Dual arrays is >100x faster than 'ad' on large fneuralHEADmaster

ad:Numeric.AD.Double / ad-dual:Numeric.ADDual.Array.Internal

Prelude> 1.129e-3 / 41.89e-6  -- neural-100
26.951539746956314
Prelude> 34.67e-3 / 156.9e-6  -- neural-180
220.9687699171447
Prelude> 79.03e-3 / 178.6e-6  -- neural-500
442.4972004479283
Prelude> 365.3e-3 / 665.5e-6  -- neural-2000
548.9105935386928

1 files changed, 14 insertions, 1 deletions

diff --git a/bench/Main.hs b/bench/Main.hs
index 1174a3a..cccc686 100644
--- a/bench/Main.hs
+++ b/bench/Main.hs
@@ -11,8 +11,10 @@ import System.Environment (getArgs)
 import System.Mem (performGC)
 
 import qualified Numeric.AD as AD
+import qualified Numeric.AD.Double as AD.Double
 
 import qualified Numeric.ADDual as ADD
+import qualified Numeric.ADDual.Array.Internal as ADDA
 import Numeric.ADDual.Examples
 
 
@@ -29,6 +31,10 @@ mainCriterion = defaultMain
   ,benchNeural 180  -- rather stably 2 GCs
   ,benchNeural 500
   ,benchNeural 2000
+  ,benchNeuralA 100
+  ,benchNeuralA 180  -- rather stably 2 GCs
+  ,benchNeuralA 500
+  ,benchNeuralA 2000
   ]
   where
     benchNeural :: Int -> Benchmark
@@ -36,7 +42,14 @@ mainCriterion = defaultMain
       env (pure (makeNeuralInput n)) $ \input ->
         bgroup ("neural-" ++ show n)
           [bench "dual" $ nf (\inp -> ADD.gradient' fneural inp 1.0) input
-          ,bench "ad" $ nf (\inp -> AD.grad fneural inp) input]
+          ,bench "ad" $ nf (\inp -> AD.grad fneural inp) input
+          ,bench "ad.Double" $ nf (\inp -> AD.Double.grad fneural inp) input]
+
+    benchNeuralA :: Int -> Benchmark
+    benchNeuralA n =
+      env (pure (makeNeuralInput_A n)) $ \input ->
+        bgroup ("neuralA-" ++ show n)
+          [bench "dual" $ nf (\inp -> ADDA.gradient' fneural_A inp 1.0) input]
 
 mainNeuralGraph :: IO ()
 mainNeuralGraph = do