Merge pull request #1400 from herbie-fp/cache-accelerators

Add a cache in front of every Rival accelerator

Author: pavpanchekha

src/syntax/generators.rkt

@@ -5,6 +5,7 @@
          ffi/unsafe)
 
 (require "../core/rival.rkt"
+         "../config.rkt"
          "types.rkt")
 
 (provide from-rival
@@ -24,14 +25,26 @@
 
 ; ----------------------- RIVAL GENERATOR ---------------------------
 
-(define-generator ((from-rival) spec ctx)
+(define/reset caches
+              '()
+              (lambda ()
+                (for ([cache (caches)])
+                  (hash-clear! cache))))
+
+(define-generator ((from-rival #:cache? [cache? #t]) spec ctx)
   (define compiler (make-real-compiler (list spec) (list ctx)))
   (define fail ((representation-bf->repr (context-repr ctx)) +nan.bf))
-  (lambda pt
+  (define (compute . pt)
     (define-values (_ exs) (real-apply compiler (list->vector pt)))
     (if exs
         (first exs)
-        fail)))
+        fail))
+  (cond
+    [cache?
+     (define cache (make-hash))
+     (caches (cons cache (caches)))
+     (lambda pt (hash-ref! cache pt (lambda () (apply compute pt))))]
+    [else compute]))
 
 ; ----------------------- FFI GENERATOR -----------------------------
 

www/doc/2.3/platforms.html

@@ -409,7 +409,10 @@ <h3>Hard-to-emulate Operations</h3>
   operation's specification. Compilation is usually <em>much</em>
   slower than with a native floating-point implementation, but for
   unusual operations that are difficult to implement otherwise, it can
-  still allow compilation to proceed.</p>
+  still allow compilation to proceed. There is an
+  optional <code>#:cache?</code> argument to <code>from-rival</code>,
+  on by default. The cache makes Herbie much faster but uses a lot of
+  memory.</p>
 
   <p>Note that <code>from-rival</code> implementations are always
   "correctly-rounded", meaning as accurate as possible. Most targets