Graphite/node-graph/libraries/wgpu-executor/src/texture_conversion.rs

use crate::{Buffer, WgpuExecutor};
use core_types::Color;
use core_types::color::SRGBA8;
use core_types::list::{Item, List};
use core_types::ops::Convert;
use core_types::transform::Footprint;
use raster_types::Image;
use raster_types::{CPU, GPU, Raster, Texture};
use wgpu::{Extent3d, TextureFormat};

/// Uploads CPU image data to a GPU texture
fn upload_to_texture(executor: &WgpuExecutor, queue: &wgpu::Queue, image: &Raster<CPU>) -> Texture {
	let rgba8_data: Vec<SRGBA8> = image.data.iter().map(|x| (*x).into()).collect();

	let texture = executor.request_texture_with_format(glam::UVec2::new(image.width, image.height), TextureFormat::Rgba8UnormSrgb);
	queue.write_texture(
		texture.as_image_copy(),
		bytemuck::cast_slice(rgba8_data.as_slice()),
		wgpu::TexelCopyBufferLayout {
			offset: 0,
			bytes_per_row: Some(4 * image.width),
			rows_per_image: Some(image.height),
		},
		Extent3d {
			width: image.width,
			height: image.height,
			depth_or_array_layers: 1,
		},
	);
	texture
}

/// Converts a Raster<GPU> texture to Raster<CPU> by downloading the underlying texture data.
///
/// Assumptions:
/// - 2D texture, mip level 0
/// - 4 bytes-per-pixel RGBA8
/// - Texture has COPY_SRC usage
struct RasterGpuToRasterCpuConverter {
	buffer: Buffer,
	width: u32,
	height: u32,
	unpadded_bytes_per_row: u32,
	padded_bytes_per_row: u32,
	_source: raster_types::Texture,
}
impl RasterGpuToRasterCpuConverter {
	fn new(executor: &WgpuExecutor, encoder: &mut wgpu::CommandEncoder, data_gpu: Raster<GPU>) -> Self {
		let texture = data_gpu.data();
		let width = texture.width();
		let height = texture.height();
		let bytes_per_pixel = 4; // RGBA8
		let unpadded_bytes_per_row = width * bytes_per_pixel;
		let align = wgpu::COPY_BYTES_PER_ROW_ALIGNMENT;
		let padded_bytes_per_row = unpadded_bytes_per_row.div_ceil(align) * align;
		let buffer_size = padded_bytes_per_row as u64 * height as u64;

		let buffer = executor.create_buffer(&wgpu::BufferDescriptor {
			label: Some("texture_download_buffer"),
			size: buffer_size,
			usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::MAP_READ,
			mapped_at_creation: false,
		});

		encoder.copy_texture_to_buffer(
			wgpu::TexelCopyTextureInfo {
				texture,
				mip_level: 0,
				origin: wgpu::Origin3d::ZERO,
				aspect: wgpu::TextureAspect::All,
			},
			wgpu::TexelCopyBufferInfo {
				buffer: &buffer,
				layout: wgpu::TexelCopyBufferLayout {
					offset: 0,
					bytes_per_row: Some(padded_bytes_per_row),
					rows_per_image: Some(height),
				},
			},
			Extent3d {
				width,
				height,
				depth_or_array_layers: 1,
			},
		);

		Self {
			buffer,
			width,
			height,
			unpadded_bytes_per_row,
			padded_bytes_per_row,
			// Keep source texture alive
			_source: data_gpu.texture.clone(),
		}
	}

	async fn convert(self, device: &wgpu::Device) -> Result<Raster<CPU>, wgpu::BufferAsyncError> {
		let buffer_slice = self.buffer.slice(..);
		let (sender, receiver) = futures::channel::oneshot::channel();
		buffer_slice.map_async(wgpu::MapMode::Read, move |result| {
			let _ = sender.send(result);
		});

		let _ = device.poll(wgpu::wgt::PollType::wait_indefinitely());

		receiver.await.expect("Failed to receive map result")?;

		let view = buffer_slice.get_mapped_range();

		let row_stride = self.padded_bytes_per_row as usize;
		let row_bytes = self.unpadded_bytes_per_row as usize;
		let mut cpu_data: Vec<Color> = Vec::with_capacity((self.width * self.height) as usize);
		for row in 0..self.height as usize {
			let start = row * row_stride;
			let row_slice = &view[start..start + row_bytes];
			for px in row_slice.chunks_exact(4) {
				cpu_data.push(SRGBA8::new(px[0], px[1], px[2], px[3]).into());
			}
		}

		drop(view);
		self.buffer.unmap();
		let cpu_image = Image {
			data: cpu_data,
			width: self.width,
			height: self.height,
			base64_string: None,
		};

		Ok(Raster::new_cpu(cpu_image))
	}
}

/// Passthrough conversion for GPU `List`s - no conversion needed
impl<'i> Convert<List<Raster<GPU>>, &'i WgpuExecutor> for List<Raster<GPU>> {
	async fn convert(self, _: Footprint, _converter: &'i WgpuExecutor) -> List<Raster<GPU>> {
		self
	}
}

/// Converts a `List<Raster<CPU>>` to `List<Raster<GPU>>` by uploading each image to a texture
impl<'i> Convert<List<Raster<GPU>>, &'i WgpuExecutor> for List<Raster<CPU>> {
	async fn convert(self, _: Footprint, executor: &'i WgpuExecutor) -> List<Raster<GPU>> {
		let queue = executor.context().queue.lock();
		let list = self
			.into_iter()
			.map(|row| {
				let (image, attributes) = row.into_parts();
				let texture = upload_to_texture(executor, &queue, &image);

				Item::from_parts(Raster::new_gpu(texture), attributes)
			})
			.collect();

		queue.submit([]);
		list
	}
}

/// Converts single CPU raster to GPU by uploading to texture
impl<'i> Convert<Raster<GPU>, &'i WgpuExecutor> for Raster<CPU> {
	async fn convert(self, _: Footprint, executor: &'i WgpuExecutor) -> Raster<GPU> {
		let queue = executor.context().queue.lock();
		let texture = upload_to_texture(executor, &queue, &self);

		queue.submit([]);
		Raster::new_gpu(texture)
	}
}

/// Passthrough conversion for CPU `List`s - no conversion needed
impl<'i> Convert<List<Raster<CPU>>, &'i WgpuExecutor> for List<Raster<CPU>> {
	async fn convert(self, _: Footprint, _converter: &'i WgpuExecutor) -> List<Raster<CPU>> {
		self
	}
}

/// Converts a `List<Raster<GPU>>` to `List<Raster<CPU>>` by downloading texture data in one go then asynchronously maps all buffers and processes the results.
impl<'i> Convert<List<Raster<CPU>>, &'i WgpuExecutor> for List<Raster<GPU>> {
	async fn convert(self, _: Footprint, executor: &'i WgpuExecutor) -> List<Raster<CPU>> {
		let device = &executor.context().device;
		let queue = &executor.context().queue;

		let mut encoder = device.create_command_encoder(&wgpu::CommandEncoderDescriptor {
			label: Some("batch_texture_download_encoder"),
		});

		let mut converters = Vec::new();
		let mut rows_meta = Vec::new();

		for row in self {
			let (element, attributes) = row.into_parts();
			converters.push(RasterGpuToRasterCpuConverter::new(executor, &mut encoder, element));
			rows_meta.push(Item::from_parts((), attributes));
		}

		queue.submit([encoder.finish()]);

		let mut map_futures = Vec::new();
		for converter in converters {
			map_futures.push(converter.convert(device));
		}

		let map_results = futures::future::try_join_all(map_futures)
			.await
			.map_err(|_| "Failed to receive map result")
			.expect("Buffer mapping communication failed");

		map_results
			.into_iter()
			.zip(rows_meta)
			.map(|(element, row)| {
				let (_, attributes) = row.into_parts();
				Item::from_parts(element, attributes)
			})
			.collect()
	}
}

/// Converts single GPU raster to CPU by downloading texture data
impl<'i> Convert<Raster<CPU>, &'i WgpuExecutor> for Raster<GPU> {
	async fn convert(self, _: Footprint, executor: &'i WgpuExecutor) -> Raster<CPU> {
		let device = &executor.context().device;
		let queue = &executor.context().queue;

		let mut encoder = device.create_command_encoder(&wgpu::CommandEncoderDescriptor {
			label: Some("single_texture_download_encoder"),
		});

		let converter = RasterGpuToRasterCpuConverter::new(executor, &mut encoder, self);

		queue.submit([encoder.finish()]);

		converter.convert(device).await.expect("Failed to download texture data")
	}
}