Add the first basic version of the GPU blend node (#1243)

* Implement Gpu Blend node

* Remove duplicate shader input

* Fix formatting

---------

Co-authored-by: Dennis Kobert <dennis@kobert.dev>

Cargo.lock

@@ -651,6 +651,7 @@ dependencies = [
  "gpu-compiler-bin-wrapper",
  "gpu-executor",
  "graph-craft",
+ "graphene-core",
  "reqwest",
  "serde_json",
  "tempfile",

editor/src/messages/portfolio/document/node_graph/node_graph_message_handler/document_node_types.rs

@@ -723,6 +723,20 @@ fn static_nodes() -> Vec<DocumentNodeType> {
 			outputs: vec![DocumentOutputType::new("Image", FrontendGraphDataType::Raster)],
 			properties: node_properties::no_properties,
 		},
+		#[cfg(feature = "gpu")]
+		DocumentNodeType {
+			name: "Blend (GPU)",
+			category: "Image Adjustments",
+			identifier: NodeImplementation::proto("graphene_std::executor::BlendGpuImageNode<_, _, _>"),
+			inputs: vec![
+				DocumentInputType::value("Image", TaggedValue::ImageFrame(ImageFrame::empty()), true),
+				DocumentInputType::value("Second", TaggedValue::ImageFrame(ImageFrame::empty()), true),
+				DocumentInputType::value("Blend Mode", TaggedValue::BlendMode(BlendMode::Normal), false),
+				DocumentInputType::value("Opacity", TaggedValue::F32(100.0), false),
+			],
+			outputs: vec![DocumentOutputType::new("Image", FrontendGraphDataType::Raster)],
+			properties: node_properties::blend_properties,
+		},
 		DocumentNodeType {
 			name: "Extract",
 			category: "Macros",

node-graph/compilation-client/Cargo.toml

@@ -12,6 +12,7 @@ serde_json = "1.0"
 graph-craft = { version = "0.1.0", path = "../graph-craft", features = [
 	"serde",
 ] }
+graphene-core = { version = "0.1.0", path = "../gcore" }
 gpu-executor = { version = "0.1.0", path = "../gpu-executor" }
 gpu-compiler-bin-wrapper = { version = "0.1.0", path = "../gpu-compiler/gpu-compiler-bin-wrapper" }
 tempfile = "3.3.0"

node-graph/compilation-client/src/main.rs

@@ -1,8 +1,11 @@
 use gpu_compiler_bin_wrapper::CompileRequest;
 use gpu_executor::{ShaderIO, ShaderInput};
 use graph_craft::concrete;
+use graph_craft::document::value::TaggedValue;
 use graph_craft::document::*;
 use graph_craft::*;
+use graphene_core::raster::adjustments::{BlendMode, BlendNode};
+use graphene_core::Color;
 
 use std::borrow::Cow;
 use std::time::Duration;
@@ -10,33 +13,21 @@ use std::time::Duration;
 fn main() {
 	let client = reqwest::blocking::Client::new();
 
-	// let network = NodeNetwork {
-	// 	inputs: vec![0],
-	// 	outputs: vec![NodeOutput::new(0, 0)],
-	// 	disabled: vec![],
-	// 	previous_outputs: None,
-	// 	nodes: [(
-	// 		0,
-	// 		DocumentNode {
-	// 			name: "Inc".into(),
-	// 			inputs: vec![NodeInput::Network(concrete!(u32))],
-	// 			implementation: DocumentNodeImplementation::Network(add_network()),
-	// 			metadata: DocumentNodeMetadata::default(),
-	// 		},
-	// 	)]
-	// 	.into_iter()
-	// 	.collect(),
-	// };
 	let network = add_network();
 	let compiler = graph_craft::executor::Compiler {};
 	let proto_network = compiler.compile_single(network, true).unwrap();
 
 	let io = ShaderIO {
-		inputs: vec![ShaderInput::StorageBuffer((), concrete!(u32))],
-		output: ShaderInput::OutputBuffer((), concrete!(&mut [u32])),
+		inputs: vec![
+			ShaderInput::StorageBuffer((), concrete!(Color)), // background image
+			ShaderInput::StorageBuffer((), concrete!(Color)), // foreground image
+			ShaderInput::StorageBuffer((), concrete!(u32)),   // width/height of the foreground image
+			ShaderInput::OutputBuffer((), concrete!(Color)),
+		],
+		output: ShaderInput::OutputBuffer((), concrete!(Color)),
 	};
 
-	let compile_request = CompileRequest::new(vec![proto_network], vec![concrete!(u32)], vec![concrete!(u32)], io);
+	let compile_request = CompileRequest::new(vec![proto_network], vec![concrete!(Color), concrete!(Color), concrete!(u32)], vec![concrete!(Color)], io);
 	let response = client
 		.post("http://localhost:3000/compile/spirv")
 		.timeout(Duration::from_secs(30))
@@ -52,27 +43,33 @@ fn add_network() -> NodeNetwork {
 		outputs: vec![NodeOutput::new(0, 0)],
 		disabled: vec![],
 		previous_outputs: None,
-		nodes: [
-			(
-				0,
-				DocumentNode {
-					name: "Dup".into(),
-					inputs: vec![NodeInput::value(value::TaggedValue::U32(5u32), false)],
-					implementation: DocumentNodeImplementation::Unresolved(NodeIdentifier::new("graphene_core::ops::IdNode")),
-					..Default::default()
-				},
-			),
-			// (
-			// 	1,
-			// 	DocumentNode {
-			// 		name: "Add".into(),
-			// 		inputs: vec![NodeInput::node(0, 0)],
-			// 		metadata: DocumentNodeMetadata::default(),
-			// 		implementation: DocumentNodeImplementation::Unresolved(NodeIdentifier::new("graphene_core::ops::AddNode")),
-			// 	},
-			// ),
-		]
+		nodes: [DocumentNode {
+			name: "Blend Image".into(),
+			inputs: vec![NodeInput::Inline(InlineRust::new(
+				format!(
+					r#"graphene_core::raster::adjustments::BlendNode::new(
+							graphene_core::value::CopiedNode::new({}),
+							graphene_core::value::CopiedNode::new({}),
+						).eval((
+							i1[_global_index.x as usize],
+							if _global_index.x < i2[2] {{
+								i0[_global_index.x as usize]
+							}} else {{
+								Color::from_rgbaf32_unchecked(0.0, 0.0, 0.0, 0.0)
+							}},
+						))"#,
+					TaggedValue::BlendMode(BlendMode::Normal).to_primitive_string(),
+					TaggedValue::F32(1.0).to_primitive_string(),
+				),
+				concrete![Color],
+			))],
+			implementation: DocumentNodeImplementation::Unresolved("graphene_core::value::CopiedNode".into()),
+			..Default::default()
+		}]
 		.into_iter()
+		.enumerate()
+		.map(|(i, n)| (i as u64, n))
 		.collect(),
+		..Default::default()
 	}
 }

node-graph/gcore/src/raster/adjustments.rs

@@ -84,6 +84,7 @@ impl BlendMode {
 #[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
 #[cfg_attr(feature = "std", derive(specta::Type))]
 #[derive(Debug, Default, Clone, Copy, Eq, PartialEq, DynAny, Hash)]
+#[repr(i32)] // TODO: Enable Int8 capability for SPRIV so that we don't need this?
 pub enum BlendMode {
 	#[default]
 	// Basic group
@@ -173,6 +174,46 @@ impl core::fmt::Display for BlendMode {
 	}
 }
 
+pub fn to_primtive_string(blend_mode: &BlendMode) -> &'static str {
+	match blend_mode {
+		BlendMode::Normal => "Normal",
+
+		BlendMode::Multiply => "Multiply",
+		BlendMode::Darken => "Darken",
+		BlendMode::ColorBurn => "ColorBurn",
+		BlendMode::LinearBurn => "LinearBurn",
+		BlendMode::DarkerColor => "DarkerColor",
+
+		BlendMode::Screen => "Screen",
+		BlendMode::Lighten => "Lighten",
+		BlendMode::ColorDodge => "ColorDodge",
+		BlendMode::LinearDodge => "LinearDodge",
+		BlendMode::LighterColor => "LighterColor",
+
+		BlendMode::Overlay => "Overlay",
+		BlendMode::SoftLight => "SoftLight",
+		BlendMode::HardLight => "HardLight",
+		BlendMode::VividLight => "VividLight",
+		BlendMode::LinearLight => "LinearLight",
+		BlendMode::PinLight => "PinLight",
+		BlendMode::HardMix => "HardMix",
+
+		BlendMode::Difference => "Difference",
+		BlendMode::Exclusion => "Exclusion",
+		BlendMode::Subtract => "Subtract",
+		BlendMode::Divide => "Divide",
+
+		BlendMode::Hue => "Hue",
+		BlendMode::Saturation => "Saturation",
+		BlendMode::Color => "Color",
+		BlendMode::Luminosity => "Luminosity",
+
+		BlendMode::InsertRed => "InsertRed",
+		BlendMode::InsertGreen => "InsertGreen",
+		BlendMode::InsertBlue => "InsertBlue",
+	}
+}
+
 #[derive(Debug, Clone, Copy, Default)]
 pub struct LuminanceNode<LuminanceCalculation> {
 	luminance_calc: LuminanceCalculation,
@@ -410,7 +451,7 @@ pub struct BlendNode<BlendMode, Opacity> {
 }
 
 #[node_macro::node_fn(BlendNode)]
-fn blend_node(input: (Color, Color), blend_mode: BlendMode, opacity: f64) -> Color {
+fn blend_node(input: (Color, Color), blend_mode: BlendMode, opacity: f32) -> Color {
 	let opacity = opacity / 100.;
 
 	let (foreground, background) = input;
@@ -452,7 +493,7 @@ fn blend_node(input: (Color, Color), blend_mode: BlendMode, opacity: f64) -> Col
 		BlendMode::InsertBlue => foreground.with_blue(background.b()),
 	};
 
-	background.alpha_blend(target_color.to_associated_alpha(opacity as f32))
+	background.alpha_blend(target_color.to_associated_alpha(opacity))
 }
 
 #[derive(Debug, Clone, Copy)]

node-graph/gcore/src/types.rs

@@ -34,6 +34,19 @@ macro_rules! concrete {
 		})
 	};
 }
+
+#[macro_export]
+macro_rules! concrete_with_name {
+	($type:ty, $name:expr) => {
+		Type::Concrete(TypeDescriptor {
+			id: Some(core::any::TypeId::of::<$type>()),
+			name: Cow::Borrowed($name),
+			size: core::mem::size_of::<$type>(),
+			align: core::mem::align_of::<$type>(),
+		})
+	};
+}
+
 #[macro_export]
 macro_rules! generic {
 	($type:ty) => {{

node-graph/gpu-compiler/gpu-compiler-bin-wrapper/src/lib.rs

@@ -17,7 +17,7 @@ pub fn compile_spirv(request: &CompileRequest, compile_dir: Option<&str>, manife
 
 	println!("calling cargo run!");
 	let non_cargo_env_vars = std::env::vars().filter(|(k, _)| k.starts_with("PATH")).collect::<Vec<_>>();
-	let mut cargo_command = std::process::Command::new("/usr/bin/cargo")
+	let mut cargo_command = std::process::Command::new("cargo")
 		.arg("run")
 		.arg("--release")
 		.arg("--manifest-path")

node-graph/gpu-compiler/src/lib.rs

@@ -71,7 +71,7 @@ pub fn construct_argument(input: &ShaderInput<()>, position: u32, binding_offset
 	let line = match input {
 		ShaderInput::Constant(constant) => format!("#[spirv({})] i{}: {}", constant_attribute(constant), position, constant.ty()),
 		ShaderInput::UniformBuffer(_, ty) => {
-			format!("#[spirv(uniform, descriptor_set = 0, binding = {})] i{}: &[{}]", position + binding_offset, position, ty,)
+			format!("#[spirv(uniform, descriptor_set = 0, binding = {})] i{}: &{}", position + binding_offset, position, ty,)
 		}
 		ShaderInput::StorageBuffer(_, ty) | ShaderInput::ReadBackBuffer(_, ty) => {
 			format!("#[spirv(storage_buffer, descriptor_set = 0, binding = {})] i{}: &[{}]", position + binding_offset, position, ty,)

node-graph/gpu-compiler/src/templates/spirv-template.rs

@@ -8,7 +8,8 @@ extern crate spirv_std;
 //pub mod gpu {
 //use super::*;
 	use spirv_std::spirv;
-	use spirv_std::glam::UVec3;
+	use spirv_std::glam;
+	use spirv_std::glam::{UVec3, Vec2, Mat2, BVec2};
 
 	#[allow(unused)]
 	#[spirv(compute(threads({{compute_threads}})))]
@@ -19,6 +20,8 @@ extern crate spirv_std;
 		{% endfor %}
 	) {
 		use graphene_core::Node;
+		use graphene_core::raster::adjustments::{BlendMode, BlendNode};
+		use graphene_core::Color;
 
 		/*
 		{% for input in input_nodes %}
@@ -34,7 +37,7 @@ extern crate spirv_std;
 
 		{% for output in output_nodes %}
 		let v = {{output}}.eval(());
-		o{{loop.index0}}[_global_index.x as usize] = v;
+		o{{loop.index0}}[(_global_index.y * i0 + _global_index.x) as usize] = v;
 		{% endfor %}
 		// TODO: Write output to buffer
 	}

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

@@ -13,6 +13,22 @@ use std::sync::Arc;
 
 type ReadBackFuture = Pin<Box<dyn Future<Output = Result<Vec<u8>>>>>;
 
+pub enum ComputePassDimensions {
+	X(u32),
+	XY(u32, u32),
+	XYZ(u32, u32, u32),
+}
+
+impl ComputePassDimensions {
+	pub fn get(&self) -> (u32, u32, u32) {
+		match self {
+			ComputePassDimensions::X(x) => (*x, 1, 1),
+			ComputePassDimensions::XY(x, y) => (*x, *y, 1),
+			ComputePassDimensions::XYZ(x, y, z) => (*x, *y, *z),
+		}
+	}
+}
+
 pub trait GpuExecutor {
 	type ShaderHandle;
 	type BufferHandle;
@@ -22,7 +38,7 @@ pub trait GpuExecutor {
 	fn create_uniform_buffer<T: ToUniformBuffer>(&self, data: T) -> Result<ShaderInput<Self::BufferHandle>>;
 	fn create_storage_buffer<T: ToStorageBuffer>(&self, data: T, options: StorageBufferOptions) -> Result<ShaderInput<Self::BufferHandle>>;
 	fn create_output_buffer(&self, len: usize, ty: Type, cpu_readable: bool) -> Result<ShaderInput<Self::BufferHandle>>;
-	fn create_compute_pass(&self, layout: &PipelineLayout<Self>, read_back: Option<Arc<ShaderInput<Self::BufferHandle>>>, instances: u32) -> Result<Self::CommandBuffer>;
+	fn create_compute_pass(&self, layout: &PipelineLayout<Self>, read_back: Option<Arc<ShaderInput<Self::BufferHandle>>>, instances: ComputePassDimensions) -> Result<Self::CommandBuffer>;
 	fn execute_compute_pipeline(&self, encoder: Self::CommandBuffer) -> Result<()>;
 	fn read_output_buffer(&self, buffer: Arc<ShaderInput<Self::BufferHandle>>) -> ReadBackFuture;
 }
@@ -129,10 +145,15 @@ pub struct StorageBufferOptions {
 }
 
 pub trait ToUniformBuffer: StaticType {
-	type UniformBufferHandle;
 	fn to_bytes(&self) -> Cow<[u8]>;
 }
 
+impl<T: StaticType + Pod + Zeroable> ToUniformBuffer for T {
+	fn to_bytes(&self) -> Cow<[u8]> {
+		Cow::Owned(bytemuck::bytes_of(self).into())
+	}
+}
+
 pub trait ToStorageBuffer: StaticType {
 	fn to_bytes(&self) -> Cow<[u8]>;
 	fn ty(&self) -> Type;
@@ -233,7 +254,12 @@ pub struct CreateComputePassNode<Executor, Output, Instances> {
 }
 
 #[node_macro::node_fn(CreateComputePassNode)]
-fn create_compute_pass_node<E: GpuExecutor + 'input>(layout: PipelineLayout<E>, executor: &'input E, output: ShaderInput<E::BufferHandle>, instances: u32) -> E::CommandBuffer {
+fn create_compute_pass_node<'any_input, E: 'any_input + GpuExecutor>(
+	layout: PipelineLayout<E>,
+	executor: &'any_input E,
+	output: ShaderInput<E::BufferHandle>,
+	instances: ComputePassDimensions,
+) -> E::CommandBuffer {
 	executor.create_compute_pass(&layout, Some(output.into()), instances).unwrap()
 }
 

node-graph/graph-craft/src/document/value.rs

@@ -3,7 +3,7 @@ use crate::executor::Any;
 pub use crate::imaginate_input::{ImaginateMaskStartingFill, ImaginateSamplingMethod, ImaginateStatus};
 use crate::proto::{Any as DAny, FutureAny};
 
-use graphene_core::raster::{BlendMode, LuminanceCalculation};
+use graphene_core::raster::{to_primtive_string, BlendMode, LuminanceCalculation};
 use graphene_core::{Color, Node, Type};
 
 pub use dyn_any::StaticType;
@@ -185,10 +185,11 @@ impl<'a> TaggedValue {
 		match self {
 			TaggedValue::None => "()".to_string(),
 			TaggedValue::String(x) => format!("\"{}\"", x),
-			TaggedValue::U32(x) => x.to_string(),
-			TaggedValue::F32(x) => x.to_string(),
-			TaggedValue::F64(x) => x.to_string(),
+			TaggedValue::U32(x) => x.to_string() + "_u32",
+			TaggedValue::F32(x) => x.to_string() + "_f32",
+			TaggedValue::F64(x) => x.to_string() + "_f64",
 			TaggedValue::Bool(x) => x.to_string(),
+			TaggedValue::BlendMode(blend_mode) => "BlendMode::".to_string() + to_primtive_string(blend_mode),
 			_ => panic!("Cannot convert to primitive string"),
 		}
 	}

node-graph/gstd/src/executor.rs

@@ -1,11 +1,11 @@
-use glam::UVec3;
-use gpu_executor::{Bindgroup, PipelineLayout, StorageBufferOptions};
+use glam::{DAffine2, DMat2, DVec2, Mat2, UVec3, Vec2};
+use gpu_executor::{Bindgroup, ComputePassDimensions, PipelineLayout, StorageBufferOptions};
 use gpu_executor::{GpuExecutor, ShaderIO, ShaderInput};
 use graph_craft::document::value::TaggedValue;
 use graph_craft::document::*;
 use graph_craft::proto::*;
+use graphene_core::raster::bbox::{AxisAlignedBbox, Bbox};
 use graphene_core::raster::*;
-use graphene_core::value::ValueNode;
 use graphene_core::*;
 use wgpu_executor::NewExecutor;
 
@@ -58,7 +58,7 @@ async fn map_gpu(image: ImageFrame<Color>, node: DocumentNode) -> ImageFrame<Col
 		nodes: [
 			DocumentNode {
 				name: "Slice".into(),
-				inputs: vec![NodeInput::Inline(InlineRust::new("i0[_global_index.x as usize]".into(), concrete![Color]))],
+				inputs: vec![NodeInput::Inline(InlineRust::new("i1[(_global_index.y * i0 + _global_index.x) as usize]".into(), concrete![Color]))],
 				implementation: DocumentNodeImplementation::Unresolved("graphene_core::value::CopiedNode".into()),
 				..Default::default()
 			},
@@ -108,10 +108,11 @@ async fn map_gpu(image: ImageFrame<Color>, node: DocumentNode) -> ImageFrame<Col
 	log::debug!("compiling shader");
 	let shader = compilation_client::compile(
 		proto_networks,
-		vec![concrete!(Color)], //, concrete!(u32)],
+		vec![concrete!(u32), concrete!(Color)], //, concrete!(u32)],
 		vec![concrete!(Color)],
 		ShaderIO {
 			inputs: vec![
+				ShaderInput::UniformBuffer((), concrete!(u32)),
 				ShaderInput::StorageBuffer((), concrete!(Color)),
 				//ShaderInput::Constant(gpu_executor::GPUConstant::GlobalInvocationId),
 				ShaderInput::OutputBuffer((), concrete!(Color)),
@@ -122,11 +123,11 @@ async fn map_gpu(image: ImageFrame<Color>, node: DocumentNode) -> ImageFrame<Col
 	.await
 	.unwrap();
 	//return ImageFrame::empty();
-	let len = image.image.data.len();
-	log::debug!("instances: {}", len);
+	let len: usize = image.image.data.len();
 
 	let executor = NewExecutor::new().await.unwrap();
 	log::debug!("creating buffer");
+	let width_uniform = executor.create_uniform_buffer(image.image.width).unwrap();
 	let storage_buffer = executor
 		.create_storage_buffer(
 			image.image.data.clone(),
@@ -138,6 +139,7 @@ async fn map_gpu(image: ImageFrame<Color>, node: DocumentNode) -> ImageFrame<Col
 			},
 		)
 		.unwrap();
+	let width_uniform = Arc::new(width_uniform);
 	let storage_buffer = Arc::new(storage_buffer);
 	let output_buffer = executor.create_output_buffer(len, concrete!(Color), false).unwrap();
 	let output_buffer = Arc::new(output_buffer);
@@ -145,7 +147,7 @@ async fn map_gpu(image: ImageFrame<Color>, node: DocumentNode) -> ImageFrame<Col
 	let readback_buffer = Arc::new(readback_buffer);
 	log::debug!("created buffer");
 	let bind_group = Bindgroup {
-		buffers: vec![storage_buffer.clone()],
+		buffers: vec![width_uniform.clone(), storage_buffer.clone()],
 	};
 
 	let shader = gpu_executor::Shader {
@@ -164,7 +166,9 @@ async fn map_gpu(image: ImageFrame<Color>, node: DocumentNode) -> ImageFrame<Col
 		output_buffer: output_buffer.clone(),
 	};
 	log::debug!("created pipeline");
-	let compute_pass = executor.create_compute_pass(&pipeline, Some(readback_buffer.clone()), len as u32).unwrap();
+	let compute_pass = executor
+		.create_compute_pass(&pipeline, Some(readback_buffer.clone()), ComputePassDimensions::XY(image.image.width, image.image.height))
+		.unwrap();
 	executor.execute_compute_pipeline(compute_pass).unwrap();
 	log::debug!("executed pipeline");
 	log::debug!("reading buffer");
@@ -243,3 +247,183 @@ fn map_gpu_single_image(input: Image<Color>, node: String) -> Image<Color> {
 	Image { data, ..input }
 }
 */
+
+#[derive(Debug, Clone, Copy)]
+pub struct BlendGpuImageNode<Background, B, O> {
+	background: Background,
+	blend_mode: B,
+	opacity: O,
+}
+
+#[node_macro::node_fn(BlendGpuImageNode)]
+async fn blend_gpu_image(foreground: ImageFrame<Color>, background: ImageFrame<Color>, blend_mode: BlendMode, opacity: f32) -> ImageFrame<Color> {
+	let foreground_size = DVec2::new(foreground.image.width as f64, foreground.image.height as f64);
+	let background_size = DVec2::new(background.image.width as f64, background.image.height as f64);
+	// Transforms a point from the background image to the forground image
+	let bg_to_fg = DAffine2::from_scale(foreground_size) * foreground.transform.inverse() * background.transform * DAffine2::from_scale(1. / background_size);
+
+	let transform_matrix: Mat2 = bg_to_fg.matrix2.as_mat2();
+	let translation: Vec2 = bg_to_fg.translation.as_vec2();
+
+	log::debug!("Executing gpu blend node!");
+	let compiler = graph_craft::executor::Compiler {};
+
+	let network = NodeNetwork {
+		inputs: vec![],
+		outputs: vec![NodeOutput::new(0, 0)],
+		nodes: [DocumentNode {
+			name: "BlendOp".into(),
+			inputs: vec![NodeInput::Inline(InlineRust::new(
+				format!(
+					r#"graphene_core::raster::adjustments::BlendNode::new(
+							graphene_core::value::CopiedNode::new({}),
+							graphene_core::value::CopiedNode::new({}),
+						).eval((
+							{{
+								let bg_point = Vec2::new(_global_index.x as f32, _global_index.y as f32);
+								let fg_point = (*i4) * bg_point + (*i5);
+
+								if !((fg_point.cmpge(Vec2::ZERO) & bg_point.cmpge(Vec2::ZERO)) == BVec2::new(true, true)) {{
+									Color::from_rgbaf32_unchecked(0.0, 0.0, 0.0, 0.0)
+								}} else {{
+									i2[((fg_point.y as u32) * i3 + (fg_point.x as u32)) as usize]
+								}}
+							}},
+							i1[(_global_index.y * i0 + _global_index.x) as usize],
+						))"#,
+					TaggedValue::BlendMode(blend_mode).to_primitive_string(),
+					TaggedValue::F32(opacity).to_primitive_string(),
+				),
+				concrete![Color],
+			))],
+			implementation: DocumentNodeImplementation::Unresolved("graphene_core::value::CopiedNode".into()),
+			..Default::default()
+		}]
+		.into_iter()
+		.enumerate()
+		.map(|(i, n)| (i as u64, n))
+		.collect(),
+		..Default::default()
+	};
+	log::debug!("compiling network");
+	let proto_networks = compiler.compile(network.clone(), true).collect();
+	log::debug!("compiling shader");
+
+	let shader = compilation_client::compile(
+		proto_networks,
+		vec![
+			concrete!(u32),
+			concrete!(Color),
+			concrete!(Color),
+			concrete!(u32),
+			concrete_with_name!(Mat2, "Mat2"),
+			concrete_with_name!(Vec2, "Vec2"),
+		],
+		vec![concrete!(Color)],
+		ShaderIO {
+			inputs: vec![
+				ShaderInput::UniformBuffer((), concrete!(u32)),                    // width of the output image
+				ShaderInput::StorageBuffer((), concrete!(Color)),                  // background image
+				ShaderInput::StorageBuffer((), concrete!(Color)),                  // foreground image
+				ShaderInput::UniformBuffer((), concrete!(u32)),                    // width of the foreground image
+				ShaderInput::UniformBuffer((), concrete_with_name!(Mat2, "Mat2")), // bg_to_fg.matrix2
+				ShaderInput::UniformBuffer((), concrete_with_name!(Vec2, "Vec2")), // bg_to_fg.translation
+				ShaderInput::OutputBuffer((), concrete!(Color)),
+			],
+			output: ShaderInput::OutputBuffer((), concrete!(Color)),
+		},
+	)
+	.await
+	.unwrap();
+	let len = background.image.data.len();
+
+	let executor = NewExecutor::new()
+		.await
+		.expect("Failed to create wgpu executor. Please make sure that webgpu is enabled for your browser.");
+	log::debug!("creating buffer");
+	let width_uniform = executor.create_uniform_buffer(background.image.width).unwrap();
+	let bg_storage_buffer = executor
+		.create_storage_buffer(
+			background.image.data.clone(),
+			StorageBufferOptions {
+				cpu_writable: false,
+				gpu_writable: true,
+				cpu_readable: false,
+				storage: true,
+			},
+		)
+		.unwrap();
+	let fg_storage_buffer = executor
+		.create_storage_buffer(
+			foreground.image.data.clone(),
+			StorageBufferOptions {
+				cpu_writable: false,
+				gpu_writable: true,
+				cpu_readable: false,
+				storage: true,
+			},
+		)
+		.unwrap();
+	let fg_width_uniform = executor.create_uniform_buffer(foreground.image.width).unwrap();
+	let transform_uniform = executor.create_uniform_buffer(transform_matrix).unwrap();
+	let translation_uniform = executor.create_uniform_buffer(translation).unwrap();
+	let width_uniform = Arc::new(width_uniform);
+	let bg_storage_buffer = Arc::new(bg_storage_buffer);
+	let fg_storage_buffer = Arc::new(fg_storage_buffer);
+	let fg_width_uniform = Arc::new(fg_width_uniform);
+	let transform_uniform = Arc::new(transform_uniform);
+	let translation_uniform = Arc::new(translation_uniform);
+	let output_buffer = executor.create_output_buffer(len, concrete!(Color), false).unwrap();
+	let output_buffer = Arc::new(output_buffer);
+	let readback_buffer = executor.create_output_buffer(len, concrete!(Color), true).unwrap();
+	let readback_buffer = Arc::new(readback_buffer);
+	log::debug!("created buffer");
+	let bind_group = Bindgroup {
+		buffers: vec![
+			width_uniform.clone(),
+			bg_storage_buffer.clone(),
+			fg_storage_buffer.clone(),
+			fg_width_uniform.clone(),
+			transform_uniform.clone(),
+			translation_uniform.clone(),
+		],
+	};
+
+	let shader = gpu_executor::Shader {
+		source: shader.spirv_binary.into(),
+		name: "gpu::eval",
+		io: shader.io,
+	};
+	log::debug!("loading shader");
+	log::debug!("shader: {:?}", shader.source);
+	let shader = executor.load_shader(shader).unwrap();
+	log::debug!("loaded shader");
+	let pipeline = PipelineLayout {
+		shader,
+		entry_point: "eval".to_string(),
+		bind_group,
+		output_buffer: output_buffer.clone(),
+	};
+	log::debug!("created pipeline");
+	let compute_pass = executor
+		.create_compute_pass(
+			&pipeline,
+			Some(readback_buffer.clone()),
+			ComputePassDimensions::XY(background.image.width as u32, background.image.height as u32),
+		)
+		.unwrap();
+	executor.execute_compute_pipeline(compute_pass).unwrap();
+	log::debug!("executed pipeline");
+	log::debug!("reading buffer");
+	let result = executor.read_output_buffer(readback_buffer).await.unwrap();
+	let colors = bytemuck::pod_collect_to_vec::<u8, Color>(result.as_slice());
+
+	ImageFrame {
+		image: Image {
+			data: colors,
+			width: background.image.width,
+			height: background.image.height,
+		},
+		transform: background.transform,
+	}
+}

node-graph/interpreted-executor/src/node_registry.rs

@@ -223,6 +223,26 @@ fn node_registry() -> HashMap<NodeIdentifier, HashMap<NodeIOTypes, NodeConstruct
 			},
 			NodeIOTypes::new(concrete!(ImageFrame<Color>), concrete!(ImageFrame<Color>), vec![value_fn!(DocumentNode)]),
 		)],
+		#[cfg(feature = "gpu")]
+		vec![(
+			NodeIdentifier::new("graphene_std::executor::BlendGpuImageNode<_, _, _>"),
+			|args| {
+				Box::pin(async move {
+					let background: DowncastBothNode<(), ImageFrame<Color>> = DowncastBothNode::new(args[0]);
+					let blend_mode: DowncastBothNode<(), BlendMode> = DowncastBothNode::new(args[1]);
+					let opacity: DowncastBothNode<(), f32> = DowncastBothNode::new(args[2]);
+					let node = graphene_std::executor::BlendGpuImageNode::new(background, blend_mode, opacity);
+					let any: DynAnyNode<ImageFrame<Color>, _, _> = graphene_std::any::DynAnyNode::new(graphene_core::value::ValueNode::new(node));
+
+					Box::pin(any) as TypeErasedPinned
+				})
+			},
+			NodeIOTypes::new(
+				concrete!(ImageFrame<Color>),
+				concrete!(ImageFrame<Color>),
+				vec![value_fn!(ImageFrame<Color>), value_fn!(BlendMode), value_fn!(f32)],
+			),
+		)],
 		vec![(
 			NodeIdentifier::new("graphene_core::structural::ComposeNode<_, _, _>"),
 			|args| {
@@ -326,7 +346,7 @@ fn node_registry() -> HashMap<NodeIdentifier, HashMap<NodeIOTypes, NodeConstruct
 				Box::pin(async move {
 					let image: DowncastBothNode<(), ImageFrame<Color>> = DowncastBothNode::new(args[0]);
 					let blend_mode: DowncastBothNode<(), BlendMode> = DowncastBothNode::new(args[1]);
-					let opacity: DowncastBothNode<(), f64> = DowncastBothNode::new(args[2]);
+					let opacity: DowncastBothNode<(), f32> = DowncastBothNode::new(args[2]);
 					let blend_node = graphene_core::raster::BlendNode::new(CopiedNode::new(blend_mode.eval(()).await), CopiedNode::new(opacity.eval(()).await));
 					let node = graphene_std::raster::BlendImageNode::new(image, FutureWrapperNode::new(ValueNode::new(blend_node)));
 					let any: DynAnyNode<ImageFrame<Color>, _, _> = graphene_std::any::DynAnyNode::new(graphene_core::value::ValueNode::new(node));
@@ -336,7 +356,7 @@ fn node_registry() -> HashMap<NodeIdentifier, HashMap<NodeIOTypes, NodeConstruct
 			NodeIOTypes::new(
 				concrete!(ImageFrame<Color>),
 				concrete!(ImageFrame<Color>),
-				vec![value_fn!(ImageFrame<Color>), value_fn!(BlendMode), value_fn!(f64)],
+				vec![value_fn!(ImageFrame<Color>), value_fn!(BlendMode), value_fn!(f32)],
 			),
 		)],
 		raster_node!(graphene_core::raster::GrayscaleNode<_, _, _, _, _, _, _>, params: [Color, f64, f64, f64, f64, f64, f64]),

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

@@ -3,7 +3,7 @@ mod executor;
 
 pub use context::Context;
 pub use executor::GpuExecutor;
-use gpu_executor::{Shader, ShaderInput, StorageBufferOptions, ToStorageBuffer, ToUniformBuffer};
+use gpu_executor::{ComputePassDimensions, Shader, ShaderInput, StorageBufferOptions, ToStorageBuffer, ToUniformBuffer};
 use graph_craft::Type;
 
 use anyhow::{bail, Result};
@@ -83,7 +83,7 @@ impl gpu_executor::GpuExecutor for NewExecutor {
 		};
 		Ok(buffer)
 	}
-	fn create_compute_pass(&self, layout: &gpu_executor::PipelineLayout<Self>, read_back: Option<Arc<ShaderInput<Self::BufferHandle>>>, instances: u32) -> Result<CommandBuffer> {
+	fn create_compute_pass(&self, layout: &gpu_executor::PipelineLayout<Self>, read_back: Option<Arc<ShaderInput<Self::BufferHandle>>>, instances: ComputePassDimensions) -> Result<CommandBuffer> {
 		let compute_pipeline = self.context.device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
 			label: None,
 			layout: None,
@@ -113,11 +113,12 @@ impl gpu_executor::GpuExecutor for NewExecutor {
 
 		let mut encoder = self.context.device.create_command_encoder(&wgpu::CommandEncoderDescriptor { label: None });
 		{
+			let dimensions = instances.get();
 			let mut cpass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor { label: None });
 			cpass.set_pipeline(&compute_pipeline);
 			cpass.set_bind_group(0, &bind_group, &[]);
 			cpass.insert_debug_marker("compute node network evaluation");
-			cpass.dispatch_workgroups(instances, 1, 1); // Number of cells to run, the (x,y,z) size of item being processed
+			cpass.dispatch_workgroups(dimensions.0, dimensions.1, dimensions.2); // Number of cells to run, the (x,y,z) size of item being processed
 		}
 		// Sets adds copy operation to command encoder.
 		// Will copy data from storage buffer on GPU to staging buffer on CPU.