Increase the size of the bin header and path buffers to accommodate the actual extents accessed by the shaders.
Also add a missing workgroup barrier, not essential to fixing this bug but found while auditing the code.
Fixes#222
We need to reduce the number of buffer bindings to 8 so it can run on
all WebGPU devices. The best candidate was to combine info and bin_data,
which are written by two different stages (draw_leaf and binning,
respectively), both read by coarse, and are unstructured, so the only
real shader change needed is to add the offset for the binning data.
I thought I was going to have to do a blit to make it fit, but
fortunately that wasn't needed.
Progress to #202
This commit reduces the workgroup shared memory of binning to fit in
16k (by packing two u16's in a u32), and moves the config binding to
uniform, from readonly storage.
Progress toward #202
* add comment for ClipInp::path_ix
* update comment for SceneBuilder drawtags
* remove copy of blend data through info buffer. Instead, update EndClip's DrawMonoid::scene_offset to point to the same scene data of the associated EndClip.
* Add comment about avoiding NaNs to inverse alpha divisions
* Change inverse alpha computations to use max(a, epsilon) rather than +
* Emit alpha component in blit fragment shader (enables support for partially transparent windows)
This just a collection of minor fixes to various things I noticed while looking over the code.
Adds a `Device::poll` call to the winit example.
Removes some lingering unnecessary `<f32>`s in the blit shaders.
Removes a duplicated bbox entry in the shared shader inclusion code and collapses nested `concat!`s.
Finally, it reintroduces the computation of separated alpha in the output of fine which I removed in a previous PR. This allows the `render_to_texture` mode to be useful for generating images that can be properly blended over other content. The blit shader has been changed to accommodate this.
A number of things weren't being computed correctly for strokes,
including bounding boxes and also incorrectly applying y_edge logic
(which is only correct for fills).
This creates a new Renderer type that offers fairly simple render_to_texture and render_to_surface methods, the latter of which handles the final blit internally. Also adds a util module with some helpers for device and surface creation.
There is a new winit example in piet-wgsl/examples/winit that shows how to make use of it all. This should be fairly trivial to adapt to glazier/xilem.
* add writeable image support to engine
* add external resources to engine
* update fine to output to a texture
* copy over original piet-gpu test scenes
* put it all in a pretty (resizable!) window
Now that wgsl-analyzer 0.6 is released, most explicit type annotations on vec can be dropped (the exception being when it is a type conversion).
Also changes mix to select when the selector is actually boolean.
This adds a new dependency on peniko, reintroduces kurbo for geometry (!), removes the now defunct types from piet-scene and updates all the test scenes to use the new types.
As stated by the README, the intention is for the shaders to be
provided under the terms of the Unlicense in addition to the MIT
and Apache-2.0 licenses.
While here, add a self-contained UNLICENSE file to replace the
external reference to unlicense.org.
This is a checkpoint of partly completed work. Much of the GPU side is done, very little of the CPU side.
For clips, the clip_els bindings (binding 6 of draw_leaf) are not added. Clip logic is missing from coarse. The overflow buffer is missing from fine, as is its size calculation in coarse (but it should work as long as the max depth fits within BLEND_STACK_SPLIT).
For gradients, the texture binding is missing (binding 6) is missing from fine, as is the infrastructure in engine to deal with texture resources, and of course porting over the logic to fill it.
The code is not tested, bugs may lurk.
Note that this is evidence in favor of https://github.com/gfx-rs/naga/issues/2098 - my code is actually wrong, and it was caught by trying to port it to run in Chrome Canary.
Still one flaw, fat lines aren't expanded with strokes in path coarse rasterization. But that's a small visual ding, and can be fixed
That said, there is some really strange stuff going on in tile_alloc. It's using storage to do a uniform broadcast (the result of bump allocation for the workgroup), which is not great at all. It should be using workgroup storage, but on my mac it behaves as if the workgroup barrier is not in place. Investigating.
