Create single Rayon threadpool and use for YIQ ops

crates/ntscrs/src/lib.rs

@@ -4,6 +4,7 @@ pub mod ntsc;
 mod random;
 pub mod settings;
 mod shift;
+pub mod thread_pool;
 pub mod yiq_fielding;
 
 #[macro_use]

crates/ntscrs/src/ntsc.rs

@@ -12,6 +12,7 @@ use crate::{
     filter::TransferFunction,
     random::{Geometric, Seeder},
     shift::{shift_row, shift_row_to, BoundaryHandling},
+    thread_pool::with_thread_pool,
     yiq_fielding::{BlitInfo, PixelFormat, YiqField, YiqOwned, YiqView},
 };
 
@@ -1368,33 +1369,7 @@ impl NtscEffect {
 
     /// Apply the effect to YIQ image data.
     pub fn apply_effect_to_yiq(&self, yiq: &mut YiqView, frame_num: usize) {
-        #[cfg(not(target_arch = "wasm32"))]
-        {
-            use std::sync::OnceLock;
-            static NUM_THREADS: OnceLock<Option<usize>> = OnceLock::new();
-            // On Windows debug builds, the stack overflows with the default stack size
-            let mut pool = rayon::ThreadPoolBuilder::new().stack_size(2 * 1024 * 1024);
-
-            // Use physical core count instead of logical core count. Hyperthreading seems to be ~20-25% slower, at least on
-            // a Ryzen 7 7700X.
-            if let Some(num_threads) = *NUM_THREADS.get_or_init(|| {
-                if std::env::var("RAYON_NUM_THREADS").is_ok() {
-                    return None;
-                }
-                Some(num_cpus::get_physical())
-            }) {
-                pool = pool.num_threads(num_threads);
-            }
-
-            let pool = pool.build().unwrap();
-            pool.scope(|_| self.apply_effect_to_all_fields(yiq, frame_num));
-        }
-        #[cfg(target_arch = "wasm32")]
-        {
-            // wasm-bindgen-rayon doesn't support custom thread pools
-            // https://github.com/RReverser/wasm-bindgen-rayon/issues/18
-            self.apply_effect_to_all_fields(yiq, frame_num);
-        }
+        with_thread_pool(|| self.apply_effect_to_all_fields(yiq, frame_num));
     }
 
     /// Apply the effect to a buffer which contains pixels in the given format.

crates/ntscrs/src/thread_pool.rs

@@ -0,0 +1,28 @@
+pub fn with_thread_pool<T: Send>(op: impl (FnOnce() -> T) + Send + Sync) -> T {
+    #[cfg(not(target_arch = "wasm32"))]
+    {
+        use std::sync::OnceLock;
+        static POOL: OnceLock<rayon::ThreadPool> = OnceLock::new();
+
+        let pool = POOL.get_or_init(|| {
+            // On Windows debug builds, the stack overflows with the default stack size
+            let mut pool = rayon::ThreadPoolBuilder::new().stack_size(2 * 1024 * 1024);
+
+            // Use physical core count instead of logical core count. Hyperthreading seems to be ~20-25% slower, at least on
+            // a Ryzen 7 7700X.
+            if std::env::var("RAYON_NUM_THREADS").is_err() {
+                pool = pool.num_threads(num_cpus::get_physical());
+            }
+
+            pool.build().unwrap()
+        });
+
+        pool.scope(|_| op())
+    }
+    #[cfg(target_arch = "wasm32")]
+    {
+        // wasm-bindgen-rayon doesn't support custom thread pools
+        // https://github.com/RReverser/wasm-bindgen-rayon/issues/18
+        op()
+    }
+}