hsbencher-1.14: HSBencher/Internal/App.hs
{-# LANGUAGE BangPatterns, NamedFieldPuns, ScopedTypeVariables, RecordWildCards, FlexibleContexts #-}
{-# LANGUAGE CPP, OverloadedStrings, TupleSections #-}
--------------------------------------------------------------------------------
-- NOTE: This is best when compiled with "ghc -threaded"
-- However, ideally for real benchmarking runs we WANT the waitForProcess below block the whole process.
-- However^2, currently [2012.05.03] when running without threads I get errors like this:
-- benchmark.run: bench_hive.log: openFile: resource busy (file is locked)
--------------------------------------------------------------------------------
{- | The Main module defining the HSBencher driver.
-}
module HSBencher.Internal.App
(defaultMainModifyConfig,
Flag(..), all_cli_options, fullUsageInfo)
where
----------------------------
-- Standard library imports
import Control.Concurrent
import qualified Control.Concurrent.Async as A
import Control.Exception (SomeException, try)
import Control.Monad.Reader
import qualified Data.ByteString.Char8 as B
import Data.IORef
import Data.List (intercalate, sortBy, intersperse, isInfixOf)
import qualified Data.Map as M
import Data.Maybe (isJust, fromJust, fromMaybe)
import Data.Version (versionBranch)
import Data.Word (Word64)
import Numeric (showFFloat)
import Prelude hiding (log)
import System.Console.GetOpt (getOpt', ArgOrder(Permute), OptDescr, usageInfo)
import System.Directory
import System.Environment (getArgs, getEnv, getProgName)
import System.Exit
import System.FilePath (splitFileName, (</>))
import System.Process (CmdSpec(..))
import Text.Printf
----------------------------
-- Additional libraries:
import qualified System.IO.Streams as Strm
import qualified System.IO.Streams.Concurrent as Strm
#ifdef USE_HYDRAPRINT
import UI.HydraPrint (hydraPrint, HydraConf(..), DeleteWinWhen(..), defaultHydraConf, hydraPrintStatic)
import Scripting.Parallel.ThreadPool (parForM)
#endif
----------------------------
-- Self imports:
import HSBencher.Types
import HSBencher.Internal.Utils
import HSBencher.Internal.Logging
import HSBencher.Internal.Config
import HSBencher.Internal.MeasureProcess
import Paths_hsbencher (version) -- Thanks, cabal!
----------------------------------------------------------------------------------------------------
hsbencherVersion :: String
hsbencherVersion = concat $ intersperse "." $ map show $
versionBranch version
-- | General usage information.
generalUsageStr :: String
generalUsageStr = unlines $
[
" ",
{-
" Many of these options can redundantly be set either when the benchmark driver is run,",
" or in the benchmark descriptions themselves. E.g. --with-ghc is just for convenience.",
"\n ENV VARS:",
-- No ENV vars currently! [2014.04.09]
" These environment variables control the behavior of the benchmark script:",
" ",
" Command line arguments take precedence over environment variables, if both apply.",
" ",
-}
" Note: This bench harness was built against hsbencher library version "++hsbencherVersion
]
--------------------------------------------------------------------------------
-- Compiling Benchmarks
--------------------------------------------------------------------------------
-- | Build a single benchmark in a single configuration.
compileOne :: (Int,Int) -> Benchmark DefaultParamMeaning -> [(DefaultParamMeaning,ParamSetting)] -> BenchM BuildResult
compileOne (iterNum,totalIters) Benchmark{target=testPath,cmdargs} cconf = do
cfg@Config{buildMethods, pathRegistry, doClean} <- ask
let (_diroffset,testRoot) = splitFileName testPath
flags = toCompileFlags cconf
paths = toCmdPaths cconf
bldid = makeBuildID testPath flags
log "\n--------------------------------------------------------------------------------"
log$ " Compiling Config "++show iterNum++" of "++show totalIters++
": "++testRoot++" (args \""++unwords cmdargs++"\") confID "++ show bldid
log "--------------------------------------------------------------------------------\n"
matches <- lift$
filterM (fmap isJust . (`filePredCheck` testPath) . canBuild) buildMethods
when (null matches) $ do
logT$ "ERROR, no build method matches path: "++testPath
logT$ " Tried methods: "++show(map methodName buildMethods)
logT$ " With file preds: "
forM buildMethods $ \ meth ->
logT$ " "++ show (canBuild meth)
lift exitFailure
logT$ printf "Found %d methods that can handle %s: %s"
(length matches) testPath (show$ map methodName matches)
let BuildMethod{methodName,clean,compile} = head matches
when (length matches > 1) $
logT$ " WARNING: resolving ambiguity, picking method: "++methodName
-- Add the static path information to the path information for this specific benchmark:
let newpathR = (M.union (M.fromList paths) pathRegistry)
when doClean $ clean newpathR bldid testPath
-- This is a bit weird... we could recast ALL fields of the Config
-- by specializing them to the benchmark particular compile
-- configuration. But right now we are doing that just for
-- pathRegistry:
let cfg2 = cfg{pathRegistry=newpathR}
-- Prefer the benchmark-local path definitions:
x <- compile cfg2 bldid flags testPath
logT$ "Compile finished, result: "++ show x
return x
--------------------------------------------------------------------------------
-- Running Benchmarks
--------------------------------------------------------------------------------
-- If the benchmark has already been compiled doCompile=False can be
-- used to skip straight to the execution.
runOne :: (Int,Int) -> BuildID -> BuildResult
-> Benchmark DefaultParamMeaning
-> [(DefaultParamMeaning,ParamSetting)] -> BenchM Bool
runOne (iterNum, totalIters) _bldid bldres
Benchmark{target=testPath, cmdargs, progname, benchTimeOut}
runconfig = do
log$ "\n--------------------------------------------------------------------------------"
log$ " Running Config "++show iterNum++" of "++show totalIters ++": "++testPath++" "++unwords cmdargs
-- " threads "++show numthreads++" (Env="++show envVars++")"
-- (1) Gather contextual information
----------------------------------------
-- Fullargs includes the runtime parameters as well as the original args:
(args, fullargs, testRoot) <- runA_gatherContext testPath cmdargs runconfig
-- (2) Now execute N trials:
----------------------------------------
-- TODO (One option woud be dynamic feedback where if the first one
-- takes a long time we don't bother doing more trials.)
(retries,nruns) <- runB_runTrials fullargs benchTimeOut bldres runconfig
-- (3) Produce output to the right places:
------------------------------------------
runC_produceOutput (args,fullargs) (retries,nruns) testRoot progname runconfig
------------------------------------------------------------
runA_gatherContext :: FilePath -> [String] -> [(a, ParamSetting)] -> ReaderT Config IO ([String], [String], FilePath)
runA_gatherContext testPath cmdargs runconfig = do
Config{shortrun, argsBeforeFlags} <- ask
let runParams = [ s | (_,RuntimeParam s) <- runconfig ]
runArgs = [ s | (_,RuntimeArg s) <- runconfig ]
args0 = cmdargs ++ runArgs
let args = if shortrun then shortArgs args0 else args0
let fullargs = if argsBeforeFlags
then args ++ runParams
else runParams ++ args
testRoot = fetchBaseName testPath
log$ nest 3 $ show$ doc$ map snd runconfig
log$ "--------------------------------------------------------------------------------\n"
pwd <- lift$ getCurrentDirectory
logT$ "(In directory "++ pwd ++")"
logT$ "Next run 'who', reporting users other than the current user. This may help with detectivework."
-- whos <- lift$ run "who | awk '{ print $1 }' | grep -v $USER"
whos <- lift$ runLines$ "who"
let whos' = map ((\ (h:_)->h) . words) whos
user <- lift$ getEnv "USER"
logT$ "Who_Output: "++ unwords (filter (/= user) whos')
return (args,fullargs,testRoot)
------------------------------------------------------------
runB_runTrials :: [String] -> Maybe Double -> BuildResult
-> [(a, ParamSetting)] -> ReaderT Config IO (Int,[RunResult])
runB_runTrials fullargs benchTimeOut bldres runconfig = do
Config{ retryFailed, trials } <- ask
let retryBudget = fromMaybe 0 retryFailed
trialLoop 1 trials retryBudget 0 []
where
trialLoop :: Int -> Int -> Int -> Int -> [RunResult] -> ReaderT Config IO (Int,[RunResult])
trialLoop ind trials retries retryAcc acc
| ind > trials = return (retryAcc, reverse acc)
| otherwise = do
Config{ runTimeOut, shortrun, harvesters } <- ask
log$ printf " Running trial %d of %d" ind trials
log " ------------------------"
let envVars = toEnvVars runconfig
let doMeasure1 cmddescr = do
SubProcess {wait,process_out,process_err} <-
lift$ measureProcess harvesters cmddescr
err2 <- lift$ Strm.map (B.append " [stderr] ") process_err
both <- lift$ Strm.concurrentMerge [process_out, err2]
mv <- echoStream (not shortrun) both
x <- lift wait
lift$ A.wait mv
logT$ " Subprocess finished and echo thread done.\n"
return x
-- I'm having problems currently [2014.07.04], where after about
-- 50 benchmarks (* 3 trials), all runs fail but there is NO
-- echo'd output. So here we try something simpler as a test.
let doMeasure2 cmddescr = do
(lines,result) <- lift$ measureProcessDBG harvesters cmddescr
mapM_ (logT . B.unpack) lines
logT $ "Subprocess completed with "++show(length lines)++" of output."
return result
-- doMeasure = doMeasure2 -- TEMP / Toggle me back later.
doMeasure = doMeasure1 -- TEMP / Toggle me back later.
this <- case bldres of
StandAloneBinary binpath -> do
-- NOTE: For now allowing rts args to include things like "+RTS -RTS", i.e. multiple tokens:
let command = binpath++" "++unwords fullargs
logT$ " Executing command: " ++ command
let timeout = if benchTimeOut == Nothing
then runTimeOut
else benchTimeOut
case timeout of
Just t -> logT$ " Setting timeout: " ++ show t
Nothing -> return ()
doMeasure CommandDescr{ command=ShellCommand command, envVars, timeout, workingDir=Nothing, tolerateError=False }
RunInPlace fn -> do
-- logT$ " Executing in-place benchmark run."
let cmd = fn fullargs envVars
logT$ " Generated in-place run command: "++show cmd
doMeasure cmd
if isError this
then if retries > 0
then do logT$ " Failed Trial! Retrying config, repeating trial "++
show ind++", "++show (retries - 1)++" retries left."
trialLoop ind trials (retries - 1) (retryAcc + 1) acc
else do logT$ " Failed Trial "++show ind++"! Out of retries, aborting remaining trials."
return (retries, this:acc)
else do -- When we advance, we reset the retry counter:
Config{ retryFailed } <- ask
trialLoop (ind+1) trials (fromMaybe 0 retryFailed) retryAcc (this:acc)
------------------------------------------------------------
runC_produceOutput :: ([String], [String]) -> (Int,[RunResult]) -> String -> Maybe String
-> [(DefaultParamMeaning, ParamSetting)] -> ReaderT Config IO Bool
runC_produceOutput (args,fullargs) (retries,nruns) testRoot progname runconfig = do
let numthreads = foldl (\ acc (x,_) ->
case x of
Threads n -> n
_ -> acc)
0 runconfig
sched = foldl (\ acc (x,_) ->
case x of
Variant s -> s
_ -> acc)
"none" runconfig
let pads n s = take (max 1 (n - length s)) $ repeat ' '
padl n x = pads n x ++ x
padr n x = x ++ pads n x
let thename = case progname of
Just s -> s
Nothing -> testRoot
Config{ keepgoing } <- ask
let exitCheck = when (any isError nruns && not keepgoing) $ do
log $ "\n Some runs were ERRORS; --keepgoing not used, so exiting now."
liftIO exitFailure
(_t1,_t2,_t3,_p1,_p2,_p3) <-
if all isError nruns then do
log $ "\n >>> MIN/MEDIAN/MAX (TIME,PROD) -- got only ERRORS: " ++show nruns
logOn [ResultsFile]$
printf "# %s %s %s %s %s" (padr 35 thename) (padr 20$ intercalate "_" fullargs)
(padr 8$ sched) (padr 3$ show numthreads) (" ALL_ERRORS"::String)
exitCheck
return ("","","","","","")
else do
exitCheck
let goodruns = filter (not . isError) nruns
-- Extract the min, median, and max:
sorted = sortBy (\ a b -> compare (gettime a) (gettime b)) goodruns
minR = head sorted
maxR = last sorted
medianR = sorted !! (length sorted `quot` 2)
let ts@[t1,t2,t3] = map (\x -> showFFloat Nothing x "")
[gettime minR, gettime medianR, gettime maxR]
prods@[p1,p2,p3] = map mshow [getprod minR, getprod medianR, getprod maxR]
mshow Nothing = "0"
mshow (Just x) = showFFloat (Just 2) x ""
-- These are really (time,prod) tuples, but a flat list of
-- scalars is simpler and readable by gnuplot:
formatted = (padl 15$ unwords $ ts)
++" "++ unwords prods -- prods may be empty!
log $ "\n >>> MIN/MEDIAN/MAX (TIME,PROD) " ++ formatted
logOn [ResultsFile]$
printf "%s %s %s %s %s" (padr 35 thename) (padr 20$ intercalate "_" fullargs)
(padr 8$ sched) (padr 3$ show numthreads) formatted
-- These should be either all Nothing or all Just:
let jittimes0 = map getjittime goodruns
misses = length (filter (==Nothing) jittimes0)
jittimes <- if misses == length goodruns
then return ""
else if misses == 0
then return $ unwords (map (show . fromJust) jittimes0)
else do log $ "WARNING: got JITTIME for some runs: "++show jittimes0
log " Zeroing those that did not report."
return $ unwords (map (show . fromMaybe 0) jittimes0)
Config{ trials } <- ask
let result =
emptyBenchmarkResult
{ _PROGNAME = case progname of
Just s -> s
Nothing -> testRoot
, _VARIANT = sched
, _ARGS = args
, _THREADS = numthreads
, _MINTIME = gettime minR
, _MEDIANTIME = gettime medianR
, _MAXTIME = gettime maxR
, _MINTIME_PRODUCTIVITY = getprod minR
, _MEDIANTIME_PRODUCTIVITY = getprod medianR
, _MEDIANTIME_ALLOCRATE = getallocrate medianR
, _MEDIANTIME_MEMFOOTPRINT = getmemfootprint medianR
, _MAXTIME_PRODUCTIVITY = getprod maxR
, _RUNTIME_FLAGS = unwords [ s | (_,RuntimeParam s) <- runconfig ]
, _ALLTIMES = unwords$ map (show . gettime) goodruns
, _ALLJITTIMES = jittimes
, _TRIALS = trials
, _RETRIES = retries
-- Should the user specify how the
-- results over many goodruns are reduced ?
-- I think so.
, _CUSTOM = custom (head goodruns) -- experimenting
}
conf <- ask
result' <- liftIO$ augmentResultWithConfig conf result
-- Upload results to plugin backends:
conf2@Config{ plugIns } <- ask
forM_ plugIns $ \ (SomePlugin p) -> do
--JS: May 21 2014, added try and case on result.
result3 <- liftIO$ try (plugUploadRow p conf2 result') :: ReaderT Config IO (Either SomeException ())
case result3 of
Left err -> logT$("plugUploadRow:Failed, error: \n"++
"------------------begin-error----------------------\n"++
show err ++
"\n-------------------end-error-----------------------\n"
)
Right () -> return ()
return ()
return (t1,t2,t3,p1,p2,p3)
-- If -keepgoing is set, we consider only errors on all runs to
-- invalidate the whole benchmark job:
return (not (all isError nruns))
--------------------------------------------------------------------------------
-- | Write the results header out stdout and to disk.
printBenchrunHeader :: BenchM ()
printBenchrunHeader = do
Config{trials, maxthreads, pathRegistry,
logOut, resultsOut, stdOut, benchversion, shortrun, gitInfo=(branch,revision,depth) } <- ask
liftIO $ do
-- let (benchfile, ver) = benchversion
let ls :: [IO String]
ls = [ e$ "# TestName Variant NumThreads MinTime MedianTime MaxTime Productivity1 Productivity2 Productivity3"
, e$ "# "
, e$ "# `date`"
, e$ "# `uname -a`"
, e$ "# Ran by: `whoami` "
, e$ "# Determined machine to have "++show maxthreads++" hardware threads."
, e$ "# "
, e$ "# Running each test for "++show trials++" trial(s)."
-- , e$ "# Benchmarks_File: " ++ benchfile
-- , e$ "# Benchmarks_Variant: " ++ if shortrun then "SHORTRUN" else whichVariant benchfile
-- , e$ "# Benchmarks_Version: " ++ show ver
, e$ "# Git_Branch: " ++ branch
, e$ "# Git_Hash: " ++ revision
, e$ "# Git_Depth: " ++ show depth
-- , e$ "# Using the following settings from environment variables:"
-- , e$ "# ENV BENCHLIST=$BENCHLIST"
-- , e$ "# ENV THREADS= $THREADS"
-- , e$ "# ENV TRIALS= $TRIALS"
-- , e$ "# ENV SHORTRUN= $SHORTRUN"
-- , e$ "# ENV KEEPGOING= $KEEPGOING"
-- , e$ "# ENV GHC= $GHC"
-- , e$ "# ENV GHC_FLAGS= $GHC_FLAGS"
-- , e$ "# ENV GHC_RTS= $GHC_RTS"
-- , e$ "# ENV ENVS= $ENVS"
, e$ "# Path registry: "++show pathRegistry
]
ls' <- sequence ls
forM_ ls' $ \line -> do
Strm.write (Just$ B.pack line) resultsOut
Strm.write (Just$ B.pack line) logOut
Strm.write (Just$ B.pack line) stdOut
return ()
where
-- This is a hack for shell expanding inside a string:
e :: String -> IO String
e s =
runSL ("echo \""++s++"\"")
-- readCommand ("echo \""++s++"\"")
-- readProcess "echo" ["\""++s++"\""] ""
----------------------------------------------------------------------------------------------------
-- Main Script
----------------------------------------------------------------------------------------------------
-- | TODO: Eventually this will make sense when all config can be read from the environment, args, files.
defaultMain :: IO ()
defaultMain = do
-- benchF = get "BENCHLIST" "benchlist.txt"
-- putStrLn$ hsbencher_tag ++ " Reading benchmark list from file: "
error "FINISHME: defaultMain requires reading benchmark list from a file. Implement it!"
-- defaultMainWithBenchmarks undefined
-- | In this version, user provides a list of benchmarks to run, explicitly.
defaultMainWithBenchmarks :: [Benchmark DefaultParamMeaning] -> IO ()
defaultMainWithBenchmarks benches = do
defaultMainModifyConfig (\ conf -> conf{ benchlist=benches })
-- | Multiple lines of usage info help docs.
fullUsageInfo :: String
fullUsageInfo =
"\nUSAGE: naked command line arguments are patterns that select the benchmarks to run.\n"++
(concat (map (uncurry usageInfo) all_cli_options)) ++
generalUsageStr
-- | Remove a plugin from the configuration based on its plugName
removePlugin :: Plugin p => p -> Config -> Config
removePlugin p cfg =
cfg { plugIns = filter byNom (plugIns cfg)}
where
byNom (SomePlugin p1) = plugName p1 /= plugName p
--------------------------------------------------------------------------------
doShowHelp :: [SomePlugin] -> IO ()
doShowHelp allplugs = do
putStrLn$ "\nUSAGE: [set ENV VARS] "++my_name++" [CMDLN OPTS]"
putStrLn$ "\nNote: \"CMDLN OPTS\" includes patterns that select which benchmarks"
putStrLn$ " to run, based on name."
mapM putStr (map (uncurry usageInfo) all_cli_options)
putStrLn ""
putStrLn $ show (length allplugs) ++ " plugins enabled: "++
show [ plugName p | SomePlugin p <- allplugs ]
putStrLn ""
forM_ allplugs $ \ (SomePlugin p) -> do
putStrLn $ "["++ plugName p++"] "++ ((uncurry usageInfo) (plugCmdOpts p))
putStrLn$ generalUsageStr
-- TODO/FIXME: Break up the giant function below. Also move to using
-- a StateT to store the Config, and reduce the verbosity of the
-- output when "Harvesting environment data".
-- | An even more flexible version allows the user to install a hook which modifies
-- the configuration just before bencharking begins. All trawling of the execution
-- environment (command line args, environment variables) happens BEFORE the user
-- sees the configuration.
--
-- This function doesn't take a benchmark list separately, because that simply
-- corresponds to the 'benchlist' field of the output 'Config'.
defaultMainModifyConfig :: (Config -> Config) -> IO ()
defaultMainModifyConfig modConfig = do
id <- myThreadId
writeIORef main_threadid id
my_name <- getProgName
cli_args <- getArgs
let (options,plainargs,_unrec,errs) = getOpt' Permute (concat$ map snd all_cli_options) cli_args
-- This ugly method avoids needing an Eq instance:
let recomp = null [ () | NoRecomp <- options]
showHelp = not$ null [ () | ShowHelp <- options]
gotVersion = not$ null [ () | ShowVersion <- options]
showBenchs = not$ null [ () | ShowBenchmarks <- options]
cabalAllowed = not$ null [ () | NoCabal <- options]
parBench = not$ null [ () | ParBench <- options]
disabled = [ s | DisablePlug s <- options ]
when gotVersion $ do
putStrLn$ "hsbencher version "++ hsbencherVersion
-- (unwords$ versionTags version)
exitSuccess
------------------------------------------------------------
putStrLn$ "\n"++hsbencher_tag++"Harvesting environment data to build Config."
conf0 <- getConfig options []
-- The list of benchmarks can optionally be narrowed to match any of the given patterns.
let conf1 = modConfig conf0
-- The phasing here is rather funny. We need to get the initial config to know
-- WHICH plugins are active. And then their individual per-plugin configs need to
-- be computed and added to the global config.
let plugnames = [ plugName p | SomePlugin p <- plugIns conf1 ]
let plugs = [ if (or [ isInfixOf d (plugName p)| d <- disabled ])
then Right (plugName p, SomePlugin p) -- Disabled
else Left (plugName p, SomePlugin p) -- Enabled
| SomePlugin p <- plugIns conf1 ]
let offplugs = [ n | Right (n, _) <- plugs ]
allplugs = [ sp | Left (_, sp) <- plugs ]
unless (null offplugs) $
putStrLn $ hsbencher_tag ++ " DISABLED plugins that were compiled/linked in: "++unwords offplugs
------------------------------------------------------------
let fullBenchList =
case conf1 of
Config{benchlist=ls} ->
(unlines [ (maybe "" (++" = ") progname) ++
(target ++ (unwords cmdargs))
| Benchmark{progname, cmdargs,target} <- ls])
when showBenchs $ do putStrLn ("All benchmarks handled by this script:\n"++fullBenchList)
exitSuccess
unless (null errs) $ do
putStrLn$ "Errors parsing command line options:"
mapM_ (putStr . (" "++)) errs
doShowHelp allplugs
exitFailure
when showHelp $ do doShowHelp allplugs; exitSuccess
------------------------------------------------------------
-- Fully populate the per-plugin configurations, folding in command line args:
--
-- Hmm, not really a strong reason to *combine* the options lists, rather we do
-- them one at a time:
let pconfs = [ (plugName p, SomePluginConf p pconf)
| (SomePlugin p) <- (plugIns conf1)
, let (_pusage,popts) = plugCmdOpts p
, let (o2,_,_,_) = getOpt' Permute popts cli_args
, let pconf = foldFlags p o2 (getMyConf p conf1)
]
let conf2 = conf1 { plugInConfs = M.fromList pconfs }
-- Combine all plugins command line options, and reparse the command line.
putStrLn$ hsbencher_tag++(show$ length allplugs)++" plugins configured ("++
concat (intersperse ", " [ plugName p | SomePlugin p <- allplugs ])
++"), now initializing them."
-- TODO/FIXME: CATCH ERRORS... should remove the plugin from the list if it errors on init.
-- JS attempted fix
conf3 <- foldM (\ cfg (SomePlugin p) ->
do result <- try (plugInitialize p cfg) :: IO (Either SomeException Config)
case result of
Left err -> do
putStrLn (hsbencher_tag++"Plugin Init FAILED! Error:\n"++show err)
return $ removePlugin p cfg
-- cannot log here, only "chatter".
Right c -> return c
) conf2 allplugs
putStrLn$ hsbencher_tag++" plugin init complete."
-------------------------------------------------------------------
-- Next prune the list of benchmarks to those selected by the user:
let cutlist = case plainargs of
[] -> benchlist conf3
patterns -> filter (\ Benchmark{target,cmdargs,progname} ->
any (\pat ->
isInfixOf pat target ||
isInfixOf pat (fromMaybe "" progname) ||
any (isInfixOf pat) cmdargs
)
patterns)
(benchlist conf3)
let conf4@Config{benchlist} = conf3{benchlist=cutlist}
------------------------------------------------------------
rootDir <- getCurrentDirectory
runReaderT
(do
unless (null plainargs) $ do
let len = (length cutlist)
logT$"There were "++show len++" benchmarks matching patterns: "++show plainargs
when (len == 0) $ do
error$ "Expected at least one pattern to match!. All benchmarks: \n"++
fullBenchList
logT$"Beginning benchmarking, root directory: "++rootDir
let globalBinDir = rootDir </> "bin"
when recomp $ do
logT$"Clearing any preexisting files in ./bin/"
lift$ do
-- runSimple "rm -f ./bin/*"
-- Yes... it's posix dependent. But right now I don't see a good way to
-- delete the contents a dir without (1) following symlinks or (2) assuming
-- either the unix package or unix shell support (rm).
--- Ok, what the heck, deleting recursively:
dde <- doesDirectoryExist globalBinDir
when dde $ removeDirectoryRecursive globalBinDir
lift$ createDirectoryIfMissing True globalBinDir
logT "Writing header for result data file:"
printBenchrunHeader
unless recomp $ log "[!!!] Skipping benchmark recompilation!"
let
benches' = map (\ b -> b { configs= compileOptsOnly (configs b) })
benchlist
cccfgs = map (enumerateBenchSpace . configs) benches' -- compile configs
cclengths = map length cccfgs
totalcomps = sum cclengths
log$ "\n--------------------------------------------------------------------------------"
logT$ "Running all benchmarks for all settings ..."
logT$ "Compiling: "++show totalcomps++" total configurations of "++ show (length benchlist)++" benchmarks"
let indent n str = unlines $ map (replicate n ' ' ++) $ lines str
printloop _ [] = return ()
printloop mp (Benchmark{target,cmdargs,configs} :tl) = do
log$ " * Benchmark/args: "++target++" "++show cmdargs
case M.lookup configs mp of
Nothing -> log$ indent 4$ show$ doc configs
Just trg0 -> log$ " ...same config space as "++show trg0
printloop (M.insertWith (\ _ x -> x) configs target mp) tl
-- log$ "Benchmarks/compile options: "++show (doc benches')
printloop M.empty benchlist
log$ "--------------------------------------------------------------------------------"
if parBench then do
unless rtsSupportsBoundThreads $ error (my_name++" was NOT compiled with -threaded. Can't do --par.")
{-
--------------------------------------------------------------------------------
-- Parallel version:
numProcs <- liftIO getNumProcessors
lift$ putStrLn$ "[!!!] Compiling in Parallel, numProcessors="++show numProcs++" ... "
when recomp $ liftIO$ do
when hasCabalFile (error "Currently, cabalized build does not support parallelism!")
(strms,barrier) <- parForM numProcs (zip [1..] pruned) $ \ outStrm (confnum,bench) -> do
outStrm' <- Strm.unlines outStrm
let conf' = conf { stdOut = outStrm' }
runReaderT (compileOne bench (confnum,length pruned)) conf'
return ()
catParallelOutput strms stdOut
res <- barrier
return ()
Config{shortrun,doFusionUpload} <- ask
if shortrun && not doFusionUpload then liftIO$ do
putStrLn$ "[!!!] Running in Parallel..."
(strms,barrier) <- parForM numProcs (zip [1..] pruned) $ \ outStrm (confnum,bench) -> do
outStrm' <- Strm.unlines outStrm
let conf' = conf { stdOut = outStrm' }
runReaderT (runOne bench (confnum,totalcomps)) conf'
catParallelOutput strms stdOut
_ <- barrier
return ()
else do
-- Non-shortrun's NEVER run multiple benchmarks at once:
forM_ (zip [1..] allruns) $ \ (confnum,bench) ->
runOne bench (confnum,totalcomps)
return ()
-}
else do
--------------------------------------------------------------------------------
-- Serial version:
-- TODO: make this a foldlM:
let allruns = map (enumerateBenchSpace . configs) benchlist
allrunsLens = map length allruns
totalruns = sum allrunsLens
let
-- Here we lazily compile benchmarks as they become required by run configurations.
runloop :: Int
-> M.Map BuildID (Int, Maybe BuildResult)
-> M.Map FilePath BuildID -- (S.Set ParamSetting)
-> [(Benchmark DefaultParamMeaning, [(DefaultParamMeaning,ParamSetting)])]
-> Bool -> BenchM Bool
runloop _ _ _ [] b = return b
runloop !iter !board !lastConfigured (nextrun:rest) allpassed = do
-- lastConfigured keeps track of what configuration was last built in
-- a directory that is used for `RunInPlace` builds.
let (bench,params) = nextrun
ccflags = toCompileFlags params
bid = makeBuildID (target bench) ccflags
case M.lookup bid board of
Nothing -> error$ "HSBencher: Internal error: Cannot find entry in map for build ID: "++show bid
Just (ccnum, Nothing) -> do
res <- compileOne (ccnum,totalcomps) bench params
let board' = M.insert bid (ccnum, Just res) board
lastC' = M.insert (target bench) bid lastConfigured
-- runloop iter board' (nextrun:rest)
b <- runOne (iter,totalruns) bid res bench params
runloop (iter+1) board' lastC' rest (allpassed && b)
Just (ccnum, Just bldres) ->
let proceed = do b <- runOne (iter,totalruns) bid bldres bench params
runloop (iter+1) board lastConfigured rest (allpassed && b)
in
case bldres of
StandAloneBinary _ -> proceed
RunInPlace _ ->
-- Here we know that some previous compile with the same BuildID inserted this here.
-- But the relevant question is whether some other config has stomped on it in the meantime.
case M.lookup (target bench) lastConfigured of
Nothing -> error$"HSBencher: Internal error, RunInPlace in the board but not lastConfigured!: "
++(target bench)++ " build id "++show bid
Just bid2 ->
if bid == bid2
then do logT$ "Skipping rebuild of in-place benchmark: "++bid
proceed
else runloop iter (M.insert bid (ccnum,Nothing) board)
lastConfigured (nextrun:rest) allpassed
-- Keeps track of what's compiled.
initBoard _ [] acc = acc
initBoard !iter ((bench,params):rest) acc =
let bid = makeBuildID (target bench) $ toCompileFlags params
base = fetchBaseName (target bench)
dfltdest = globalBinDir </> base ++"_"++bid in
case M.lookup bid acc of
Just _ -> initBoard iter rest acc
Nothing ->
let elm = if recomp
then (iter, Nothing)
else (iter, Just (StandAloneBinary dfltdest))
in
initBoard (iter+1) rest (M.insert bid elm acc)
zippedruns = (concat$ zipWith (\ b cfs -> map (b,) cfs) benchlist allruns)
unless recomp $ logT$ "Recompilation disabled, assuming standalone binaries are in the expected places!"
let startBoard = initBoard 1 zippedruns M.empty
Config{skipTo, runOnly} <- ask
(ix,runs') <- case skipTo of
Nothing -> return (1,zippedruns)
Just ix -> do logT$" !!! WARNING: SKIPPING AHEAD in configuration space; jumping to: "++show ix
return (ix, drop (ix-1) zippedruns)
runs'' <- case runOnly of
Nothing -> return runs'
Just num -> do logT$" !!! WARNING: TRUNCATING config space to only run "++show num++" configs."
return (take num runs')
win <- runloop ix startBoard M.empty runs'' True
unless win $ do
log$ "\n--------------------------------------------------------------------------------"
log " Finished benchmarks, but some errored out, marking this job as a failure."
log$ "--------------------------------------------------------------------------------"
liftIO$ exitFailure
return ()
{-
do Config{logOut, resultsOut, stdOut} <- ask
liftIO$ Strm.write Nothing logOut
liftIO$ Strm.write Nothing resultsOut
-}
log$ "\n--------------------------------------------------------------------------------"
log " Finished with all benchmark configurations. Success."
log$ "--------------------------------------------------------------------------------"
liftIO$ exitSuccess
)
conf4
-- Several different options for how to display output in parallel:
catParallelOutput :: [Strm.InputStream B.ByteString] -> Strm.OutputStream B.ByteString -> IO ()
catParallelOutput strms stdOut = do
case 4 of
#ifdef USE_HYDRAPRINT
-- First option is to create N window panes immediately.
1 -> do
hydraPrintStatic defaultHydraConf (zip (map show [1..]) strms)
2 -> do
srcs <- Strm.fromList (zip (map show [1..]) strms)
hydraPrint defaultHydraConf{deleteWhen=Never} srcs
#endif
-- This version interleaves their output lines (ugly):
3 -> do
strms2 <- mapM Strm.lines strms
interleaved <- Strm.concurrentMerge strms2
Strm.connect interleaved stdOut
-- This version serializes the output one worker at a time:
4 -> do
strms2 <- mapM Strm.lines strms
merged <- Strm.concatInputStreams strms2
-- Strm.connect (head strms) stdOut
Strm.connect merged stdOut
----------------------------------------------------------------------------------------------------
-- * GENERIC HELPER ROUTINES
----------------------------------------------------------------------------------------------------
-- These should go in another module.......
didComplete :: RunResult -> Bool
didComplete RunCompleted{} = True
didComplete _ = False
isError :: RunResult -> Bool
isError ExitError{} = True
isError _ = False
getprod :: RunResult -> Maybe Double
getprod RunCompleted{productivity} = productivity
getprod RunTimeOut{} = Nothing
getprod x = error$"Cannot get productivity from: "++show x
getallocrate :: RunResult -> Maybe Word64
getallocrate RunCompleted{allocRate} = allocRate
getallocrate _ = Nothing
getmemfootprint :: RunResult -> Maybe Word64
getmemfootprint RunCompleted{memFootprint} = memFootprint
getmemfootprint _ = Nothing
gettime :: RunResult -> Double
gettime RunCompleted{realtime} = realtime
gettime RunTimeOut{} = posInf
gettime x = error$"Cannot get realtime from: "++show x
getjittime :: RunResult -> Maybe Double
getjittime RunCompleted{jittime} = jittime
getjittime _ = Nothing
posInf :: Double
posInf = 1/0
-- Compute a cut-down version of a benchmark's args list that will do
-- a short (quick) run. The way this works is that benchmarks are
-- expected to run and do something quick if they are invoked with no
-- arguments. (A proper benchmarking run, therefore, requires larger
-- numeric arguments be supplied.)
--
shortArgs :: [String] -> [String]
shortArgs _ls = []
-- shortArgs [] = []
-- DISABLING:
-- HOWEVER: there's a further hack here which is that leading
-- non-numeric arguments are considered qualitative (e.g. "monad" vs
-- "sparks") rather than quantitative and are not pruned by this
-- function.
-- shortArgs (h:tl) | isNumber h = []
-- | otherwise = h : shortArgs tl
----------------------------------------------------------------------------------------------------
nest :: Int -> String -> String
nest n str = remlastNewline $ unlines $
map (replicate n ' ' ++) $
lines str
where
remlastNewline str =
case reverse str of
'\n':rest -> reverse rest
_ -> str