Use vello_shaders crate to load and preprocess WGSL

.github/copyright.sh

@@ -7,7 +7,7 @@
 #   -g "!src/special_directory"
 
 # Check all the standard Rust source files
-output=$(rg "^// Copyright (19|20)[\d]{2} (.+ and )?the Vello Authors( and .+)?$\n^// SPDX-License-Identifier: Apache-2\.0 OR MIT$\n\n" --files-without-match --multiline -g "*.rs" -g "!{shader,src/cpu_shader}" .)
+output=$(rg "^// Copyright (19|20)[\d]{2} (.+ and )?the Vello Authors( and .+)?$\n^// SPDX-License-Identifier: Apache-2\.0 OR MIT$\n\n" --files-without-match --multiline -g "*.rs" -g "!{shader,crates/shaders/src/cpu}" .)
 
 if [ -n "$output" ]; then
 	echo -e "The following files lack the correct copyright header:\n"
@@ -20,7 +20,7 @@ if [ -n "$output" ]; then
 fi
 
 # Check all the shaders, both WGSL and CPU shaders in Rust, as they also have Unlicense
-output=$(rg "^// Copyright (19|20)[\d]{2} (.+ and )?the Vello Authors( and .+)?$\n^// SPDX-License-Identifier: Apache-2\.0 OR MIT OR Unlicense$\n\n" --files-without-match --multiline -g "{shader,src/cpu_shader}/**/*.{rs,wgsl}" .)
+output=$(rg "^// Copyright (19|20)[\d]{2} (.+ and )?the Vello Authors( and .+)?$\n^// SPDX-License-Identifier: Apache-2\.0 OR MIT OR Unlicense$\n\n" --files-without-match --multiline -g "{shader,crates/shaders/src/cpu}/**/*.{rs,wgsl}" .)
 
 if [ -n "$output" ]; then
         echo -e "The following shader files lack the correct copyright header:\n"

Cargo.toml

@@ -47,14 +47,15 @@ default = ["wgpu"]
 # bump-allocated GPU memory.
 # TODO: Turn this into a runtime option used at resolve time and remove the feature.
 bump_estimate = ["vello_encoding/bump_estimate"]
-hot_reload = []
+hot_reload = ["vello_shaders/compile"]
 buffer_labels = []
 
 [lints]
 workspace = true
 
 [dependencies]
 vello_encoding = { workspace = true }
+vello_shaders = { workspace = true }
 bytemuck = { workspace = true }
 skrifa = { workspace = true }
 peniko = { workspace = true }
@@ -71,6 +72,7 @@ clippy.semicolon_if_nothing_returned = "warn"
 
 [workspace.dependencies]
 vello_encoding = { version = "0.1.0", path = "crates/encoding" }
+vello_shaders = { version = "0.1.0", path = "crates/shaders" }
 bytemuck = { version = "1.15.0", features = ["derive"] }
 skrifa = "0.19.0"
 peniko = "0.1.0"

crates/shaders/Cargo.toml

@@ -10,7 +10,7 @@ repository.workspace = true
 publish = false # Remove this when the package is ready for publishing
 
 [features]
-default = ["compile", "wgsl", "msl", "full"]
+default = ["wgsl", "full", "cpu"]
 compile = ["naga", "thiserror"]
 
 # Enables the complete imaging model. When this feature is disabled, the fine rasterization
@@ -22,12 +22,17 @@ full = []
 wgsl = []
 msl = []
 
+# Enable the CPU versions of the shaders
+cpu = ["bytemuck", "vello_encoding"]
+
 [lints]
 workspace = true
 
 [dependencies]
+bytemuck = { workspace = true, optional = true }
 naga = { version = "0.19.2", features = ["wgsl-in", "msl-out",], optional = true }
 thiserror = { version = "1.0.58", optional = true }
+vello_encoding = { workspace = true, optional = true }
 
 [build-dependencies]
 naga = { version = "0.19.2",  features = ["wgsl-in", "msl-out",] }

crates/shaders/src/compile/mod.rs

@@ -139,7 +139,7 @@ impl ShaderInfo {
         } else {
             Default::default()
         };
-        println!("{:?}", permutation_map);
+        //println!("{permutation_map:?}");
         let imports = preprocess::get_imports(shader_dir);
         let mut info = HashMap::default();
         let defines: HashSet<_> = if cfg!(feature = "full") {

src/cpu_dispatch.rs → crates/shaders/src/cpu.rs

@@ -4,7 +4,48 @@
 //! Support for CPU implementations of compute shaders.
 //!
 //! Note that while this CPU implementation is useful for testing and debugging,
-//! a full CPU fallback for targets without wgpu hasn't been implemented yet.
+//! a full CPU fallback as an alternative to GPU shaders is not provided.
+
+// Allow un-idiomatic Rust to more closely match shaders
+#![allow(clippy::needless_range_loop)]
+#![allow(clippy::too_many_arguments)]
+
+mod backdrop;
+mod bbox_clear;
+mod binning;
+mod clip_leaf;
+mod clip_reduce;
+mod coarse;
+mod draw_leaf;
+mod draw_reduce;
+mod euler;
+mod fine;
+mod flatten;
+mod path_count;
+mod path_count_setup;
+mod path_tiling;
+mod path_tiling_setup;
+mod pathtag_reduce;
+mod pathtag_scan;
+mod tile_alloc;
+mod util;
+
+pub use backdrop::backdrop;
+pub use bbox_clear::bbox_clear;
+pub use binning::binning;
+pub use clip_leaf::clip_leaf;
+pub use clip_reduce::clip_reduce;
+pub use coarse::coarse;
+pub use draw_leaf::draw_leaf;
+pub use draw_reduce::draw_reduce;
+pub use flatten::flatten;
+pub use path_count::path_count;
+pub use path_count_setup::path_count_setup;
+pub use path_tiling::path_tiling;
+pub use path_tiling_setup::path_tiling_setup;
+pub use pathtag_reduce::pathtag_reduce;
+pub use pathtag_scan::pathtag_scan;
+pub use tile_alloc::tile_alloc;
 
 use std::cell::{Ref, RefCell, RefMut};
 use std::ops::{Deref, DerefMut};
@@ -123,3 +164,34 @@ pub struct CpuTexture {
     // In RGBA format. May expand in the future.
     pub pixels: Vec<u32>,
 }
+
+// Common internal definitions
+
+const PTCL_INITIAL_ALLOC: u32 = 64;
+
+// Tags for PTCL commands
+const CMD_END: u32 = 0;
+const CMD_FILL: u32 = 1;
+//const CMD_STROKE: u32 = 2;
+const CMD_SOLID: u32 = 3;
+const CMD_COLOR: u32 = 5;
+const CMD_LIN_GRAD: u32 = 6;
+const CMD_RAD_GRAD: u32 = 7;
+const CMD_SWEEP_GRAD: u32 = 8;
+const CMD_IMAGE: u32 = 9;
+const CMD_BEGIN_CLIP: u32 = 10;
+const CMD_END_CLIP: u32 = 11;
+const CMD_JUMP: u32 = 12;
+
+// The following are computed in draw_leaf from the generic gradient parameters
+// encoded in the scene, and stored in the gradient's info struct, for
+// consumption during fine rasterization.
+
+// Radial gradient kinds
+const RAD_GRAD_KIND_CIRCULAR: u32 = 1;
+const RAD_GRAD_KIND_STRIP: u32 = 2;
+const RAD_GRAD_KIND_FOCAL_ON_CIRCLE: u32 = 3;
+const RAD_GRAD_KIND_CONE: u32 = 4;
+
+// Radial gradient flags
+const RAD_GRAD_SWAPPED: u32 = 1;

src/cpu_shader/backdrop.rs → crates/shaders/src/cpu/backdrop.rs

@@ -3,7 +3,7 @@
 
 use vello_encoding::{ConfigUniform, Path, Tile};
 
-use crate::cpu_dispatch::CpuBinding;
+use super::CpuBinding;
 
 fn backdrop_main(config: &ConfigUniform, paths: &[Path], tiles: &mut [Tile]) {
     for drawobj_ix in 0..config.layout.n_draw_objects {

src/cpu_shader/bbox_clear.rs → crates/shaders/src/cpu/bbox_clear.rs

@@ -3,7 +3,7 @@
 
 use vello_encoding::{ConfigUniform, PathBbox};
 
-use crate::cpu_dispatch::CpuBinding;
+use super::CpuBinding;
 
 fn bbox_clear_main(config: &ConfigUniform, path_bboxes: &mut [PathBbox]) {
     for i in 0..(config.layout.n_paths as usize) {

src/cpu_shader/binning.rs → crates/shaders/src/cpu/binning.rs

@@ -3,7 +3,7 @@
 
 use vello_encoding::{BinHeader, BumpAllocators, ConfigUniform, DrawMonoid, PathBbox};
 
-use crate::cpu_dispatch::CpuBinding;
+use super::CpuBinding;
 
 const WG_SIZE: usize = 256;
 const TILE_WIDTH: usize = 16;

src/cpu_shader/clip_leaf.rs → crates/shaders/src/cpu/clip_leaf.rs

@@ -3,7 +3,7 @@
 
 use vello_encoding::{Clip, ConfigUniform, DrawMonoid, PathBbox};
 
-use crate::cpu_dispatch::CpuBinding;
+use super::CpuBinding;
 
 struct ClipStackElement {
     // index of draw object

src/cpu_shader/clip_reduce.rs → crates/shaders/src/cpu/clip_reduce.rs

@@ -3,7 +3,7 @@
 
 use vello_encoding::{Clip, ClipBic, ClipElement, PathBbox};
 
-use crate::cpu_dispatch::CpuBinding;
+use super::CpuBinding;
 
 const WG_SIZE: usize = 256;
 

src/cpu_shader/coarse.rs → crates/shaders/src/cpu/coarse.rs

@@ -6,11 +6,9 @@ use vello_encoding::{
     DRAW_INFO_FLAGS_FILL_RULE_BIT,
 };
 
-use crate::cpu_dispatch::CpuBinding;
-
 use super::{
-    CMD_BEGIN_CLIP, CMD_COLOR, CMD_END, CMD_END_CLIP, CMD_FILL, CMD_IMAGE, CMD_JUMP, CMD_LIN_GRAD,
-    CMD_RAD_GRAD, CMD_SOLID, CMD_SWEEP_GRAD, PTCL_INITIAL_ALLOC,
+    CpuBinding, CMD_BEGIN_CLIP, CMD_COLOR, CMD_END, CMD_END_CLIP, CMD_FILL, CMD_IMAGE, CMD_JUMP,
+    CMD_LIN_GRAD, CMD_RAD_GRAD, CMD_SOLID, CMD_SWEEP_GRAD, PTCL_INITIAL_ALLOC,
 };
 
 const N_TILE_X: usize = 16;

src/cpu_shader/draw_leaf.rs → crates/shaders/src/cpu/draw_leaf.rs

@@ -3,12 +3,10 @@
 
 use vello_encoding::{Clip, ConfigUniform, DrawMonoid, DrawTag, Monoid, PathBbox};
 
-use crate::cpu_dispatch::CpuBinding;
-
-use super::util::{read_draw_tag_from_scene, Transform, Vec2};
 use super::{
-    RAD_GRAD_KIND_CIRCULAR, RAD_GRAD_KIND_CONE, RAD_GRAD_KIND_FOCAL_ON_CIRCLE, RAD_GRAD_KIND_STRIP,
-    RAD_GRAD_SWAPPED,
+    util::{read_draw_tag_from_scene, Transform, Vec2},
+    CpuBinding, RAD_GRAD_KIND_CIRCULAR, RAD_GRAD_KIND_CONE, RAD_GRAD_KIND_FOCAL_ON_CIRCLE,
+    RAD_GRAD_KIND_STRIP, RAD_GRAD_SWAPPED,
 };
 
 const WG_SIZE: usize = 256;

src/cpu_shader/draw_reduce.rs → crates/shaders/src/cpu/draw_reduce.rs

@@ -3,9 +3,7 @@
 
 use vello_encoding::{ConfigUniform, DrawMonoid, DrawTag, Monoid};
 
-use crate::cpu_dispatch::CpuBinding;
-
-use super::util::read_draw_tag_from_scene;
+use super::{util::read_draw_tag_from_scene, CpuBinding};
 
 const WG_SIZE: usize = 256;
 

src/cpu_shader/fine.rs → crates/shaders/src/cpu/fine.rs

@@ -3,9 +3,7 @@
 
 use vello_encoding::{ConfigUniform, PathSegment, Tile};
 
-use crate::cpu_dispatch::CpuTexture;
-
-use super::{CMD_COLOR, CMD_END, CMD_FILL, CMD_JUMP, CMD_SOLID, PTCL_INITIAL_ALLOC};
+use super::{CpuTexture, CMD_COLOR, CMD_END, CMD_FILL, CMD_JUMP, CMD_SOLID, PTCL_INITIAL_ALLOC};
 
 // These should also move into a common area
 const TILE_WIDTH: usize = 16;

src/cpu_shader/flatten.rs → crates/shaders/src/cpu/flatten.rs

@@ -3,11 +3,13 @@
 
 use std::f32::consts::FRAC_1_SQRT_2;
 
-use super::euler::{
-    espc_int_approx, espc_int_inv_approx, CubicParams, EulerParams, EulerSeg, TANGENT_THRESH,
+use super::{
+    euler::{
+        espc_int_approx, espc_int_inv_approx, CubicParams, EulerParams, EulerSeg, TANGENT_THRESH,
+    },
+    util::{Transform, Vec2, ROBUST_EPSILON},
+    CpuBinding,
 };
-use super::util::{Transform, Vec2, ROBUST_EPSILON};
-use crate::cpu_dispatch::CpuBinding;
 use vello_encoding::math::f16_to_f32;
 use vello_encoding::{
     BumpAllocators, ConfigUniform, LineSoup, Monoid, PathBbox, PathMonoid, PathTag, Style,

src/cpu_shader/path_count.rs → crates/shaders/src/cpu/path_count.rs

@@ -3,9 +3,10 @@
 
 use vello_encoding::{BumpAllocators, LineSoup, Path, SegmentCount, Tile};
 
-use crate::cpu_dispatch::CpuBinding;
-
-use super::util::{span, Vec2, ONE_MINUS_ULP, ROBUST_EPSILON};
+use super::{
+    util::{span, Vec2, ONE_MINUS_ULP, ROBUST_EPSILON},
+    CpuBinding,
+};
 
 const TILE_SCALE: f32 = 1.0 / 16.0;
 

src/cpu_shader/path_count_setup.rs → crates/shaders/src/cpu/path_count_setup.rs

@@ -3,7 +3,7 @@
 
 use vello_encoding::{BumpAllocators, IndirectCount};
 
-use crate::cpu_dispatch::CpuBinding;
+use super::CpuBinding;
 
 const WG_SIZE: usize = 256;
 

src/cpu_shader/path_tiling.rs → crates/shaders/src/cpu/path_tiling.rs

@@ -3,10 +3,10 @@
 
 use vello_encoding::{BumpAllocators, LineSoup, Path, PathSegment, SegmentCount, Tile};
 
-use crate::cpu_dispatch::CpuBinding;
-use crate::cpu_shader::util::{ONE_MINUS_ULP, ROBUST_EPSILON};
-
-use super::util::{span, Vec2};
+use super::{
+    util::{span, Vec2, ONE_MINUS_ULP, ROBUST_EPSILON},
+    CpuBinding,
+};
 
 const TILE_WIDTH: u32 = 16;
 const TILE_HEIGHT: u32 = 16;

src/cpu_shader/path_tiling_setup.rs → crates/shaders/src/cpu/path_tiling_setup.rs

@@ -3,7 +3,7 @@
 
 use vello_encoding::{BumpAllocators, IndirectCount};
 
-use crate::cpu_dispatch::CpuBinding;
+use super::CpuBinding;
 
 const WG_SIZE: usize = 256;
 

src/cpu_shader/pathtag_reduce.rs → crates/shaders/src/cpu/pathtag_reduce.rs

@@ -3,7 +3,7 @@
 
 use vello_encoding::{ConfigUniform, Monoid, PathMonoid};
 
-use crate::cpu_dispatch::CpuBinding;
+use super::CpuBinding;
 
 const WG_SIZE: usize = 256;
 

src/cpu_shader/pathtag_scan.rs → crates/shaders/src/cpu/pathtag_scan.rs

@@ -3,7 +3,7 @@
 
 use vello_encoding::{ConfigUniform, Monoid, PathMonoid};
 
-use crate::cpu_dispatch::CpuBinding;
+use super::CpuBinding;
 
 const WG_SIZE: usize = 256;
 

src/cpu_shader/tile_alloc.rs → crates/shaders/src/cpu/tile_alloc.rs

@@ -3,7 +3,7 @@
 
 use vello_encoding::{BumpAllocators, ConfigUniform, DrawTag, Path, Tile};
 
-use crate::cpu_dispatch::CpuBinding;
+use super::CpuBinding;
 
 const TILE_WIDTH: usize = 16;
 const TILE_HEIGHT: usize = 16;

crates/shaders/src/lib.rs

@@ -5,11 +5,12 @@ mod types;
 
 #[cfg(feature = "compile")]
 pub mod compile;
-
-pub use types::{BindType, BindingInfo, WorkgroupBufferInfo};
+#[cfg(feature = "cpu")]
+pub mod cpu;
 
 #[cfg(feature = "msl")]
 pub use types::msl;
+pub use types::{BindType, BindingInfo, WorkgroupBufferInfo};
 
 use std::borrow::Cow;