libdispatch based ParallelJobs is not enough parallel

https://bugs.webkit.org/show_bug.cgi?id=66378

Reviewed by Zoltan Herczeg.

Use the appropriate libdispatch API for our use case.
Throw away the hard coded limit of parallel threads
and use dispatch_apply with the default priority normal
queue istead of using our own custom serial queue (which
was a misuse of the API). Enabling PARALLEL_JOBS is now
a 60% win (2.63x as fast) on the methanol benchmark
(https://gitorious.org/methanol) with an SVG centric test set
while the old implementation was almost identical (less than 5% win).

* wtf/ParallelJobsLibdispatch.h:
(WTF::ParallelEnvironment::ParallelEnvironment):
(WTF::ParallelEnvironment::execute):

git-svn-id: http://svn.webkit.org/repository/webkit/trunk@96492 268f45cc-cd09-0410-ab3c-d52691b4dbfc

diff --git a/Source/JavaScriptCore/ChangeLog b/Source/JavaScriptCore/ChangeLog
index dbb91fc..044ddfc 100644 (file)
--- a/Source/JavaScriptCore/ChangeLog
+++ b/Source/JavaScriptCore/ChangeLog
@@ -1,3 +1,23 @@
+2011-10-03  Balazs Kelemen  <kbalazs@webkit.org>
+
+        libdispatch based ParallelJobs is not enough parallel
+        https://bugs.webkit.org/show_bug.cgi?id=66378
+
+        Reviewed by Zoltan Herczeg.
+
+        Use the appropriate libdispatch API for our use case.
+        Throw away the hard coded limit of parallel threads
+        and use dispatch_apply with the default priority normal
+        queue istead of using our own custom serial queue (which
+        was a misuse of the API). Enabling PARALLEL_JOBS is now
+        a 60% win (2.63x as fast) on the methanol benchmark
+        (https://gitorious.org/methanol) with an SVG centric test set
+        while the old implementation was almost identical (less than 5% win).
+
+        * wtf/ParallelJobsLibdispatch.h:
+        (WTF::ParallelEnvironment::ParallelEnvironment):
+        (WTF::ParallelEnvironment::execute):
+
 2011-10-02  Zoltan Herczeg  <zherczeg@webkit.org>
 
         [Qt]REGRESSION(r95912): It made sputnik tests flakey

diff --git a/Source/JavaScriptCore/wtf/ParallelJobsLibdispatch.h b/Source/JavaScriptCore/wtf/ParallelJobsLibdispatch.h
index 687c771..ca7d9a4 100644 (file)
--- a/Source/JavaScriptCore/wtf/ParallelJobsLibdispatch.h
+++ b/Source/JavaScriptCore/wtf/ParallelJobsLibdispatch.h
@@ -34,23 +34,18 @@
 
 namespace WTF {
 
-static const int maxParallelThreads = 2;
-
 class ParallelEnvironment {
     WTF_MAKE_FAST_ALLOCATED;
 public:
     typedef void (*ThreadFunction)(void*);
 
-    ParallelEnvironment(ThreadFunction threadFunction, size_t sizeOfParameter, int requestedJobNumber) :
-        m_threadFunction(threadFunction),
-        m_sizeOfParameter(sizeOfParameter)
+    ParallelEnvironment(ThreadFunction threadFunction, size_t sizeOfParameter, int requestedJobNumber)
+        : m_threadFunction(threadFunction)
+        , m_sizeOfParameter(sizeOfParameter)
+        , m_numberOfJobs(requestedJobNumber)
     {
-        if (!requestedJobNumber || requestedJobNumber > maxParallelThreads)
-            requestedJobNumber = maxParallelThreads;
-
-        ASSERT(requestedJobNumber > 0);
-
-        m_numberOfJobs = requestedJobNumber;
+        // We go with the requested number of jobs. libdispatch will distribute the work optimally.
+        ASSERT_ARG(requestedJobNumber, requestedJobNumber > 0);
     }
 
     int numberOfJobs()
@@ -60,16 +55,9 @@ public:
 
     void execute(unsigned char* parameters)
     {
-        // libdispatch is NOT supported inside a template
-        dispatch_queue_t parallelJobsQueue = dispatch_queue_create("ParallelJobs", 0);
-
-        for (int i = 0; i < m_numberOfJobs - 1; ++i) {
-            dispatch_async(parallelJobsQueue, ^{(*m_threadFunction)(parameters);});
-            parameters += m_sizeOfParameter;
-        }
+        static dispatch_queue_t globalQueue = dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0);
 
-        // The work for the main thread. Wait until all jobs are done.
-        dispatch_sync(parallelJobsQueue, ^{(*m_threadFunction)(parameters);});
+        dispatch_apply(m_numberOfJobs, globalQueue, ^(size_t i) { (*m_threadFunction)(parameters + (m_sizeOfParameter * i)); });
     }
 
 private: