Shaders: runtime and shader node codegen (#2985)

* shader-rt: initial

* shader-rt: fix recursion when generating shader node

* shader-rt: replace gpu node's args and ret types with `Raster<GPU>`

* shader-rt: properly cfg out the gpu node

* shader-rt: fix `impl Context` in the wrong places

* shader-rt: disable gpu blend node, needs two images

* shader-rt: connect shader runtime

* shader-rt: pass WgpuExecutor by reference

* shader-rt: correct bindings with derpy arg buffer

* shader-rt: manual pipeline layout, fixing errors when bindings got DCE'd

* shader-rt: correct RT format, working invert gpu node

* shader-rt: cleanup codegen with common sym struct

* shader-rt: correct arg buffer handling

* shader-nodes feature: put shader nodes behind feature gate

* shader-nodes feature: rename any `gpu_node` to `shader-node`

* shaders-rt: fix wgpu label name

* shaders-rt: explain fullscreen_vertex coordinates with a drawing

Cargo.lock

@@ -2191,6 +2191,7 @@ dependencies = [
  "specta",
  "spirv-std",
  "tokio",
+ "wgpu-executor",
 ]
 
 [[package]]

node-graph/gcore-shaders/src/blending.rs

@@ -66,7 +66,7 @@ impl AlphaBlending {
 }
 
 #[repr(i32)]
-#[derive(Debug, Default, Clone, Copy, Eq, PartialEq, Hash)]
+#[derive(Debug, Default, Clone, Copy, Eq, PartialEq, Hash, bytemuck::NoUninit)]
 #[cfg_attr(feature = "std", derive(dyn_any::DynAny, specta::Type, serde::Serialize, serde::Deserialize))]
 pub enum BlendMode {
 	// Basic group

node-graph/gcore/src/raster_types.rs

@@ -137,7 +137,7 @@ mod gpu {
 
 	#[derive(Clone, Debug, PartialEq, Hash)]
 	pub struct GPU {
-		texture: wgpu::Texture,
+		pub texture: wgpu::Texture,
 	}
 
 	impl Sealed for Raster<GPU> {}

node-graph/graster-nodes/Cargo.toml

@@ -17,6 +17,7 @@ default = ["std"]
 shader-nodes = [
 	"std",
 	"dep:graphene-raster-nodes-shaders",
+	"dep:wgpu-executor",
 ]
 std = [
 	"dep:graphene-core",
@@ -39,6 +40,7 @@ node-macro = { workspace = true }
 # Local std dependencies
 dyn-any = { workspace = true, optional = true }
 graphene-core = { workspace = true, optional = true }
+wgpu-executor = { workspace = true, optional = true }
 graphene-raster-nodes-shaders = { path = "./shaders", optional = true }
 
 # Workspace dependencies

node-graph/graster-nodes/src/adjustments.rs

@@ -30,7 +30,7 @@ use num_traits::float::Float;
 // https://www.adobe.com/devnet-apps/photoshop/fileformatashtml/#:~:text=%27clrL%27%20%3D%20Color%20Lookup
 // https://www.adobe.com/devnet-apps/photoshop/fileformatashtml/#:~:text=Color%20Lookup%20(Photoshop%20CS6
 
-#[derive(Debug, Default, Clone, Copy, Eq, PartialEq, Hash, node_macro::ChoiceType)]
+#[derive(Debug, Default, Clone, Copy, Eq, PartialEq, Hash, node_macro::ChoiceType, bytemuck::NoUninit)]
 #[cfg_attr(feature = "std", derive(dyn_any::DynAny, specta::Type, serde::Serialize, serde::Deserialize))]
 #[widget(Dropdown)]
 #[repr(u32)]
@@ -560,7 +560,7 @@ pub enum RedGreenBlue {
 }
 
 /// Color Channel
-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Hash, node_macro::ChoiceType)]
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Hash, node_macro::ChoiceType, bytemuck::NoUninit)]
 #[cfg_attr(feature = "std", derive(dyn_any::DynAny, specta::Type, serde::Serialize, serde::Deserialize))]
 #[widget(Radio)]
 #[repr(u32)]

node-graph/graster-nodes/src/blending_nodes.rs

@@ -132,7 +132,7 @@ pub fn apply_blend_mode(foreground: Color, background: Color, blend_mode: BlendM
 	}
 }
 
-#[node_macro::node(category("Raster"), shader_node(PerPixelAdjust))]
+#[node_macro::node(category("Raster"), cfg(feature = "std"))]
 fn blend<T: Blend<Color> + Send>(
 	_: impl Ctx,
 	#[implementations(

node-graph/graster-nodes/src/fullscreen_vertex.rs

@@ -0,0 +1,29 @@
+use glam::{Vec2, Vec4};
+use spirv_std::spirv;
+
+/// webgpu NDC is like OpenGL: (-1.0 .. 1.0, -1.0 .. 1.0, 0.0 .. 1.0)
+/// https://www.w3.org/TR/webgpu/#coordinate-systems
+///
+/// So to make a fullscreen triangle around a box at (-1..1):
+///
+/// ```norun
+/// 3  +
+///    |\
+/// 2  |  \
+///    |    \
+/// 1  +-----+
+///    |     |\
+/// 0  |  0  |  \
+///    |     |    \
+/// -1 +-----+-----+
+///   -1  0  1  2  3
+/// ```
+const FULLSCREEN_VERTICES: [Vec2; 3] = [Vec2::new(-1., -1.), Vec2::new(-1., 3.), Vec2::new(3., -1.)];
+
+#[spirv(vertex)]
+pub fn fullscreen_vertex(#[spirv(vertex_index)] vertex_index: u32, #[spirv(position)] gl_position: &mut Vec4) {
+	// broken on edition 2024 branch
+	// let vertex = unsafe { *FULLSCREEN_VERTICES.index_unchecked(vertex_index as usize) };
+	let vertex = FULLSCREEN_VERTICES[vertex_index as usize];
+	*gl_position = Vec4::from((vertex, 0., 1.));
+}

node-graph/graster-nodes/src/lib.rs

@@ -4,6 +4,11 @@ pub mod adjust;
 pub mod adjustments;
 pub mod blending_nodes;
 pub mod cubic_spline;
+pub mod fullscreen_vertex;
+
+/// required by shader macro
+#[cfg(feature = "shader-nodes")]
+pub use graphene_raster_nodes_shaders::WGSL_SHADER;
 
 #[cfg(feature = "std")]
 pub mod curve;

node-graph/node-macro/src/codegen.rs

@@ -295,7 +295,8 @@ pub(crate) fn generate_node_code(parsed: &ParsedNodeFn) -> syn::Result<TokenStre
 
 	let cfg = crate::shader_nodes::modify_cfg(attributes);
 	let node_input_accessor = generate_node_input_references(parsed, fn_generics, &field_idents, &graphene_core, &identifier, &cfg);
-	let shader_entry_point = attributes.shader_node.as_ref().map(|n| n.codegen_shader_entry_point(parsed)).unwrap_or(Ok(TokenStream2::new()))?;
+	let ShaderTokens { shader_entry_point, gpu_node } = attributes.shader_node.as_ref().map(|n| n.codegen(parsed)).unwrap_or(Ok(ShaderTokens::default()))?;
+
 	Ok(quote! {
 		/// Underlying implementation for [#struct_name]
 		#[inline]
@@ -387,6 +388,8 @@ pub(crate) fn generate_node_code(parsed: &ParsedNodeFn) -> syn::Result<TokenStre
 		}
 
 		#shader_entry_point
+
+		#gpu_node
 	})
 }
 
@@ -589,7 +592,7 @@ fn generate_register_node_impl(parsed: &ParsedNodeFn, field_names: &[&Ident], st
 	})
 }
 
-use crate::shader_nodes::CodegenShaderEntryPoint;
+use crate::shader_nodes::{ShaderCodegen, ShaderTokens};
 use syn::visit_mut::VisitMut;
 use syn::{GenericArgument, Lifetime, Type};
 

node-graph/node-macro/src/parsing.rs

@@ -39,7 +39,7 @@ pub(crate) struct ParsedNodeFn {
 	pub(crate) description: String,
 }
 
-#[derive(Debug, Default)]
+#[derive(Debug, Default, Clone)]
 pub(crate) struct NodeFnAttributes {
 	pub(crate) category: Option<LitStr>,
 	pub(crate) display_name: Option<LitStr>,
@@ -144,7 +144,7 @@ pub struct NodeParsedField {
 	pub implementations: Punctuated<Implementation, Comma>,
 }
 
-#[derive(Debug)]
+#[derive(Clone, Debug)]
 pub(crate) struct Input {
 	pub(crate) pat_ident: PatIdent,
 	pub(crate) ty: Type,
@@ -663,7 +663,7 @@ pub fn new_node_fn(attr: TokenStream2, item: TokenStream2) -> TokenStream2 {
 }
 
 impl ParsedNodeFn {
-	fn replace_impl_trait_in_input(&mut self) {
+	pub fn replace_impl_trait_in_input(&mut self) {
 		if let Type::ImplTrait(impl_trait) = self.input.ty.clone() {
 			let ident = Ident::new("_Input", impl_trait.span());
 			let mut bounds = impl_trait.bounds;

node-graph/node-macro/src/shader_nodes/mod.rs

@@ -3,24 +3,38 @@ use crate::shader_nodes::per_pixel_adjust::PerPixelAdjust;
 use proc_macro2::{Ident, TokenStream};
 use quote::quote;
 use strum::VariantNames;
-use syn::Error;
 use syn::parse::{Parse, ParseStream};
+use syn::punctuated::Punctuated;
+use syn::{Error, Token};
 
 pub mod per_pixel_adjust;
 
 pub const STD_FEATURE_GATE: &str = "std";
+pub const SHADER_NODES_FEATURE_GATE: &str = "shader-nodes";
 
 pub fn modify_cfg(attributes: &NodeFnAttributes) -> TokenStream {
-	match (&attributes.cfg, &attributes.shader_node) {
-		(Some(cfg), Some(_)) => quote!(#[cfg(all(#cfg, feature = #STD_FEATURE_GATE))]),
-		(Some(cfg), None) => quote!(#[cfg(#cfg)]),
-		(None, Some(_)) => quote!(#[cfg(feature = #STD_FEATURE_GATE)]),
-		(None, None) => quote!(),
-	}
+	let feature_gate = match &attributes.shader_node {
+		// shader node cfg is done on the mod
+		Some(ShaderNodeType::ShaderNode) => quote!(),
+		Some(_) => quote!(feature = #STD_FEATURE_GATE),
+		None => quote!(),
+	};
+	let cfgs: Punctuated<_, Token![,]> = match &attributes.cfg {
+		None => [&feature_gate].into_iter().collect(),
+		Some(cfg) => [cfg, &feature_gate].into_iter().collect(),
+	};
+	quote!(#[cfg(all(#cfgs))])
 }
 
-#[derive(Debug, VariantNames)]
+#[derive(Debug, Clone, VariantNames)]
 pub(crate) enum ShaderNodeType {
+	/// Marker for this node being in a gpu node crate, but not having a gpu implementation. This is distinct from not
+	/// declaring `shader_node` at all, as it will wrap the CPU node with a `#[cfg(feature = "std")]` feature gate.
+	None,
+	/// Marker for this node being a generated gpu node implementation, that should not emit anything to prevent
+	/// recursively generating more gpu nodes. But it still counts as a gpu node and will get the
+	/// `#[cfg(feature = "std")]` feature gate around it's impl.
+	ShaderNode,
 	PerPixelAdjust(PerPixelAdjust),
 }
 
@@ -28,24 +42,37 @@ impl Parse for ShaderNodeType {
 	fn parse(input: ParseStream) -> syn::Result<Self> {
 		let ident: Ident = input.parse()?;
 		Ok(match ident.to_string().as_str() {
+			"None" => ShaderNodeType::None,
 			"PerPixelAdjust" => ShaderNodeType::PerPixelAdjust(PerPixelAdjust::parse(input)?),
 			_ => return Err(Error::new_spanned(&ident, format!("attr 'shader_node' must be one of {:?}", Self::VARIANTS))),
 		})
 	}
 }
 
-pub trait CodegenShaderEntryPoint {
-	fn codegen_shader_entry_point(&self, parsed: &ParsedNodeFn) -> syn::Result<TokenStream>;
+pub trait ShaderCodegen {
+	fn codegen(&self, parsed: &ParsedNodeFn) -> syn::Result<ShaderTokens>;
 }
 
-impl CodegenShaderEntryPoint for ShaderNodeType {
-	fn codegen_shader_entry_point(&self, parsed: &ParsedNodeFn) -> syn::Result<TokenStream> {
-		if parsed.is_async {
-			return Err(Error::new_spanned(&parsed.fn_name, "Shader nodes must not be async"));
+impl ShaderCodegen for ShaderNodeType {
+	fn codegen(&self, parsed: &ParsedNodeFn) -> syn::Result<ShaderTokens> {
+		match self {
+			ShaderNodeType::None | ShaderNodeType::ShaderNode => (),
+			_ => {
+				if parsed.is_async {
+					return Err(Error::new_spanned(&parsed.fn_name, "Shader nodes must not be async"));
+				}
+			}
 		}
 
 		match self {
-			ShaderNodeType::PerPixelAdjust(x) => x.codegen_shader_entry_point(parsed),
+			ShaderNodeType::None | ShaderNodeType::ShaderNode => Ok(ShaderTokens::default()),
+			ShaderNodeType::PerPixelAdjust(x) => x.codegen(parsed),
 		}
 	}
 }
+
+#[derive(Clone, Default)]
+pub struct ShaderTokens {
+	pub shader_entry_point: TokenStream,
+	pub gpu_node: TokenStream,
+}

node-graph/node-macro/src/shader_nodes/per_pixel_adjust.rs

@@ -1,11 +1,15 @@
-use crate::parsing::{ParsedFieldType, ParsedNodeFn, RegularParsedField};
-use crate::shader_nodes::CodegenShaderEntryPoint;
+use crate::parsing::{Input, NodeFnAttributes, ParsedField, ParsedFieldType, ParsedNodeFn, RegularParsedField};
+use crate::shader_nodes::{SHADER_NODES_FEATURE_GATE, ShaderCodegen, ShaderNodeType, ShaderTokens};
+use convert_case::{Case, Casing};
+use proc_macro_crate::FoundCrate;
 use proc_macro2::{Ident, TokenStream};
 use quote::{ToTokens, format_ident, quote};
 use std::borrow::Cow;
 use syn::parse::{Parse, ParseStream};
+use syn::punctuated::Punctuated;
+use syn::{PatIdent, Type, parse_quote};
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct PerPixelAdjust {}
 
 impl Parse for PerPixelAdjust {
@@ -14,53 +18,113 @@ impl Parse for PerPixelAdjust {
 	}
 }
 
-impl CodegenShaderEntryPoint for PerPixelAdjust {
-	fn codegen_shader_entry_point(&self, parsed: &ParsedNodeFn) -> syn::Result<TokenStream> {
+impl ShaderCodegen for PerPixelAdjust {
+	fn codegen(&self, parsed: &ParsedNodeFn) -> syn::Result<ShaderTokens> {
 		let fn_name = &parsed.fn_name;
-		let gpu_mod = format_ident!("{}_gpu", parsed.fn_name);
-		let spirv_image_ty = quote!(Image2d);
 
-		// bindings for images start at 1
-		let mut binding_cnt = 0;
-		let params = parsed
-			.fields
-			.iter()
-			.map(|f| {
-				let ident = &f.pat_ident;
-				match &f.ty {
-					ParsedFieldType::Node { .. } => Err(syn::Error::new_spanned(ident, "PerPixelAdjust shader nodes cannot accept other nodes as generics")),
-					ParsedFieldType::Regular(RegularParsedField { gpu_image: false, ty, .. }) => Ok(Param {
-						ident: Cow::Borrowed(&ident.ident),
-						ty: Cow::Owned(ty.to_token_stream()),
-						param_type: ParamType::Uniform,
-					}),
-					ParsedFieldType::Regular(RegularParsedField { gpu_image: true, .. }) => {
-						binding_cnt += 1;
-						Ok(Param {
-							ident: Cow::Owned(format_ident!("image_{}", &ident.ident)),
-							ty: Cow::Borrowed(&spirv_image_ty),
-							param_type: ParamType::Image { binding: binding_cnt },
-						})
+		let mut params;
+		let has_uniform;
+		{
+			// categorize params
+			params = parsed
+				.fields
+				.iter()
+				.map(|f| {
+					let ident = &f.pat_ident;
+					match &f.ty {
+						ParsedFieldType::Node { .. } => Err(syn::Error::new_spanned(ident, "PerPixelAdjust shader nodes cannot accept other nodes as generics")),
+						ParsedFieldType::Regular(RegularParsedField { gpu_image: false, ty, .. }) => Ok(Param {
+							ident: Cow::Borrowed(&ident.ident),
+							ty: ty.to_token_stream(),
+							param_type: ParamType::Uniform,
+						}),
+						ParsedFieldType::Regular(RegularParsedField { gpu_image: true, .. }) => {
+							let param = Param {
+								ident: Cow::Owned(format_ident!("image_{}", &ident.ident)),
+								ty: quote!(Image2d),
+								param_type: ParamType::Image { binding: 0 },
+							};
+							Ok(param)
+						}
 					}
-				}
-			})
-			.collect::<syn::Result<Vec<_>>>()?;
+				})
+				.collect::<syn::Result<Vec<_>>>()?;
 
-		let uniform_members = params
+			has_uniform = params.iter().any(|p| matches!(p.param_type, ParamType::Uniform));
+
+			// assign image bindings
+			// if an arg_buffer exists, bindings for images start at 1 to leave 0 for arg buffer
+			let mut binding_cnt = if has_uniform { 1 } else { 0 };
+			for p in params.iter_mut() {
+				match &mut p.param_type {
+					ParamType::Image { binding } => {
+						*binding = binding_cnt;
+						binding_cnt += 1;
+					}
+					ParamType::Uniform => {}
+				}
+			}
+		}
+
+		let entry_point_mod = format_ident!("{}_gpu_entry_point", fn_name);
+		let entry_point_name_ident = format_ident!("ENTRY_POINT_NAME");
+		let entry_point_name = quote!(#entry_point_mod::#entry_point_name_ident);
+		let uniform_struct_ident = format_ident!("Uniform");
+		let uniform_struct = quote!(#entry_point_mod::#uniform_struct_ident);
+		let shader_node_mod = format_ident!("{}_shader_node", fn_name);
+
+		let codegen = PerPixelAdjustCodegen {
+			parsed,
+			params,
+			has_uniform,
+			entry_point_mod,
+			entry_point_name_ident,
+			entry_point_name,
+			uniform_struct_ident,
+			uniform_struct,
+			shader_node_mod,
+		};
+
+		Ok(ShaderTokens {
+			shader_entry_point: codegen.codegen_shader_entry_point()?,
+			gpu_node: codegen.codegen_gpu_node()?,
+		})
+	}
+}
+
+pub struct PerPixelAdjustCodegen<'a> {
+	parsed: &'a ParsedNodeFn,
+	params: Vec<Param<'a>>,
+	has_uniform: bool,
+	entry_point_mod: Ident,
+	entry_point_name_ident: Ident,
+	entry_point_name: TokenStream,
+	uniform_struct_ident: Ident,
+	uniform_struct: TokenStream,
+	shader_node_mod: Ident,
+}
+
+impl PerPixelAdjustCodegen<'_> {
+	fn codegen_shader_entry_point(&self) -> syn::Result<TokenStream> {
+		let fn_name = &self.parsed.fn_name;
+		let uniform_members = self
+			.params
 			.iter()
 			.filter_map(|Param { ident, ty, param_type }| match param_type {
 				ParamType::Image { .. } => None,
 				ParamType::Uniform => Some(quote! {#ident: #ty}),
 			})
 			.collect::<Vec<_>>();
-		let image_params = params
+		let image_params = self
+			.params
 			.iter()
 			.filter_map(|Param { ident, ty, param_type }| match param_type {
 				ParamType::Image { binding } => Some(quote! {#[spirv(descriptor_set = 0, binding = #binding)] #ident: &#ty}),
 				ParamType::Uniform => None,
 			})
 			.collect::<Vec<_>>();
-		let call_args = params
+		let call_args = self
+			.params
 			.iter()
 			.map(|Param { ident, param_type, .. }| match param_type {
 				ParamType::Image { .. } => quote!(Color::from_vec4(#ident.fetch_with(texel_coord, lod(0)))),
@@ -69,8 +133,11 @@ impl CodegenShaderEntryPoint for PerPixelAdjust {
 			.collect::<Vec<_>>();
 		let context = quote!(());
 
+		let entry_point_mod = &self.entry_point_mod;
+		let entry_point_name = &self.entry_point_name_ident;
+		let uniform_struct_ident = &self.uniform_struct_ident;
 		Ok(quote! {
-			pub mod #gpu_mod {
+			pub mod #entry_point_mod {
 				use super::*;
 				use graphene_core_shaders::color::Color;
 				use spirv_std::spirv;
@@ -78,8 +145,12 @@ impl CodegenShaderEntryPoint for PerPixelAdjust {
 				use spirv_std::image::{Image2d, ImageWithMethods};
 				use spirv_std::image::sample_with::lod;
 
-				pub struct Uniform {
-					#(#uniform_members),*
+				pub const #entry_point_name: &str = core::concat!(core::module_path!(), "::entry_point");
+
+				#[repr(C)]
+				#[derive(Copy, Clone, bytemuck::NoUninit)]
+				pub struct #uniform_struct_ident {
+					#(pub #uniform_members),*
 				}
 
 				#[spirv(fragment)]
@@ -96,11 +167,162 @@ impl CodegenShaderEntryPoint for PerPixelAdjust {
 			}
 		})
 	}
+
+	fn codegen_gpu_node(&self) -> syn::Result<TokenStream> {
+		let gcore = match &self.parsed.crate_name {
+			FoundCrate::Itself => format_ident!("crate"),
+			FoundCrate::Name(name) => format_ident!("{name}"),
+		};
+
+		// adapt fields for gpu node
+		let raster_gpu: Type = parse_quote!(#gcore::table::Table<#gcore::raster_types::Raster<#gcore::raster_types::GPU>>);
+		let mut fields = self
+			.parsed
+			.fields
+			.iter()
+			.map(|f| match &f.ty {
+				ParsedFieldType::Regular(reg @ RegularParsedField { gpu_image: true, .. }) => Ok(ParsedField {
+					pat_ident: PatIdent {
+						mutability: None,
+						by_ref: None,
+						..f.pat_ident.clone()
+					},
+					ty: ParsedFieldType::Regular(RegularParsedField {
+						ty: raster_gpu.clone(),
+						implementations: Punctuated::default(),
+						..reg.clone()
+					}),
+					..f.clone()
+				}),
+				ParsedFieldType::Regular(RegularParsedField { gpu_image: false, .. }) => Ok(ParsedField {
+					pat_ident: PatIdent {
+						mutability: None,
+						by_ref: None,
+						..f.pat_ident.clone()
+					},
+					..f.clone()
+				}),
+				ParsedFieldType::Node { .. } => Err(syn::Error::new_spanned(&f.pat_ident, "PerPixelAdjust shader nodes cannot accept other nodes as generics")),
+			})
+			.collect::<syn::Result<Vec<_>>>()?;
+
+		// insert wgpu_executor field
+		let wgpu_executor = format_ident!("__wgpu_executor");
+		fields.push(ParsedField {
+			pat_ident: PatIdent {
+				attrs: vec![],
+				by_ref: None,
+				mutability: None,
+				ident: parse_quote!(#wgpu_executor),
+				subpat: None,
+			},
+			name: None,
+			description: "".to_string(),
+			widget_override: Default::default(),
+			ty: ParsedFieldType::Regular(RegularParsedField {
+				ty: parse_quote!(&'a WgpuExecutor),
+				exposed: false,
+				value_source: Default::default(),
+				number_soft_min: None,
+				number_soft_max: None,
+				number_hard_min: None,
+				number_hard_max: None,
+				number_mode_range: None,
+				implementations: Default::default(),
+				gpu_image: false,
+			}),
+			number_display_decimal_places: None,
+			number_step: None,
+			unit: None,
+		});
+
+		// find exactly one gpu_image field, runtime doesn't support more than 1 atm
+		let gpu_image_field = {
+			let mut iter = fields.iter().filter(|f| matches!(f.ty, ParsedFieldType::Regular(RegularParsedField { gpu_image: true, .. })));
+			match (iter.next(), iter.next()) {
+				(Some(v), None) => Ok(v),
+				(Some(_), Some(more)) => Err(syn::Error::new_spanned(&more.pat_ident, "No more than one parameter must be annotated with `#[gpu_image]`")),
+				(None, _) => Err(syn::Error::new_spanned(&self.parsed.fn_name, "At least one parameter must be annotated with `#[gpu_image]`")),
+			}?
+		};
+		let gpu_image = &gpu_image_field.pat_ident.ident;
+
+		// uniform buffer struct construction
+		let has_uniform = self.has_uniform;
+		let uniform_buffer = if has_uniform {
+			let uniform_struct = &self.uniform_struct;
+			let uniform_members = self
+				.params
+				.iter()
+				.filter_map(|p| match p.param_type {
+					ParamType::Image { .. } => None,
+					ParamType::Uniform => Some(p.ident.as_ref()),
+				})
+				.collect::<Vec<_>>();
+			quote!(Some(&super::#uniform_struct {
+				#(#uniform_members),*
+			}))
+		} else {
+			// explicit generics placed here cause it's easier than explicitly writing `run_per_pixel_adjust::<()>`
+			quote!(Option::<&()>::None)
+		};
+
+		// node function body
+		let entry_point_name = &self.entry_point_name;
+		let body = quote! {
+			{
+				#wgpu_executor.shader_runtime.run_per_pixel_adjust(&::wgpu_executor::shader_runtime::per_pixel_adjust_runtime::Shaders {
+					wgsl_shader: crate::WGSL_SHADER,
+					fragment_shader_name: super::#entry_point_name,
+					has_uniform: #has_uniform,
+				}, #gpu_image, #uniform_buffer).await
+			}
+		};
+
+		// call node codegen
+		let mut parsed_node_fn = ParsedNodeFn {
+			vis: self.parsed.vis.clone(),
+			attributes: NodeFnAttributes {
+				shader_node: Some(ShaderNodeType::ShaderNode),
+				..self.parsed.attributes.clone()
+			},
+			fn_name: self.shader_node_mod.clone(),
+			struct_name: format_ident!("{}", self.shader_node_mod.to_string().to_case(Case::Pascal)),
+			mod_name: self.shader_node_mod.clone(),
+			fn_generics: vec![parse_quote!('a: 'n)],
+			where_clause: None,
+			input: Input {
+				pat_ident: self.parsed.input.pat_ident.clone(),
+				ty: parse_quote!(impl #gcore::context::Ctx),
+				implementations: Default::default(),
+			},
+			output_type: raster_gpu,
+			is_async: true,
+			fields,
+			body,
+			crate_name: self.parsed.crate_name.clone(),
+			description: "".to_string(),
+		};
+		parsed_node_fn.replace_impl_trait_in_input();
+		let gpu_node_impl = crate::codegen::generate_node_code(&parsed_node_fn)?;
+
+		// wrap node in `mod #gpu_node_mod`
+		let shader_node_mod = &self.shader_node_mod;
+		Ok(quote! {
+			#[cfg(feature = #SHADER_NODES_FEATURE_GATE)]
+			mod #shader_node_mod {
+				use super::*;
+				use wgpu_executor::WgpuExecutor;
+
+				#gpu_node_impl
+			}
+		})
+	}
 }
 
 struct Param<'a> {
 	ident: Cow<'a, Ident>,
-	ty: Cow<'a, TokenStream>,
+	ty: TokenStream,
 	param_type: ParamType,
 }
 

node-graph/wgpu-executor/src/lib.rs

@@ -1,6 +1,8 @@
 mod context;
+pub mod shader_runtime;
 pub mod texture_upload;
 
+use crate::shader_runtime::ShaderRuntime;
 use anyhow::Result;
 pub use context::Context;
 use dyn_any::StaticType;
@@ -18,6 +20,7 @@ use wgpu::{Origin3d, SurfaceConfiguration, TextureAspect};
 pub struct WgpuExecutor {
 	pub context: Context,
 	vello_renderer: Mutex<Renderer>,
+	pub shader_runtime: ShaderRuntime,
 }
 
 impl std::fmt::Debug for WgpuExecutor {
@@ -195,6 +198,7 @@ impl WgpuExecutor {
 		.ok()?;
 
 		Some(Self {
+			shader_runtime: ShaderRuntime::new(&context),
 			context,
 			vello_renderer: vello_renderer.into(),
 		})

node-graph/wgpu-executor/src/shader_runtime/mod.rs

@@ -0,0 +1,20 @@
+use crate::Context;
+use crate::shader_runtime::per_pixel_adjust_runtime::PerPixelAdjustShaderRuntime;
+
+pub mod per_pixel_adjust_runtime;
+
+pub const FULLSCREEN_VERTEX_SHADER_NAME: &str = "fullscreen_vertexfullscreen_vertex";
+
+pub struct ShaderRuntime {
+	context: Context,
+	per_pixel_adjust: PerPixelAdjustShaderRuntime,
+}
+
+impl ShaderRuntime {
+	pub fn new(context: &Context) -> Self {
+		Self {
+			context: context.clone(),
+			per_pixel_adjust: PerPixelAdjustShaderRuntime::new(),
+		}
+	}
+}

node-graph/wgpu-executor/src/shader_runtime/per_pixel_adjust_runtime.rs

@@ -0,0 +1,243 @@
+use crate::Context;
+use crate::shader_runtime::{FULLSCREEN_VERTEX_SHADER_NAME, ShaderRuntime};
+use bytemuck::NoUninit;
+use futures::lock::Mutex;
+use graphene_core::raster_types::{GPU, Raster};
+use graphene_core::table::{Table, TableRow};
+use std::borrow::Cow;
+use std::collections::HashMap;
+use wgpu::util::{BufferInitDescriptor, DeviceExt};
+use wgpu::{
+	BindGroupDescriptor, BindGroupEntry, BindGroupLayoutDescriptor, BindGroupLayoutEntry, BindingResource, BindingType, Buffer, BufferBinding, BufferBindingType, BufferUsages, ColorTargetState, Face,
+	FragmentState, FrontFace, LoadOp, Operations, PipelineLayoutDescriptor, PolygonMode, PrimitiveState, PrimitiveTopology, RenderPassColorAttachment, RenderPassDescriptor, RenderPipelineDescriptor,
+	ShaderModuleDescriptor, ShaderSource, ShaderStages, StoreOp, TextureDescriptor, TextureDimension, TextureFormat, TextureSampleType, TextureViewDescriptor, TextureViewDimension, VertexState,
+};
+
+pub struct PerPixelAdjustShaderRuntime {
+	// TODO: PerPixelAdjustGraphicsPipeline already contains the key as `name`
+	pipeline_cache: Mutex<HashMap<String, PerPixelAdjustGraphicsPipeline>>,
+}
+
+impl PerPixelAdjustShaderRuntime {
+	pub fn new() -> Self {
+		Self {
+			pipeline_cache: Mutex::new(HashMap::new()),
+		}
+	}
+}
+
+impl ShaderRuntime {
+	pub async fn run_per_pixel_adjust<T: NoUninit>(&self, shaders: &Shaders<'_>, textures: Table<Raster<GPU>>, args: Option<&T>) -> Table<Raster<GPU>> {
+		let mut cache = self.per_pixel_adjust.pipeline_cache.lock().await;
+		let pipeline = cache
+			.entry(shaders.fragment_shader_name.to_owned())
+			.or_insert_with(|| PerPixelAdjustGraphicsPipeline::new(&self.context, &shaders));
+
+		let arg_buffer = args.map(|args| {
+			let device = &self.context.device;
+			device.create_buffer_init(&BufferInitDescriptor {
+				label: Some(&format!("{} arg buffer", pipeline.name.as_str())),
+				usage: BufferUsages::STORAGE,
+				contents: bytemuck::bytes_of(args),
+			})
+		});
+		pipeline.dispatch(&self.context, textures, arg_buffer)
+	}
+}
+
+pub struct Shaders<'a> {
+	pub wgsl_shader: &'a str,
+	pub fragment_shader_name: &'a str,
+	pub has_uniform: bool,
+}
+
+pub struct PerPixelAdjustGraphicsPipeline {
+	name: String,
+	has_uniform: bool,
+	pipeline: wgpu::RenderPipeline,
+}
+
+impl PerPixelAdjustGraphicsPipeline {
+	pub fn new(context: &Context, info: &Shaders) -> Self {
+		let device = &context.device;
+		let name = info.fragment_shader_name.to_owned();
+
+		let fragment_name = &name;
+		let fragment_name = &fragment_name[(fragment_name.find("::").unwrap() + 2)..];
+		// TODO workaround to naga removing `:`
+		let fragment_name = fragment_name.replace(":", "");
+		let shader_module = device.create_shader_module(ShaderModuleDescriptor {
+			label: Some(&format!("PerPixelAdjust {} wgsl shader", name)),
+			source: ShaderSource::Wgsl(Cow::Borrowed(info.wgsl_shader)),
+		});
+
+		let entries: &[_] = if info.has_uniform {
+			&[
+				BindGroupLayoutEntry {
+					binding: 0,
+					visibility: ShaderStages::FRAGMENT,
+					ty: BindingType::Buffer {
+						ty: BufferBindingType::Storage { read_only: true },
+						has_dynamic_offset: false,
+						min_binding_size: None,
+					},
+					count: None,
+				},
+				BindGroupLayoutEntry {
+					binding: 1,
+					visibility: ShaderStages::FRAGMENT,
+					ty: BindingType::Texture {
+						sample_type: TextureSampleType::Float { filterable: false },
+						view_dimension: TextureViewDimension::D2,
+						multisampled: false,
+					},
+					count: None,
+				},
+			]
+		} else {
+			&[BindGroupLayoutEntry {
+				binding: 0,
+				visibility: ShaderStages::FRAGMENT,
+				ty: BindingType::Texture {
+					sample_type: TextureSampleType::Float { filterable: false },
+					view_dimension: TextureViewDimension::D2,
+					multisampled: false,
+				},
+				count: None,
+			}]
+		};
+		let pipeline_layout = device.create_pipeline_layout(&PipelineLayoutDescriptor {
+			label: Some(&format!("PerPixelAdjust {} PipelineLayout", name)),
+			bind_group_layouts: &[&device.create_bind_group_layout(&BindGroupLayoutDescriptor {
+				label: Some(&format!("PerPixelAdjust {} BindGroupLayout 0", name)),
+				entries,
+			})],
+			push_constant_ranges: &[],
+		});
+
+		let pipeline = device.create_render_pipeline(&RenderPipelineDescriptor {
+			label: Some(&format!("PerPixelAdjust {} Pipeline", name)),
+			layout: Some(&pipeline_layout),
+			vertex: VertexState {
+				module: &shader_module,
+				entry_point: Some(FULLSCREEN_VERTEX_SHADER_NAME),
+				compilation_options: Default::default(),
+				buffers: &[],
+			},
+			primitive: PrimitiveState {
+				topology: PrimitiveTopology::TriangleList,
+				strip_index_format: None,
+				front_face: FrontFace::Ccw,
+				cull_mode: Some(Face::Back),
+				unclipped_depth: false,
+				polygon_mode: PolygonMode::Fill,
+				conservative: false,
+			},
+			depth_stencil: None,
+			multisample: Default::default(),
+			fragment: Some(FragmentState {
+				module: &shader_module,
+				entry_point: Some(&fragment_name),
+				compilation_options: Default::default(),
+				targets: &[Some(ColorTargetState {
+					format: TextureFormat::Rgba8UnormSrgb,
+					blend: None,
+					write_mask: Default::default(),
+				})],
+			}),
+			multiview: None,
+			cache: None,
+		});
+		Self {
+			pipeline,
+			name,
+			has_uniform: info.has_uniform,
+		}
+	}
+
+	pub fn dispatch(&self, context: &Context, textures: Table<Raster<GPU>>, arg_buffer: Option<Buffer>) -> Table<Raster<GPU>> {
+		assert_eq!(self.has_uniform, arg_buffer.is_some());
+		let device = &context.device;
+		let name = self.name.as_str();
+
+		let mut cmd = device.create_command_encoder(&wgpu::CommandEncoderDescriptor {
+			label: Some(&format!("{name} cmd encoder")),
+		});
+		let out = textures
+			.iter()
+			.map(|instance| {
+				let tex_in = &instance.element.texture;
+				let view_in = tex_in.create_view(&TextureViewDescriptor::default());
+				let format = tex_in.format();
+
+				let entries: &[_] = if let Some(arg_buffer) = arg_buffer.as_ref() {
+					&[
+						BindGroupEntry {
+							binding: 0,
+							resource: BindingResource::Buffer(BufferBinding {
+								buffer: arg_buffer,
+								offset: 0,
+								size: None,
+							}),
+						},
+						BindGroupEntry {
+							binding: 1,
+							resource: BindingResource::TextureView(&view_in),
+						},
+					]
+				} else {
+					&[BindGroupEntry {
+						binding: 0,
+						resource: BindingResource::TextureView(&view_in),
+					}]
+				};
+				let bind_group = device.create_bind_group(&BindGroupDescriptor {
+					label: Some(&format!("{name} bind group")),
+					// `get_bind_group_layout` allocates unnecessary memory, we could create it manually to not do that
+					layout: &self.pipeline.get_bind_group_layout(0),
+					entries,
+				});
+
+				let tex_out = device.create_texture(&TextureDescriptor {
+					label: Some(&format!("{name} texture out")),
+					size: tex_in.size(),
+					mip_level_count: 1,
+					sample_count: 1,
+					dimension: TextureDimension::D2,
+					format,
+					usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST | wgpu::TextureUsages::COPY_SRC | wgpu::TextureUsages::RENDER_ATTACHMENT,
+					view_formats: &[format],
+				});
+
+				let view_out = tex_out.create_view(&TextureViewDescriptor::default());
+				let mut rp = cmd.begin_render_pass(&RenderPassDescriptor {
+					label: Some(&format!("{name} render pipeline")),
+					color_attachments: &[Some(RenderPassColorAttachment {
+						view: &view_out,
+						resolve_target: None,
+						ops: Operations {
+							// should be dont_care but wgpu doesn't expose that
+							load: LoadOp::Clear(wgpu::Color::BLACK),
+							store: StoreOp::Store,
+						},
+					})],
+					depth_stencil_attachment: None,
+					timestamp_writes: None,
+					occlusion_query_set: None,
+				});
+				rp.set_pipeline(&self.pipeline);
+				rp.set_bind_group(0, Some(&bind_group), &[]);
+				rp.draw(0..3, 0..1);
+
+				TableRow {
+					element: Raster::new(GPU { texture: tex_out }),
+					transform: *instance.transform,
+					alpha_blending: *instance.alpha_blending,
+					source_node_id: *instance.source_node_id,
+				}
+			})
+			.collect::<Table<_>>();
+		context.queue.submit([cmd.finish()]);
+		out
+	}
+}