Merge pull request #9200 from penpot/azazeln28-build-time-performance-optimizations

🔧 Change build and cargo settings

frontend/src/debug.cljs

@@ -135,6 +135,14 @@
       (wasm.mem/free)
       text)))
 
+(defn ^:export wasmRenderStats
+  []
+  (let [module wasm/internal-module
+        f      (when module (unchecked-get module "_render_stats"))]
+    (if (fn? f)
+      (wasm.h/call module "_render_stats")
+      (js/console.warn "[debug] render-wasm module not ready or missing _render_stats"))))
+
 (defn ^:export wasmAtlasConsole
   "Logs the current render-wasm atlas as an image in the JS console (if present)."
   []

render-wasm/Cargo.toml

@@ -10,6 +10,7 @@ build = "build.rs"
 
 [features]
 default = []
+stats = []
 profile = ["profile-macros", "profile-raf"]
 profile-macros = []
 profile-raf = []
@@ -36,4 +37,6 @@ thiserror = "2.0.18"
 uuid = { version = "1.11.0", features = ["v4", "js"] }
 
 [profile.release]
-opt-level = "s"
+opt-level = 3
+lto = "fat"
+strip = true

render-wasm/_build_env

@@ -31,8 +31,8 @@ export EM_MALLOC="dlmalloc"
 
 export EMCC_CFLAGS="--no-entry \
     --js-library src/js/wapi.js \
-    -sASSERTIONS=1 \
+    -sMALLOC=$EM_MALLOC \
-    -sALLOW_TABLE_GROWTH=1 \
+    -sALLOW_TABLE_GROWTH=0 \
     -sALLOW_MEMORY_GROWTH=1 \
     -sINITIAL_HEAP=$EM_INITIAL_HEAP \
     -sMEMORY_GROWTH_GEOMETRIC_STEP=$EM_MEMORY_GROWTH_GEOMETRIC_STEP \
@@ -50,13 +50,13 @@ export CARGO_PARAMS="${@:2}";
 
 if [ "$BUILD_MODE" = "release" ]; then
     export CARGO_PARAMS="--release $CARGO_PARAMS"
-    export EMCC_CFLAGS="-Os $EMCC_CFLAGS"
+    export EMCC_CFLAGS="-O3 -sASSERTIONS=0 $EMCC_CFLAGS"
 else
     # TODO: Extra parameters that could be good to look into:
     #        -gseparate-dwarf
     #        -gsplit-dwarf
     #        -gsource-map
-    export EMCC_CFLAGS="-g $EMCC_CFLAGS -sVERBOSE=1 -sMALLOC=$EM_MALLOC"
+    export EMCC_CFLAGS="-g -sASSERTIONS=1 -sVERBOSE=1 $EMCC_CFLAGS"
 fi
 
 function clean {

render-wasm/src/main.rs

@@ -1052,6 +1052,13 @@ pub extern "C" fn render_shape_pixels(
     })
 }
 
+#[no_mangle]
+pub extern "C" fn render_stats() {
+    with_state!(state, {
+        state.render_state.print_stats();
+    })
+}
+
 fn main() {
     #[cfg(target_arch = "wasm32")]
     init_gl!();

render-wasm/src/render.rs

@@ -15,6 +15,7 @@ mod ui;
 
 use skia_safe::{self as skia, Matrix, RRect, Rect};
 use std::borrow::Cow;
+use std::collections::HashMap;
 use std::collections::HashSet;
 
 use gpu_state::GpuState;
@@ -289,9 +290,47 @@ fn sort_z_index(tree: ShapesPoolRef, element: &Shape, children_ids: Vec<Uuid>) -
     }
 }
 
+struct RenderStats {
+    pub counts: HashMap<Uuid, i32>,
+}
+
+#[allow(dead_code)]
+impl RenderStats {
+    pub fn new() -> Self {
+        Self {
+            counts: HashMap::new(),
+        }
+    }
+
+    fn count(&mut self, id: Uuid) -> i32 {
+        let counter = self.counts.entry(id).or_insert(0);
+        *counter += 1;
+        *counter
+    }
+
+    fn clear(&mut self) {
+        self.counts.clear();
+    }
+
+    #[allow(dead_code)]
+    fn get(&self, id: &Uuid) -> Option<&i32> {
+        self.counts.get(id)
+    }
+
+    fn print(&self) {
+        let mut sum: i32 = 0;
+        for (&id, &count) in self.counts.iter() {
+            println!("{}: {}", id, count);
+            sum += count;
+        }
+        println!("{}: {}", self.counts.len(), sum);
+    }
+}
+
 pub(crate) struct RenderState {
     gpu_state: GpuState,
     pub options: RenderOptions,
+    stats: RenderStats,
     pub surfaces: Surfaces,
     pub fonts: FontStore,
     pub viewbox: Viewbox,
@@ -388,6 +427,7 @@ impl RenderState {
         Ok(RenderState {
             gpu_state: gpu_state.clone(),
             options,
+            stats: RenderStats::new(),
             surfaces,
             fonts,
             viewbox,
@@ -876,6 +916,9 @@ impl RenderState {
         outset: Option<f32>,
         target_surface: SurfaceId,
     ) -> Result<()> {
+        #[cfg(feature = "stats")]
+        self.stats.count(shape.id);
+
         let surface_ids = fills_surface_id as u32
             | strokes_surface_id as u32
             | innershadows_surface_id as u32
@@ -1690,6 +1733,9 @@ impl RenderState {
         timestamp: i32,
         sync_render: bool,
     ) -> Result<()> {
+        #[cfg(feature = "stats")]
+        self.stats.clear();
+
         let _start = performance::begin_timed_log!("start_render_loop");
         let scale = self.get_scale();
 
@@ -3443,6 +3489,10 @@ impl RenderState {
     pub fn set_view(&mut self, zoom: f32, x: f32, y: f32) {
         self.viewbox.set_all(zoom, x, y);
     }
+
+    pub fn print_stats(&self) {
+        self.stats.print();
+    }
 }
 
 impl Drop for RenderState {

render-wasm/src/render/gpu_state.rs

@@ -1,5 +1,8 @@
 use crate::error::{Error, Result};
-use skia_safe::gpu::{self, gl::FramebufferInfo, gl::TextureInfo, DirectContext};
+use skia_safe::gpu::{
+    self, ganesh::context_options::Enable, gl::FramebufferInfo, gl::TextureInfo, ContextOptions,
+    DirectContext,
+};
 use skia_safe::{self as skia, ISize};
 
 #[derive(Debug, Clone)]
@@ -13,7 +16,16 @@ impl GpuState {
         let interface = gpu::gl::Interface::new_native().ok_or(Error::CriticalError(
             "Failed to create GL interface".to_string(),
         ))?;
-        let context = gpu::direct_contexts::make_gl(interface, None).ok_or(
+
+        // We tweak some options to enhance performance.
+        let mut context_options = ContextOptions::default();
+        // context_options.reduce_ops_task_splitting = Enable::Yes;
+        context_options.skip_gl_error_checks = Enable::Yes;
+        // context_options.runtime_program_cache_size = 1024;
+        // context_options.allow_multiple_glyph_cache_textures = Enable::Yes;
+        // context_options.allow_path_mask_caching = false;
+
+        let context = gpu::direct_contexts::make_gl(interface, Some(&context_options)).ok_or(
             Error::CriticalError("Failed to create GL context".to_string()),
         )?;
         let framebuffer_info = {

render-wasm/src/render/strokes.rs

@@ -231,8 +231,8 @@ fn draw_stroke_on_path(
     if let Some(pt) = path_transform {
         canvas.concat(pt);
     }
-    let skia_path = path.to_skia_path(svg_attrs);
 
+    let skia_path = path.to_skia_path(svg_attrs);
     match stroke.render_kind(is_open) {
         StrokeKind::Inner => {
             draw_inner_stroke_path(canvas, &skia_path, &draw_paint, blur, antialias);

render-wasm/src/shapes/paths.rs

@@ -1,7 +1,6 @@
-use skia_safe::{self as skia, Matrix};
-
 use crate::math;
 use crate::shapes::svg_attrs::{FillRule, SvgAttrs};
+use skia_safe::{self as skia, Matrix};
 
 mod subpaths;