LUT caching registry for CORDIC tables

crates/cord-cordic/src/compiler.rs

@@ -1,4 +1,5 @@
 use cord_trig::{TrigGraph, TrigOp};
+use crate::lut::CordicTable;
 use crate::ops::*;
 
 /// A compiled CORDIC program ready for binary serialization or execution.
@@ -77,12 +78,14 @@ impl CORDICProgram {
             }
         }).collect();
 
+        let table = CordicTable::for_word_bits(config.word_bits);
+
         CORDICProgram {
             word_bits: config.word_bits,
             instructions,
             output: graph.output,
-            atan_table: atan_table(config.word_bits),
-            gain: cordic_gain(config.word_bits, frac_bits),
+            atan_table: table.atan.to_vec(),
+            gain: table.gain,
         }
     }
 

crates/cord-cordic/src/eval.rs

@@ -1,38 +1,28 @@
+use std::sync::Arc;
 use cord_trig::{TrigGraph, TrigOp};
+use crate::lut::CordicTable;
 
 /// CORDIC evaluator: evaluates a TrigGraph using only integer
 /// shifts, adds, and comparisons. No floating point trig.
 ///
 /// Proof that the entire pipeline compiles down to
-/// binary arithmetic — shift, add, compare, repeat.
+/// binary arithmetic -- shift, add, compare, repeat.
 pub struct CORDICEvaluator {
     word_bits: u8,
     frac_bits: u8,
-    atan_table: Vec<i64>,
+    atan_table: Arc<[i64]>,
     gain: i64,
 }
 
 impl CORDICEvaluator {
     pub fn new(word_bits: u8) -> Self {
-        let frac_bits = word_bits - 1;
-        let iterations = word_bits;
-
-        // Precompute atan(2^-i) as fixed-point
-        let atan_table: Vec<i64> = (0..iterations)
-            .map(|i| {
-                let angle = (2.0f64).powi(-(i as i32)).atan();
-                (angle * (1i64 << frac_bits) as f64).round() as i64
-            })
-            .collect();
-
-        // CORDIC gain K = product of 1/sqrt(1 + 2^{-2i})
-        let mut k = 1.0f64;
-        for i in 0..iterations {
-            k *= 1.0 / (1.0 + (2.0f64).powi(-2 * i as i32)).sqrt();
+        let table = CordicTable::for_word_bits(word_bits);
+        CORDICEvaluator {
+            word_bits,
+            frac_bits: table.frac_bits,
+            atan_table: table.atan,
+            gain: table.gain,
         }
-        let gain = (k * (1i64 << frac_bits) as f64).round() as i64;
-
-        CORDICEvaluator { word_bits, frac_bits, atan_table, gain }
     }
 
     /// Convert f64 to fixed-point.

crates/cord-cordic/src/lib.rs

@@ -10,6 +10,8 @@
 pub mod compiler;
 pub mod ops;
 pub mod eval;
+pub mod lut;
 
 pub use compiler::CORDICProgram;
 pub use eval::CORDICEvaluator;
+pub use lut::CordicTable;

crates/cord-cordic/src/lut.rs

@@ -0,0 +1,114 @@
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex, OnceLock};
+
+/// Pre-computed CORDIC table: atan values + gain for a given iteration count.
+#[derive(Debug, Clone)]
+pub struct CordicTable {
+    pub atan: Arc<[i64]>,
+    pub gain: i64,
+    pub iterations: u8,
+    pub frac_bits: u8,
+}
+
+/// Global table registry. Keyed by (iterations, frac_bits) so any
+/// combination of word width and iteration count shares a single
+/// allocation. Thread-safe via interior Mutex on the map only;
+/// once an Arc is cloned out, no lock is held during evaluation.
+struct LutRegistry {
+    tables: Mutex<HashMap<(u8, u8), CordicTable>>,
+}
+
+static REGISTRY: OnceLock<LutRegistry> = OnceLock::new();
+
+fn registry() -> &'static LutRegistry {
+    REGISTRY.get_or_init(|| LutRegistry {
+        tables: Mutex::new(HashMap::new()),
+    })
+}
+
+impl CordicTable {
+    /// Retrieve or compute the CORDIC table for the given parameters.
+    /// Repeated calls with the same (iterations, frac_bits) return
+    /// a clone of the same Arc'd data -- no recomputation.
+    pub fn get(iterations: u8, frac_bits: u8) -> Self {
+        let reg = registry();
+        let key = (iterations, frac_bits);
+
+        let mut map = reg.tables.lock().unwrap();
+        if let Some(cached) = map.get(&key) {
+            return cached.clone();
+        }
+
+        let table = Self::compute(iterations, frac_bits);
+        map.insert(key, table.clone());
+        table
+    }
+
+    /// Convenience: derive iterations and frac_bits from word_bits
+    /// using the standard CORDIC convention (iterations = word_bits,
+    /// frac_bits = word_bits - 1).
+    pub fn for_word_bits(word_bits: u8) -> Self {
+        Self::get(word_bits, word_bits - 1)
+    }
+
+    fn compute(iterations: u8, frac_bits: u8) -> Self {
+        let scale = (1i64 << frac_bits) as f64;
+
+        let atan: Arc<[i64]> = (0..iterations)
+            .map(|i| {
+                let angle = (2.0f64).powi(-(i as i32)).atan();
+                (angle * scale).round() as i64
+            })
+            .collect::<Vec<_>>()
+            .into();
+
+        let mut k = 1.0f64;
+        for i in 0..iterations {
+            k *= 1.0 / (1.0 + (2.0f64).powi(-2 * i as i32)).sqrt();
+        }
+        let gain = (k * scale).round() as i64;
+
+        CordicTable { atan, gain, iterations, frac_bits }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::ops;
+
+    #[test]
+    fn cached_matches_raw() {
+        for bits in [8u8, 16, 32, 48, 64] {
+            let table = CordicTable::for_word_bits(bits);
+            let raw_atan = ops::atan_table(bits);
+            let raw_gain = ops::cordic_gain(bits, bits - 1);
+
+            assert_eq!(table.atan.len(), raw_atan.len(), "len mismatch at {bits}");
+            assert_eq!(&*table.atan, &raw_atan[..], "atan mismatch at {bits}");
+            assert_eq!(table.gain, raw_gain, "gain mismatch at {bits}");
+        }
+    }
+
+    #[test]
+    fn deduplication() {
+        let a = CordicTable::for_word_bits(32);
+        let b = CordicTable::for_word_bits(32);
+        assert!(Arc::ptr_eq(&a.atan, &b.atan));
+    }
+
+    #[test]
+    fn different_widths_distinct() {
+        let a = CordicTable::for_word_bits(16);
+        let b = CordicTable::for_word_bits(32);
+        assert!(!Arc::ptr_eq(&a.atan, &b.atan));
+        assert_ne!(a.atan.len(), b.atan.len());
+    }
+
+    #[test]
+    fn custom_iteration_count() {
+        let t = CordicTable::get(16, 31);
+        assert_eq!(t.atan.len(), 16);
+        assert_eq!(t.frac_bits, 31);
+    }
+}