Picks up https://github.com/kvark/blade/pull/118Fixes#10351
Seeing that Zed loaded with Blade repository is consuming 260Mb of RAM.
We can tune this to be lower, but ultimately it doesn't matter: this
memory isn't wasted, it's just pools for memory and descriptors, which
may be used by bigger and more complex views of Zed.
Release Notes:
- N/A
## What
Addresses a long-standing issue of doing the blending operations in sRGB
space. Currently, the input HSL colors are provided in sRGB space and
converted to linear in the vertex shader. Conversion back to sRGB, which
is required on most platforms today, happens at the very end of the
pipeline when writing into sRGB render target.
Note-1: in the future we may consider doing HSL -> sRGB -> Linear
transform on CPU before feeding into shaders. However, I don't expect
any significant difference here given that we are likely bound by fill
rate and pixel shaders, anyway.
Note-2: the graphics stack is programmed to detect if the platform
supports presenting in linear color space and avoids converting to sRGB
at the end in this case. However, on my Z13 laptop this isn't supported
by the RADV driver.
Closes#7684Closes#11462
@jansol please confirm if you can!
## Comparison
Screenshot of the Glazier theme before the change:
Same theme after the change:
Release Notes:
- N/A
Picks up https://github.com/kvark/blade/pull/113 and a bunch of other
fixes.
Should prevent the exclusive full-screen on Vulkan - related to #9728
cc @kazatsuyu
Note: this PR doesn't enable transparency, this is left to follow-up
Associates every window with its own refresh event. Removes the use of
X11 present.
Alternative to #8592.
Instead of doing the rendering on idle and then involving a hack for
polling X11 events, this PR just tries to do the rendering inside the
main loop. This guarantees that we continue to poll for events after the
draw, and not get screwed by the driver talking to X11 via the same file
descriptor.
Release Notes:
- N/A
Following-up on #8330
Invocation
```bash
cargo-machete --with-metadata --skip-target-dir --fix
````
There is more stuff to fix, but it chokes on `async-lock`:
```
cargo-machete found the following unused dependencies in /x/Code/zed:
rpc -- /x/Code/zed/crates/rpc/Cargo.toml:
async_lock
prost_build
serde_derive
Error: Dependency async_lock not found
```
Release Notes:
- N/A
With this change, interaction with Zed is actually real-time and usable
🚀🎉
The gist of it is - trying to process all of the input events before
rendering anything.
Release Notes:
- N/A
**Note**: this can be further improved in a follow-up.
Currently, once the input and runnables are processed, we'd try to draw
+ render a frame.
Presentation starts with acquiring a new frame. We currently have FIFO
presentation method, so acquiring a frame is blocking on that swapchain
image to become available. As the result, presentation takes around 16
ms, most of which is just busy wait.
Ideally, we'd be able to process more input in this time frame, instead.
**Note2**: it's a bit laggy in Debug for me, but that's just because of
the extra-long `draw` times, which is unrelated to rendering (or
platform support, for the matter). I'm curious how come on MacOS the
`draw()` times in Debug are more modest.
[Profiling](https://crates.io/crates/profiling) crate allows easy
integration with various profiler tools. The best thing is - annotations
compile to nothing unless you request a specific feature.
For example, I used this command to enable Tracy support:
```bash
cargo run --features profiling/profile-with-tracy
```
At the same time I had Tracy tool open and waiting for connection. It
gathered nice stats from the run:
Release Notes:
- N/A
With `send_and_check_request` we'd be blocking both the main loop and
the caller. `send_request` is only going to be blocking on the main loop
when processing the request.
Release Notes:
- N/A
Based on a flamegraph from `perf`/`hotspot`, we are spending 40% of time
redrawing, another 40% of time downloading stuff (i.e. rust toolchain),
and the rest on text rendering, layout and such. This is with Vulkan
Validation (see https://github.com/zed-industries/zed/pull/8044).
I'm also wondering if it would be better with #7758, but regardless we
should have no problem rendering at 60-120 fps and processing user
input. More follow-ups are expected here.
Generally the BladeAtlas logic has been deferring all the texture
initializations and updates till `begin_frame`. This doesn't work for
path rasterization, since a texture needs to be allocated after
`begin_frame` and used immediately.
Fixed validation error:
> UNASSIGNED-CoreValidation-DrawState-InvalidImageLayout(ERROR / SPEC):
msgNum: 1303270965 - Validation Error: [
UNASSIGNED-CoreValidation-DrawState-InvalidImageLayout ] Object 0:
handle = 0x60ce301b9010, name = main, type =
VK_OBJECT_TYPE_COMMAND_BUFFER; Object 1: handle = 0x51820000000007b,
name = atlas, type = VK_OBJECT_TYPE_IMAGE; | MessageID = 0x4dae5635 |
vkQueueSubmit(): pSubmits[0].pCommandBuffers[0] command buffer
VkCommandBuffer 0x60ce301b9010[main] expects VkImage
0x51820000000007b[atlas] (subresource: aspectMask 0x1 array layer 0, mip
level 0) to be in layout VK_IMAGE_LAYOUT_GENERAL--instead, current
layout is VK_IMAGE_LAYOUT_UNDEFINED.
Objects: 2
[0] 0x60ce301b9010, type: 6, name: main
[1] 0x51820000000007b, type: 10, name: atlas
Release Notes:
- N/A
Depends on https://github.com/zed-industries/font-kit/pull/2 and
https://github.com/kvark/blade/pull/77
This change enables Blade to be also used on MacOS. It will also make it
easier to use it on Windows.
What works: most of the things. Zed loads as fast and appears equally
responsive to the current renderer.
<img width="306" alt="Screenshot 2024-02-11 at 12 09 15 AM"
src="https://github.com/zed-industries/zed/assets/107301/66d82f45-5ea2-4e2b-86c6-5b3ed333c827">
Things missing:
- [x] video streaming. ~~Requires a bit of plumbing on both Blade and
Zed sides, but all fairly straightforward.~~
- verified with a local setup
- [x] resize. ~~Not sure where exactly to hook up the reaction on the
window size change. Once we know where, the fix is one line.~~
- [ ] fine-tune CA Layer
- this isn't a blocker for merging the PR, but it would be a blocker if
we wanted to switch to the new path by default
- [ ] rebase on latest, get the dependency merged (need review/merge of
https://github.com/zed-industries/font-kit/pull/2!)
Update: I implemented resize support as well as "surface" rendering on
the Blade path (which will be useful on Linux/Windows later on). I
haven't tested the latter though - not sure how to get something
streaming. Would appreciate some help! I don't think this should be a
blocker to this PR, anyway.
The only little piece that's missing for the Blade on MacOS path to be
full-featured is fine-tuning the CALayer configuration. Zed does a lot
of careful logic in configuring the layer, such as switching the
"present with transaction" on/off intermittently, which Blade path
doesn't have yet.
Release Notes:
- N/A
---------
Co-authored-by: Mikayla <mikayla@zed.dev>
Alternative to #7758, which doesn't involve adding a new trait method
`request_draw`.
Somehow, my whole screen goes blinking black with this when moving the
window, so not ready for landing.
Release Notes:
- N/A
---------
Co-authored-by: Mikayla Maki <mikayla@zed.dev>
Fix found by @h3mosphere (thanks!)
Solves the Vulkan validation on start on some platforms about the
mismatched surface size, e.g.
```
VUID-VkSwapchainCreateInfoKHR-imageExtent-01274(ERROR / SPEC): msgNum: 2094043421 - Validation Error: [ VUID-VkSwapchainCreateInfoKHR-imageExtent-01274 ] Object 0: handle = 0x55dff99554c0, type = VK_OBJECT_TYPE_DEVICE; | MessageID = 0x7cd0911d | vkCreateSwapchainKHR() called with imageExtent = (1920,1080), which is outside the bounds returned by vkGetPhysicalDeviceSurfaceCapabilitiesKHR(): currentExtent = (1920,1016), minImageExtent = (1920,1016), maxImageExtent = (1920,1016). The Vulkan spec states: imageExtent must be between minImageExtent and maxImageExtent, inclusive, where minImageExtent and maxImageExtent are members of the VkSurfaceCapabilitiesKHR structure returned by vkGetPhysicalDeviceSurfaceCapabilitiesKHR for the surface (https://www.khronos.org/registry/vulkan/specs/1.3-extensions/html/vkspec.html#VUID-VkSwapchainCreateInfoKHR-imageExtent-01274)
```
Release Notes:
- N/A
Co-authored-by: Mikayla Maki <mikayla@zed.dev>
It allows us to receive messages from the dispatcher,
which breaks us out of waiting and lets us execute
main thread runnables as a part of the main loop.
Link error:
= note: /nix/store/idiaraknw071d20nlqp49s18gbvw4wa0-binutils-2.40/bin/ld: /x/Code/zed/target/x86_64-unknown-linux-gnu/debug/deps/libtree_sitter_haskell-7323f782ad886c6d.rlib(scanner.o): in function `state_new':
/home/kvark/.cargo/git/checkouts/tree-sitter-haskell-74c278e7a2ef8d7d/cf98de2/src/scanner.c:218: multiple definition of `state_new'; /x/Code/zed/target/x86_64-unknown-linux-gnu/debug/deps/libtree_sitter_purescript-b0a95fb604a5817c.rlib(scanner.o):/home/kvark/.cargo/git/checkouts/tree-sitter-purescript-88dd3ec656c48026/a37140f/src/scanner.c:218: first defined here
Dzmitry Malyshau
