For now bring cilk up and down each time, this seems to avoid issue #4

AccelerateHS · Feb 12, 2014 · 5e99ec2 · 5e99ec2
1 parent 410617d
commit 5e99ec2
Show file tree

Hide file tree

Showing 3 changed files with 30 additions and 6 deletions.
diff --git a/icc-opencl/DEVLOG.md b/icc-opencl/DEVLOG.md
@@ -59,6 +59,16 @@ like ICC cannot fail gracefully in this case.  In this case would be
 perfectly happy if I could have a __declspec(vector) fail to vectorize
 but still compile.
 
+[2014.02.12] {Segfaults may actually be a Cilk bug}
+----------------------------------------------------
+
+I doubt Cilk gets used much in this repeated-dlopen-mode.  I started
+seeing this error:
+
+   "Cilk runtime error: thread creation (3) failed: 11" #4 
+
+For which I filed issue #4:
+
+   https://github.com/AccelerateHS/accelerate-backend-kit/issues/4
 
 
-
diff --git a/icc-opencl/Data/Array/Accelerate/Cilk/JITRuntime.hs b/icc-opencl/Data/Array/Accelerate/Cilk/JITRuntime.hs
@@ -9,8 +9,6 @@ module Data.Array.Accelerate.Cilk.JITRuntime (run, runNamed, rawRunIO) where
 
 import           Data.Array.Accelerate (Acc)
 import qualified Data.Array.Accelerate.Array.Sugar as Sug
-import qualified Data.Array.Accelerate.BackendKit.IRs.SimpleAcc   as S
-import           Data.Array.Accelerate.BackendKit.IRs.SimpleAcc (Type(..), Const(..))
 import qualified Data.Array.Accelerate.BackendKit.SimpleArray     as SA
 import           Data.Array.Accelerate.BackendKit.Utils.Helpers (maybtrace, dbg)
 import           Data.Array.Accelerate.BackendKit.CompilerPipeline (phase0, phase1, phase2, repackAcc)
@@ -46,6 +44,8 @@ import Data.Array.Accelerate.BackendKit.Phase3.ToGPUIR           (convertToGPUIR
 import Data.Array.Accelerate.BackendKit.Phase3.DesugarGenerate   (desugarGenerate)
 import Data.Array.Accelerate.BackendKit.Phase3.FuseGenReduce     (fuseGenReduce)
 import Data.Array.Accelerate.BackendKit.CompilerPipeline         (runPass)
+import qualified Data.Array.Accelerate.BackendKit.IRs.SimpleAcc   as S
+import           Data.Array.Accelerate.BackendKit.IRs.SimpleAcc (Type(..), Const(..))
 import qualified Data.Array.Accelerate.BackendKit.IRs.CLike     as C
 import qualified Data.Array.Accelerate.BackendKit.IRs.GPUIR     as G
 import           Data.Array.Accelerate.BackendKit.IRs.Metadata   (ArraySizeEstimate, FreeVars)
@@ -136,7 +136,12 @@ rawRunIO pm name prog = do
          CilkParallel -> pickCC (error "ICC not found!  Need it for Cilk backend.")
 
   let suppress = if dbg>0 then " -g " else " -w " -- No warnings leaking through to the user.
+-- #define EXE_OUTPUT
+#ifdef EXE_OUTPUT
+      ccCmd = cc++suppress++" -lcilkrts -std=c99 "++thisprog++".c -o "++thisprog++".exe"
+#else
       ccCmd = cc++suppress++" -shared -fPIC -std=c99 "++thisprog++".c -o "++thisprog++".so"
+#endif
   dbgPrint 2 $ "[JIT]   Compiling with: "++ ccCmd
 
   t1 <- getCurrentTime 
@@ -148,10 +153,10 @@ rawRunIO pm name prog = do
     ExitSuccess -> return ()
     ExitFailure c -> error$"C Compiler failed with code "++show c
   -- Run the program and capture its output:
-#if 0     
+#ifdef EXE_OUTPUT
   dbgPrint 1$ "[JIT] Invoking external executable: "++ thisprog++".exe"  
   result <- readProcess ("./"++thisprog++".exe") [] ""
-  let elts = tyToElts (S.progType prog)
+  let elts = tyToElts (C.progType prog)
   dbgPrint 1$ "[JIT] Executable completed, parsing results, element types "++
      show elts++", "++show (length result)++" characters:\n "++take 80 result
   return$ parseMultiple result elts

diff --git a/icc-opencl/Data/Array/Accelerate/Shared/EmitC.hs b/icc-opencl/Data/Array/Accelerate/Shared/EmitC.hs
@@ -94,6 +94,7 @@ instance EmitBackend CEmitter where
     case e of
       CEmitter Sequential   _ -> return ()
       CEmitter CilkParallel _ -> do include "cilk/cilk.h"
+                                    include "cilk/cilk_api_linux.h"
                                     include "cilk/reducer.h"
     mapM_ (emitLine . strToSyn) $ P.lines headerCode 
 
@@ -247,7 +248,7 @@ mainBody isCMain e prog@GPUProg{progBinds, progResults, sizeEnv} = do
         allResults = standardResultOrder (progResults)
         allArrays  = map P.fst allResults
         allUses    = S.fromList $ map (\(a,b,c) -> a) useBinds
-        body       = do            
+        body0      = do
            comm "First we EXECUTE the program by executing each array op in order:"
            mapM_ (execBind e prog) (L.zip [0..] progBinds)
 
@@ -274,6 +275,14 @@ mainBody isCMain e prog@GPUProg{progBinds, progResults, sizeEnv} = do
                if S.member vr allUses P.|| elem vr allArrays
                then return ()
                else freeCStorage ty (varSyn vr)
+
+        -- [2014.02.12] For debugging purposes let's bring Cilk up and down explicitly each time:
+        body = 
+         case e of
+           CEmitter Sequential   _ -> body0
+           CEmitter CilkParallel _ -> do emitStmt$ function "__cilkrts_init" []
+                                         body0
+                                         emitStmt$ function "__cilkrts_end_cilk" []
     _ <- if isCMain
          then rawFunDef "int" "main" [] (do body; return_ 0)
          else rawFunDef "void" "MainProg" [("struct ArgRecord*",globalArgs), ("struct ResultRecord*",globalResults)] body