Merge pull request #374 from linebender/cpu_shader

Add CPU shaders

crates/encoding/src/path.rs

@@ -227,8 +227,11 @@ pub struct Path {
 pub struct Tile {
     /// Accumulated backdrop at the left edge of the tile.
     pub backdrop: i32,
-    /// Index of first path segment.
-    pub segments: u32,
+    /// An enum that holds either the count of the number of path
+    /// segments in this tile, or an index to the beginning of an
+    /// allocated slice of `PathSegment` objects. In the latter case,
+    /// the bits are inverted.
+    pub segment_count_or_ix: u32,
 }
 
 /// Encoder for path segments.

shader/draw_leaf.wgsl

@@ -108,7 +108,6 @@ fn main(
         // let x1 = f32(bbox.x1);
         // let y1 = f32(bbox.y1);
         // let bbox_f = vec4(x0, y0, x1, y1);
-        let fill_mode = u32(bbox.linewidth >= 0.0);
         var transform = Transform();
         var linewidth = bbox.linewidth;
         if linewidth >= 0.0 || tag_word == DRAWTAG_FILL_LIN_GRADIENT || tag_word == DRAWTAG_FILL_RAD_GRADIENT ||

src/cpu_dispatch.rs

@@ -4,10 +4,12 @@
 //! Support for CPU implementations of compute shaders.
 
 use std::{
-    cell::{RefCell, RefMut},
-    ops::Deref,
+    cell::{Ref, RefCell, RefMut},
+    ops::{Deref, DerefMut},
 };
 
+use bytemuck::Pod;
+
 #[derive(Clone, Copy)]
 pub enum CpuBinding<'a> {
     Buffer(&'a [u8]),
@@ -16,39 +18,88 @@ pub enum CpuBinding<'a> {
     Texture(&'a CpuTexture),
 }
 
-pub enum CpuBufGuard<'a> {
-    Slice(&'a [u8]),
-    Interior(RefMut<'a, Vec<u8>>),
+pub enum TypedBufGuard<'a, T: ?Sized> {
+    Slice(&'a T),
+    Interior(Ref<'a, T>),
+}
+
+pub enum TypedBufGuardMut<'a, T: ?Sized> {
+    Slice(&'a mut T),
+    Interior(RefMut<'a, T>),
 }
 
-impl<'a> Deref for CpuBufGuard<'a> {
-    type Target = [u8];
+impl<'a, T: ?Sized> Deref for TypedBufGuard<'a, T> {
+    type Target = T;
 
     fn deref(&self) -> &Self::Target {
         match self {
-            CpuBufGuard::Slice(s) => s,
-            CpuBufGuard::Interior(r) => r,
+            TypedBufGuard::Slice(s) => s,
+            TypedBufGuard::Interior(r) => r,
         }
     }
 }
 
-impl<'a> CpuBufGuard<'a> {
-    /// Get a mutable reference to the buffer.
-    ///
-    /// Panics if the underlying resource is read-only.
-    pub fn as_mut(&mut self) -> &mut [u8] {
+impl<'a, T: ?Sized> Deref for TypedBufGuardMut<'a, T> {
+    type Target = T;
+
+    fn deref(&self) -> &Self::Target {
         match self {
-            CpuBufGuard::Interior(r) => &mut *r,
-            _ => panic!("tried to borrow immutable buffer as mutable"),
+            TypedBufGuardMut::Slice(s) => s,
+            TypedBufGuardMut::Interior(r) => r,
+        }
+    }
+}
+
+impl<'a, T: ?Sized> DerefMut for TypedBufGuardMut<'a, T> {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        match self {
+            TypedBufGuardMut::Slice(s) => s,
+            TypedBufGuardMut::Interior(r) => r,
         }
     }
 }
 
 impl<'a> CpuBinding<'a> {
-    pub fn as_buf(&self) -> CpuBufGuard {
+    pub fn as_typed<T: Pod>(&self) -> TypedBufGuard<T> {
+        match self {
+            CpuBinding::Buffer(b) => TypedBufGuard::Slice(bytemuck::from_bytes(b)),
+            CpuBinding::BufferRW(b) => {
+                TypedBufGuard::Interior(Ref::map(b.borrow(), |buf| bytemuck::from_bytes(buf)))
+            }
+            _ => panic!("resource type mismatch"),
+        }
+    }
+
+    pub fn as_typed_mut<T: Pod>(&self) -> TypedBufGuardMut<T> {
+        match self {
+            CpuBinding::Buffer(_) => panic!("can't borrow external buffer mutably"),
+            CpuBinding::BufferRW(b) => {
+                TypedBufGuardMut::Interior(RefMut::map(b.borrow_mut(), |buf| {
+                    bytemuck::from_bytes_mut(buf)
+                }))
+            }
+            _ => panic!("resource type mismatch"),
+        }
+    }
+
+    pub fn as_slice<T: Pod>(&self) -> TypedBufGuard<[T]> {
+        match self {
+            CpuBinding::Buffer(b) => TypedBufGuard::Slice(bytemuck::cast_slice(b)),
+            CpuBinding::BufferRW(b) => {
+                TypedBufGuard::Interior(Ref::map(b.borrow(), |buf| bytemuck::cast_slice(buf)))
+            }
+            _ => panic!("resource type mismatch"),
+        }
+    }
+
+    pub fn as_slice_mut<T: Pod>(&self) -> TypedBufGuardMut<[T]> {
         match self {
-            CpuBinding::Buffer(b) => CpuBufGuard::Slice(b),
-            CpuBinding::BufferRW(b) => CpuBufGuard::Interior(b.borrow_mut()),
+            CpuBinding::Buffer(_) => panic!("can't borrow external buffer mutably"),
+            CpuBinding::BufferRW(b) => {
+                TypedBufGuardMut::Interior(RefMut::map(b.borrow_mut(), |buf| {
+                    bytemuck::cast_slice_mut(buf)
+                }))
+            }
             _ => panic!("resource type mismatch"),
         }
     }

src/cpu_shader/backdrop.rs

@@ -0,0 +1,30 @@
+// Copyright 2023 The Vello authors
+// SPDX-License-Identifier: Apache-2.0 OR MIT OR Unlicense
+
+use vello_encoding::{ConfigUniform, Path, Tile};
+
+use crate::cpu_dispatch::CpuBinding;
+
+fn backdrop_main(config: &ConfigUniform, paths: &[Path], tiles: &mut [Tile]) {
+    for drawobj_ix in 0..config.layout.n_draw_objects {
+        let path = paths[drawobj_ix as usize];
+        let width = path.bbox[2] - path.bbox[0];
+        let height = path.bbox[3] - path.bbox[1];
+        let base = path.tiles;
+        for y in 0..height {
+            let mut sum = 0;
+            for x in 0..width {
+                let tile = &mut tiles[(base + y * width + x) as usize];
+                sum += tile.backdrop;
+                tile.backdrop = sum;
+            }
+        }
+    }
+}
+
+pub fn backdrop(_n_wg: u32, resources: &[CpuBinding]) {
+    let config = resources[0].as_typed();
+    let paths = resources[1].as_slice();
+    let mut tiles = resources[2].as_slice_mut();
+    backdrop_main(&config, &paths, &mut tiles);
+}

src/cpu_shader/bbox_clear.rs

@@ -0,0 +1,21 @@
+// Copyright 2023 The Vello authors
+// SPDX-License-Identifier: Apache-2.0 OR MIT OR Unlicense
+
+use vello_encoding::{ConfigUniform, PathBbox};
+
+use crate::cpu_dispatch::CpuBinding;
+
+fn bbox_clear_main(config: &ConfigUniform, path_bboxes: &mut [PathBbox]) {
+    for i in 0..(config.layout.n_paths as usize) {
+        path_bboxes[i].x0 = 0x7fff_ffff;
+        path_bboxes[i].y0 = 0x7fff_ffff;
+        path_bboxes[i].x1 = -0x8000_0000;
+        path_bboxes[i].y1 = -0x8000_0000;
+    }
+}
+
+pub fn bbox_clear(_n_wg: u32, resources: &[CpuBinding]) {
+    let config = resources[0].as_typed();
+    let mut path_bboxes = resources[1].as_slice_mut();
+    bbox_clear_main(&config, &mut path_bboxes);
+}

src/cpu_shader/binning.rs

@@ -0,0 +1,128 @@
+// Copyright 2023 The Vello authors
+// SPDX-License-Identifier: Apache-2.0 OR MIT OR Unlicense
+
+use vello_encoding::{BinHeader, BumpAllocators, ConfigUniform, DrawMonoid, PathBbox};
+
+use crate::cpu_dispatch::CpuBinding;
+
+const WG_SIZE: usize = 256;
+const TILE_WIDTH: usize = 16;
+const TILE_HEIGHT: usize = 16;
+const N_TILE_X: usize = 16;
+const N_TILE_Y: usize = 16;
+const SX: f32 = 1.0 / ((N_TILE_X * TILE_WIDTH) as f32);
+const SY: f32 = 1.0 / ((N_TILE_Y * TILE_HEIGHT) as f32);
+
+fn bbox_intersect(a: [f32; 4], b: [f32; 4]) -> [f32; 4] {
+    [
+        a[0].max(b[0]),
+        a[1].max(b[1]),
+        a[2].min(b[2]),
+        a[3].min(b[3]),
+    ]
+}
+
+fn binning_main(
+    n_wg: u32,
+    config: &ConfigUniform,
+    draw_monoids: &[DrawMonoid],
+    path_bbox_buf: &[PathBbox],
+    clip_bbox_buf: &[[f32; 4]],
+    intersected_bbox: &mut [[f32; 4]],
+    bump: &mut BumpAllocators,
+    bin_data: &mut [u32],
+    bin_header: &mut [BinHeader],
+) {
+    for wg in 0..n_wg as usize {
+        let mut counts = [0; WG_SIZE];
+        let mut bboxes = [[0, 0, 0, 0]; WG_SIZE];
+        let width_in_bins =
+            ((config.width_in_tiles + N_TILE_X as u32 - 1) / N_TILE_X as u32) as i32;
+        let height_in_bins =
+            ((config.height_in_tiles + N_TILE_Y as u32 - 1) / N_TILE_Y as u32) as i32;
+        for local_ix in 0..WG_SIZE {
+            let element_ix = wg * WG_SIZE + local_ix;
+            let mut x0 = 0;
+            let mut y0 = 0;
+            let mut x1 = 0;
+            let mut y1 = 0;
+            if element_ix < config.layout.n_draw_objects as usize {
+                let draw_monoid = draw_monoids[element_ix];
+                let mut clip_bbox = [-1e9, -1e9, 1e9, 1e9];
+                if draw_monoid.clip_ix > 0 {
+                    assert!(draw_monoid.clip_ix - 1 < config.layout.n_clips);
+                    clip_bbox = clip_bbox_buf[draw_monoid.clip_ix as usize - 1];
+                }
+                let path_bbox = path_bbox_buf[draw_monoid.path_ix as usize];
+                let pb = [
+                    path_bbox.x0 as f32,
+                    path_bbox.y0 as f32,
+                    path_bbox.x1 as f32,
+                    path_bbox.y1 as f32,
+                ];
+                let bbox = bbox_intersect(clip_bbox, pb);
+                intersected_bbox[element_ix] = bbox;
+                if bbox[0] < bbox[2] && bbox[1] < bbox[3] {
+                    x0 = (bbox[0] * SX).floor() as i32;
+                    y0 = (bbox[1] * SY).floor() as i32;
+                    x1 = (bbox[2] * SX).ceil() as i32;
+                    y1 = (bbox[3] * SY).ceil() as i32;
+                }
+            }
+            x0 = x0.clamp(0, width_in_bins);
+            y0 = y0.clamp(0, height_in_bins);
+            x1 = x1.clamp(0, width_in_bins);
+            y1 = y1.clamp(0, height_in_bins);
+            for y in y0..y1 {
+                for x in x0..x1 {
+                    counts[(y * width_in_bins + x) as usize] += 1;
+                }
+            }
+            bboxes[local_ix] = [x0, y0, x1, y1];
+        }
+        let mut chunk_offset = [0; WG_SIZE];
+        for local_ix in 0..WG_SIZE {
+            let global_ix = wg * WG_SIZE + local_ix;
+            chunk_offset[local_ix] = bump.binning;
+            bump.binning += counts[local_ix];
+            bin_header[global_ix] = BinHeader {
+                element_count: counts[local_ix],
+                chunk_offset: chunk_offset[local_ix],
+            };
+        }
+        for local_ix in 0..WG_SIZE {
+            let element_ix = wg * WG_SIZE + local_ix;
+            let bbox = bboxes[local_ix];
+            for y in bbox[1]..bbox[3] {
+                for x in bbox[0]..bbox[2] {
+                    let bin_ix = (y * width_in_bins + x) as usize;
+                    let ix = config.layout.bin_data_start + chunk_offset[bin_ix];
+                    bin_data[ix as usize] = element_ix as u32;
+                    chunk_offset[bin_ix] += 1;
+                }
+            }
+        }
+    }
+}
+
+pub fn binning(n_wg: u32, resources: &[CpuBinding]) {
+    let config = resources[0].as_typed();
+    let draw_monoids = resources[1].as_slice();
+    let path_bbox_buf = resources[2].as_slice();
+    let clip_bbox_buf = resources[3].as_slice();
+    let mut intersected_bbox = resources[4].as_slice_mut();
+    let mut bump = resources[5].as_typed_mut();
+    let mut bin_data = resources[6].as_slice_mut();
+    let mut bin_header = resources[7].as_slice_mut();
+    binning_main(
+        n_wg,
+        &config,
+        &draw_monoids,
+        &path_bbox_buf,
+        &clip_bbox_buf,
+        &mut intersected_bbox,
+        &mut bump,
+        &mut bin_data,
+        &mut bin_header,
+    );
+}