[Boost-commit] svn:boost r78179 - in trunk/libs/context: doc performance

Subject: [Boost-commit] svn:boost r78179 - in trunk/libs/context: doc performance
From: oliver.kowalke_at_[hidden]
Date: 2012-04-24 08:36:56

Author: olli
Date: 2012-04-24 08:36:55 EDT (Tue, 24 Apr 2012)
New Revision: 78179
URL: http://svn.boost.org/trac/boost/changeset/78179

Log:
context: documentation update, performance test with boost.function

Text files modified: 
   trunk/libs/context/doc/performance.qbk         |    12 ++++++++----                            
   trunk/libs/context/doc/rationale.qbk           |    17 +++++------------                       
   trunk/libs/context/performance/performance.cpp |    27 +++++++++++++++++++++++++++             
   3 files changed, 40 insertions(+), 16 deletions(-)

Modified: trunk/libs/context/doc/performance.qbk
==============================================================================
--- trunk/libs/context/doc/performance.qbk	(original)
+++ trunk/libs/context/doc/performance.qbk	2012-04-24 08:36:55 EDT (Tue, 24 Apr 2012)
@@ -15,19 +15,23 @@
 'variant=release cxxflags=-DBOOST_DISABLE_ASSERTS'.
 
 The numbers in the table are the number of cycles per iteration, based upon an
-average computed over 1000 iterations.
+average computed over 256 iterations.
 
 [table Perfomance of context switch
-    [[Platform] [ucontext_t] [fcontext_t]]
+    [[Platform] [ucontext_t] [fcontext_t with fpu] [fcontext_t without fpu] [boost::function]]
     [
         [AMD Athlon 64 DualCore 4400+ (32bit Linux)]
-        [1771 cycles]
-        [146 cycles]
+        [846 cycles]
+        [65 cycles]
+        [46 cycles]
+        [43 cycles]
     ]
     [
         [Intel Core2 Quad (64bit Linux)]
         [1739 cycles]
         [117 cycles]
+        [117 cycles]
+        [117 cycles]
     ]
 ]
 

Modified: trunk/libs/context/doc/rationale.qbk
==============================================================================
--- trunk/libs/context/doc/rationale.qbk	(original)
+++ trunk/libs/context/doc/rationale.qbk	2012-04-24 08:36:55 EDT (Tue, 24 Apr 2012)
@@ -35,6 +35,11 @@
 
 [note Context switches do not preserve the signal mask on UNIX systems.]
 
+Because the assembler code uses the byte layout of __fcontext__ to access 
+its members __fcontext_t must be a POD. This requires that __fcontext__ has only a
+default constructor, no visibility keywords (e.g. private, public, protected), no
+virtual methods and all members and base clases are PODs too.
+
 
 [section Other APIs ]
 
@@ -103,9 +108,6 @@
 [footnote [@boost:/libs/context/doc/pdf/calling-conventions.pdf 'Calling Conventions',
 Agner Fog]].
 
-Conclusion: because caller as well as callee preserve FpCsr and MxCsr
-__jump_fcontext__ is not required to preserve FpCsr and MxCsr.
-
 
 [heading x86_64]
 
@@ -123,9 +125,6 @@
 [footnote [@http://http://msdn.microsoft.com/en-us/library/ms235300.aspx MSDN article
 'FpCsr']].
 
-Conclusion: because caller as well as callee preserve FpCsr and MxCsr
-__jump_fcontext__ are not required to preserve FpCsr and MxCsr.
-
 "The MMX and floating-point stack registers (MM0-MM7/ST0-ST7) are preserved across
 context switches. There is no explicit calling convention for these registers."
 [footnote [@http://msdn.microsoft.com/en-us/library/a32tsf7t%28VS.80%29.aspx MSDN article
@@ -135,9 +134,6 @@
 [footnote [@boost:/libs/context/doc/pdf/calling-conventions.pdf 'Calling Conventions',
 Agner Fog]].
 
-Conclusion: because no explicit calling convetion exists and MS compiler do not use those
-registers, they will not be preserved.
-
 "XMM6-XMM15 must be preserved"
 [footnote [@http://msdn.microsoft.com/en-us/library/9z1stfyw%28v=vs.100%29.aspx MSDN
 article 'Register Usage']]
@@ -151,9 +147,6 @@
 [footnote [@boost:/libs/context/doc/pdf/x86_64-sysv.pdf SysV ABI AMD64 Architecture
 Processor Supplement Draft Version 0.99.4, 3.2.1]].
 
-Conclusion: FpCsr and MxCsr must be preserved.
-
-
 [endsect]
 
 

Modified: trunk/libs/context/performance/performance.cpp
==============================================================================
--- trunk/libs/context/performance/performance.cpp	(original)
+++ trunk/libs/context/performance/performance.cpp	2012-04-24 08:36:55 EDT (Tue, 24 Apr 2012)
@@ -13,6 +13,7 @@
 
 #include <boost/assert.hpp>
 #include <boost/bind.hpp>
+#include <boost/function.hpp>
 #include <boost/config.hpp>
 #include <boost/context/all.hpp>
 #include <boost/preprocessor/repetition/repeat_from_to.hpp>
@@ -30,6 +31,9 @@
 
 bool preserve_fpu = true;
 
+#define CALL_FUNCTION(z,n,unused) \
+    fn();
+
 #define CALL_UCONTEXT(z,n,unused) \
     ::swapcontext( & ucm, & uc);
 
@@ -41,6 +45,9 @@
 #endif
 ctx::fcontext_t fc, fcm;
 
+static void f3()
+{ }
+
 #ifndef BOOST_WINDOWS
 static void f2()
 {
@@ -55,6 +62,24 @@
         ctx::jump_fcontext( & fc, & fcm, 7, preserve_fpu);
 }
 
+cycle_t test_function( cycle_t ov)
+{
+    boost::function< void() > fn( boost::bind( f3) );
+    // cache warum-up
+BOOST_PP_REPEAT_FROM_TO( 0, BOOST_PP_LIMIT_MAG, CALL_FUNCTION, ~)
+
+    cycle_t start( cycles() );
+BOOST_PP_REPEAT_FROM_TO( 0, BOOST_PP_LIMIT_MAG, CALL_FUNCTION, ~)
+    cycle_t total( cycles() - start);
+
+    // we have two jumps and two measuremt-overheads
+    total -= ov; // overhead of measurement
+    total /= BOOST_PP_LIMIT_MAG; // per call
+    total /= 2; // 2x jump_to c1->c2 && c2->c1
+
+    return total;
+}
+
 #ifndef BOOST_WINDOWS
 cycle_t test_ucontext( cycle_t ov)
 {
@@ -142,6 +167,8 @@
         res = test_ucontext( ov);
         std::cout << "ucontext: average of " << res << " cycles per switch" << std::endl;
 #endif
+        res = test_function( ov);
+        std::cout << "boost::function: average of " << res << " cycles per switch" << std::endl;
 
         return EXIT_SUCCESS;
     }