`draw` had three branches, one per shader, which is the dynamic dispatch this
was asking for. A primitive now brings a `PrimitiveRender`: it states the
layout its shader reads, uploads whatever it owns, and records its own draws.
`GlyphRender` owns the atlas and binds it once for a list; `ImageRender` owns
the images and binds one per instance; the default owns nothing and draws them
all in one call. The renderer sets the pipeline, the shared group, the list's
data and the vertex buffer, and knows nothing else about what it is drawing.
Measured before committing to it, since dispatch per list is the cost. Wall
time on this machine swings 2x between runs of one binary, so the comparison
is instructions retired, which is stable to 0.1%: at 256 layers drawing 8
rects, 8 glyphs and 2 images each, 7.4074e9 against 7.4145e9, and at 1024
layers 27.467e9 against 27.498e9. Both are 0.1%, which is 6 instructions per
list drawn -- one indirect call. Recording a list into the pass costs wgpu
about 5,400.
`tests/draw_cost.rs` is that measurement, kept.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`Primitive::SAMPLES` says what a primitive samples and how often it has to be
bound: `Atlas` once for a list, `Image` for one instance alone. Group 2 is
that, whichever it is, so a rect's pipeline has no texture and no sampler in
its layout and the prelude hands out neither.
The two differ only in the view dimension and in what binds them, so they
share `sampled_layout` and `sampled_group`; `GpuPages` owns its bind group
again and rebuilds it when the array grows.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The window uniform and the mask array were still `None`, which is what puts a
binding on wgpu-core's late-sized list: `check_late_buffer_bindings` runs from
`is_ready` on every draw and compares each such binding's bound size against
the naga-derived minimum for the shader global. Stating the size filters the
binding out of that list, and moves the same comparison to bind group and
pipeline creation.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The update loop zips all three and asserts up front that no list is left
without a pipeline, instead of indexing the pipelines to get that guarantee.
The comment on the data layout claimed a `None` minimum takes its value from
the first pipeline built against the layout. That is not documented and does
not reproduce -- one shared layout with `None` renders the tabs example
correctly today. What is documented is that a stated size is checked when the
bind group and pipeline are created, and `None` is checked on every draw, so
that is what the comment says now.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`Painter::primitive` takes `impl PrimitiveLike`: a primitive, or something
that yields one and does whatever else drawing it needs. A `&TextureHandle`
yields a `TexturePrimitive` and retains its share on the way through, so
`texture`, `texture_within` and `texture_at` are gone and an image is drawn
like anything else.
I said last round that the blanket impl would collide with the one for
`&TextureHandle` under coherence. It does not: `Primitive` is ours, so no
crate can add the impl that would overlap, and rustc accepts both.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Nothing seeds the registry any more, so a kind's id is decided by the first
draw and no order within a layer can be relied on even by accident. A ui that
draws no images now pays for no image pipeline, and a layer's list vector only
reaches the highest kind that layer draws.
The atlas and the sampler are still bound for every draw, but are declared by
the two shaders that read them rather than by the prelude, which is now only
what every primitive uses.
`TexturePrimitive` gets a `From<&TextureHandle>`; `Painter::texture_at` stays
because the share of the handle is what keeps the slot from being freed while
it is drawn, which a `Pod` primitive cannot hold.
Checked on the headless rig that the layers carry the ordering rather than the
ids: with a bare text drawn first, so glyph registers before rect, a stacked
label still draws over its background. Also re-ran an image alone in a layer,
now the only primitive a ui registers, and a four-layer atlas.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A `Primitive` now carries its own WGSL, and `PrimitiveRegistry` keys ids by
`TypeId`, so `Painter::primitive` takes only the value and the `RECT`,
`GLYPH` and `TEXTURE` constants are gone. Registering is what a first draw
does; the built-ins are seeded up front so first-draw order cannot decide
anything about them.
What a primitive samples is no longer something every registration states.
The glyph atlas and the one sampler moved into the shared group, which is
where a mask texture would go too, so a rect's pipeline has no texture in
its layout at all. Only a primitive whose type sets `TEXTURE` gets an image
group, and that is also what records the slot at write time -- so nothing
reads a `u32` back out of the instance payload.
Verified on the headless rig: the tabs example, two images added at runtime,
an image alone in a layer, and text spanning a four-layer atlas after the
array grew twice.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`write_texture` differed from `write` by one argument, which is what the
generic parameter was already for, so textures register as a primitive
with a `TexturePrimitive` holding the slot. `write_texture`,
`InstanceKind` and the layer's separate texture list are gone;
`DrawLayers` is back to `write` and `free`, with the kind carried in
`PrimitiveInst` as it carries everything else.
What differs between a texture and a rect is only what it samples, so
that is what registration says: `PrimitiveTexture::Atlas` binds the
shared atlas once for the layer, `PerInstance` binds the texture its own
data names and draws one instance at a time. One loop over a layer's
lists, one match on that.
`Pod` is back to being a supertrait of `Primitive` rather than the bound
itself. The guarantee is that a `PrimitiveKind<P>` is only minted by
`register::<P>` and `write` takes the kind and the value together, so a
primitive always has a list of its own to go in and the write does not
check anything: a list takes its stride from the type it was made for
instead of inferring it from the first write and asserting on the rest.
Also from reviewing this: a layer's lists and their buffers are created
only when that layer draws that primitive, so a primitive nobody uses no
longer costs two buffers in every layer -- which matters more now the set
is open-ended. `ListBuffers::update` takes the two things it uses rather
than the whole pipeline.
Verified again over all five cases: an image alone in a layer, three
images added and one deleted, the masked text-edit tab, the text-layout
tab and the default tab.
Reviewing the previous fix, which was itself unreviewed. `update` gave
each list the bind group layout of the pipeline that draws it by zipping
the layer's lists against `self.primitives`, but built those pipelines
afterwards -- so on any pass where one did not exist yet the zip yielded
nothing, and those lists kept no bind group and drew nothing. Measured
on startup: four layers had content while `self.primitives` was still
empty. It only looked right because those layers were marked dirty again
on a later frame and rebuilt then.
`build_pipelines` now runs before the layers, and the pairing is indexed
rather than zipped, so a primitive drawn before it was registered panics
instead of silently leaving its list unbuilt.
Three defects, two of them invisible to every case I had run.
An image alone in a layer failed validation. The texture pipeline never
bound group 1, and every earlier case happened to have a rect in the
same layer, which left one bound from the primitive draw -- so the bug
was hidden by the tests passing.
A `min_binding_size: None` binding takes its minimum from the first
pipeline built against that bind group layout, so one shared group 1
layout held every primitive to the largest. Rect and glyph coexisted
only because glyph is the bigger of the two; the texture slots, at four
bytes, did not. Each primitive now gets its own layout with its entry
size stated, which is also why `PrimitiveRegistry` records the stride.
Because the pipeline layouts now differ per primitive, a pipeline change
drops the bound groups, so group 2 moves after `set_pipeline`.
An empty list still built a bind group over a buffer too small for one
entry, which the stated minimum would now reject. It gets no bind group,
and nothing draws it.
Also from the read-through: `UiRenderNode` kept a `Device` beside the
one `update` is handed, `PrimitiveRegistry::default` registered inside
an `assert_eq!`, and `mask_idx` was an unqualified integer varying where
`idx` beside it was `flat`.
The `primitives!` macro, `PrimitiveData`, `PrimitiveVec`,
`PrimitiveBuffers`, the `Primitive` trait and the shader's dispatch
switch are gone. A primitive is now a registration: its WGSL and,
implicitly, the size of the entry that WGSL reads. Everything else --
its instance list, its free list, its buffers, its bind group and its
pipeline -- follows from that, so adding one is a `register` call and a
shader file, with nothing per-type to remember and no cross-type
dispatch to extend.
Nothing dispatches dynamically. Push, free, renumber, upload and draw
are identical for every primitive; what differs is the entry size and
the pipeline, which are data. So `InstanceList` carries a runtime
stride and its instances' data as bytes, and one concrete type serves
every primitive and the textures. Measured against a typed list it
costs 0.2ns per write, where a trait object costs 1.6ns.
Because each type has its own list, an instance's index is also its
data index: `@builtin(instance_index)` replaces the `idx` field, the
`binding` field goes with the switch, and `PrimitiveInstance` drops from
28 bytes to 20. `PrimitiveVec`'s free list merges into the instance
list's, so an instance and its data are freed by one `swap_remove`
rather than two arenas kept in step.
`shader.wgsl` becomes `shader/prelude.wgsl` plus one file per primitive.
The prelude carries the window, masks, sampled texture, vertex shader
and `masked()`, and is compiled ahead of each primitive's own source --
which is also what a caller's own primitive would be. Masks move back
into group 0 beside the window uniform, since every pipeline shares one
layout.
Within a layer, types now draw in registration order: a rect under a
glyph under a texture. Order within a layer was never meaningful --
freeing an instance swaps another into its place -- so this replaces an
accident with a defined order, and backgrounds land under their content.
Verified with a headless run per case: text over its own rect and an
image over its own rect in one layer, a masked stack clipping, the
text-layout tab, and adding three images and deleting one.
Interleaving textures with primitives was solving a problem that does not
exist: within a layer, order is already undefined because freeing an
instance swap-removes it, and layering is what layers are for. So the run
batching is gone.
A layer is now `LayerDraws`: a `Primitives` and a texture `InstanceList`
side by side, with `updated` covering both. `Primitives` holds only
primitives again -- its instance list plus the group-1 data those
instances read -- and `InstanceList` is the shared push/free/apply_free
the two lists have in common rather than a second copy of it.
`PrimitiveHandle` names which list with `InstanceKind`, and
`PrimitiveChange` carries the same, since the two index independently.
The handle no longer carries a group-1 index at all: a primitive
instance already records where its entry is.
The renderer gives each layer a second instance buffer and draws its
textures one at a time after the instanced draw, each binding its own
group 2.
Review fixes alongside: `GlyphAtlas::allocate` returns the `PageUpload`
it reserved instead of a bare tuple; a page or image region uploads
through a new `write_region`, which passes the row stride to
`write_texture` rather than copying the rectangle out first.
Verified by replaying taps into the `tabs` example: three images added
and one deleted leaves two drawn with two live texture slots, and the
masked text-edit tab still clips, with the images freed on tab switch.
Review response: `GlyphInfo` goes back to two `vec2<f32>`, which is what
the uvs are, and `GlyphPrimitive` takes `#[repr(C, align(8))]` to match.
That leaves four padding bytes, which the `primitives!` macro's
`unsafe impl Pod` accepts. Measured at 32 bytes, align 8, the same as
WGSL's layout for the struct.
Rework of the review on #11. Pages and standalone images were one
`Textures` manager separated by a `TextureKind` tag, and images were a
second instance list beside `Primitives::instances`. The tag forced
`image_index()`/`layer()` to panic on the wrong kind of handle, and the
second list forced an `is_image` branch through `free`, `region_mut`,
`apply_free` and `PrimitiveChange`.
Pages are now their own thing. `GlyphAtlas` owns its page images
outright and hands the renderer dirty rectangles; `GpuPages` owns the
array texture they upload to. `Textures` is standalone images only, so
`TextureHandle` has one kind, `slot()` cannot be wrong, and nothing
needs a free list that skips pages. `GlyphAtlas::insert` no longer takes
a `Textures`, which drops that parameter from `TextData::render` and
`SizeCtx` too.
Images go back through the one instance list. A texture instance is an
ordinary `PrimitiveInstance` whose `idx` names a texture rather than a
group-1 entry, which `PrimitiveHandle::data_idx: Option` records.
`RenderLayer::plan_draws` batches the layer's instances into runs
sharing a bind group, so a ui with no images still plans a single draw,
and an image draws in instance order rather than on top of its layer.
Group 2 is now one `texture_2d_array` and a sampler, bound per run: the
atlas for rects and glyphs, or one image viewed as an array of one. That
removes the second texture binding and the 1x1 null view that had to
fill it. Masks move to group 3, so resizing that buffer no longer stales
every texture bind group, and `GpuTextures` no longer reports whether
the caller must rebuild one.
`GlyphPrimitive` drops its manual pad: the WGSL struct now declares the
uvs as scalars, which matches the Rust layout exactly. `#[repr(C,
align(8))]` would have left real padding bytes, which `bytemuck::Pod`
forbids.
Verified with a headless run of the `tabs` example and of a scratch
example mixing images, rects, glyphs and a mask in one layer; 52 glyphs
at size 300 grew the atlas array from 1 to 4 layers with every earlier
page still sampling correctly.
The renderer bound every texture through one
`binding_array<texture_2d<f32>>` indexed per primitive. That needs
`VK_EXT_descriptor_indexing`, which a real share of Android GPUs do not
have, so the shape did not run there at all.
Split the two things being bound, since they want opposite treatment:
- Glyph atlas pages become layers of one `texture_2d_array`. A glyph
primitive carries a layer rather than a view/sampler index pair, and a
layer index is an ordinary sampling operand -- no extension. Growing the
atlas recreates the array with headroom and copies the old layers across
GPU-side.
- A standalone image gets its own texture and its own bind group, and
draws in its own call. It no longer needs a per-instance entry in
`PrimitiveData` at all: the bind group has already picked the texture.
`Primitives` therefore keeps images in a list of their own, with
`PrimitiveChange::is_image` saying which list a renumbering belongs to --
the two have independent index spaces, so `(layer, inst_idx)` alone would
collide between them.
Verified on this machine's real GPU (Venus onto an RX 7900 XT, confirmed
by the loaded ICD rather than assumed): the `tabs` example renders
byte-identical screenshots before and after, both for a text-and-rect tab
and for one holding a standalone image.
`update` redrew everything when `resized` was set, but `needs_redraw` --
which is what decides whether to request a frame at all -- did not know
about `resized`. A condition in one and not the other is a frame nobody
asks for and a stale window. The two share one `needs_redraw_all` now.
Latent on Wayland, because winit requests a redraw after a resize by
itself; a resize changes neither the root nor any widget, so nothing else
here would have asked. It stops being latent on Android, where the
surface work will not have winit underneath it and every rotation and
keyboard open is a resize.
This is not a fix for the startup defect recorded in RUST.md, where the
window keeps its pre-configure layout: that reproduces with this change
in place, and the frame it needs is requested and drawn.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
irisandClaude Opus 5