Hash a model's bytes once, at open, not on every acquire

A border fill acquires the filler once a tile, and each acquire hashed
the 28 MB model twice — 60 ms a tile, a third of the tile's run on a
throttled TensorRT. The Model keeps its hash from open.

core/dr-inference-engine/src/engines.rs

@@ -29,7 +29,12 @@ pub fn hash(bytes: &[u8]) -> u64 {
 
 /// The cache entry for `bytes` compiled on `rung`.
 pub fn key(rung: Rung, bytes: &[u8]) -> String {
-    format!("{}:{:016x}", rung.label(), hash(bytes))
+    key_of(rung, hash(bytes))
+}
+
+/// The same, from a hash already taken.
+pub fn key_of(rung: Rung, hash: u64) -> String {
+    format!("{}:{:016x}", rung.label(), hash)
 }
 
 /// Where QNN's compiled context for `bytes` lives.

core/dr-inference-engine/src/lib.rs

@@ -185,6 +185,9 @@ pub struct Model {
     role: Role,
     form: Form,
     bytes: Arc<[u8]>,
+    /// `engines::hash` of the bytes, taken once: an acquire per tile of a
+    /// border fill must not hash 28 MB each time.
+    hash: u64,
 }
 
 /// A loaded session, held for one `run` and its output decoding.
@@ -203,7 +206,7 @@ impl Model {
     /// one run; a scan and a develop click can want the same detector at
     /// once, and the second waits on the first.
     pub fn acquire(&self) -> Result<Acquired, Error> {
-        acquire(self.role, self.form, &self.bytes)
+        acquire(self.role, self.form, &self.bytes, self.hash)
     }
 
     pub fn form(&self) -> Form {
@@ -247,17 +250,17 @@ fn registry() -> &'static Mutex<Registry> {
     })
 }
 
-fn acquire(role: Role, form: Form, bytes: &Arc<[u8]>) -> Result<Acquired, Error> {
+fn acquire(role: Role, form: Form, bytes: &Arc<[u8]>, hash: u64) -> Result<Acquired, Error> {
     api::ensure_installed();
     let (rung, cfg) = {
         let s = state().lock().unwrap();
         let selected = current_rung(&s);
         (
-            effective_rung(&s, selected, role, form, bytes),
+            effective_rung(&s, selected, role, form, hash),
             s.config.clone(),
         )
     };
-    let key = format!("{role:?}:{}", engines::key(rung, bytes));
+    let key = format!("{role:?}:{}", engines::key_of(rung, hash));
 
     if let Some(entry) = registry().lock().unwrap().get(&key).cloned() {
         *entry.last_used.lock().unwrap() = Instant::now();
@@ -456,20 +459,26 @@ pub fn form_suffix(form: Form) -> &'static str {
 /// this crate existed.
 pub fn open(role: Role, form: Form, bytes: &[u8]) -> Result<Model, Error> {
     let bytes: Arc<[u8]> = Arc::from(bytes);
-    acquire(role, form, &bytes)?;
+    let hash = engines::hash(&bytes);
-    Ok(Model { role, form, bytes })
+    acquire(role, form, &bytes, hash)?;
+    Ok(Model {
+        role,
+        form,
+        bytes,
+        hash,
+    })
 }
 
 /// Where a request lands: the selected rung unless the role's precision rule,
 /// the form on offer, or a missing engine says one lower (§6 step 4).
-fn effective_rung(s: &State, selected: Rung, role: Role, form: Form, bytes: &[u8]) -> Rung {
+fn effective_rung(s: &State, selected: Rung, role: Role, form: Form, hash: u64) -> Rung {
     let mut rung = selected;
     if !rung.serves(role) || rung.form(role) != form {
         // The embedder on a Hexagon device, or an f32 detector where the int8
         // sibling was missing: neither can go to the NPU.
         rung = rung.fallback();
     }
-    if rung.compiles() && !s.cache.compiled.contains(&engines::key(rung, bytes)) {
+    if rung.compiles() && !s.cache.compiled.contains(&engines::key_of(rung, hash)) {
         rung = rung.fallback();
     }
     rung
@@ -570,16 +579,34 @@ mod tests {
             wanted: 0,
         };
         assert_eq!(
-            effective_rung(&s, Rung::Hexagon, Role::Embedder, Form::F32, b""),
+            effective_rung(
+                &s,
+                Rung::Hexagon,
+                Role::Embedder,
+                Form::F32,
+                engines::hash(b"")
+            ),
             Rung::Cpu
         );
         assert_eq!(
-            effective_rung(&s, Rung::Hexagon, Role::Detector, Form::F32, b""),
+            effective_rung(
+                &s,
+                Rung::Hexagon,
+                Role::Detector,
+                Form::F32,
+                engines::hash(b"")
+            ),
             Rung::Cpu
         );
         // An int8 detector whose context is not compiled yet: also the CPU.
         assert_eq!(
-            effective_rung(&s, Rung::Hexagon, Role::Detector, Form::Int8, b""),
+            effective_rung(
+                &s,
+                Rung::Hexagon,
+                Role::Detector,
+                Form::Int8,
+                engines::hash(b"")
+            ),
             Rung::Cpu
         );
     }