Integration of transactional memory support into a data-flow extension of OpenMP

[ismail]

On Fri, 2011-05-06 at 17:51 +0200, Patrick Marlier wrote:
>>>Dear Patrick,
>>> I am currently searching for a pipeline in a TM-Benchmarks but I can
>>>not overlap the concept of transactions with pipeline in current TM
>>>benchmarks. I think I will write a new one.

> Dear Ismail,
> 
> On Fri, 6 May 2011, ismail wrote:
> >> TinySTM 1.0 uses implicit transaction descriptor by default (same as
> >> 0.9.9) but you can compile it with EXPLICIT_TX_PARAMETER (Makefile) and
> >> then the transaction descriptor is explicit.
> >> But functions are the same from 0.9.9 to 1.0.0 so I don't really see the
> >> problem.
> > --- Yes  . thanks  for update , I remembered . Also I wrote the version wrong it is "0.9.0b1 release".  There should be a parameter like EXPLICIT_TX_PARAMETER in trans-mem. Did you need  a kind of parameter
> > --- while compiling the trans-mem branch ? For example to use trans-mem to transactify LeeTM (I am especially talking about the MEMSET and MEMCPY funcs. of trans-mem do not have transaction identifier while accessing to the mem.).
> > --- I am currently using the explicit function calls when I need transaction identifier as the following.
> 
> Humm... libitm is always using implicit transaction descriptor and there 
> is no such flag as in TinySTM.
>>>I am clear with tx-iden. and tx-desc. Do you think that not having
explicit tx_desc. is a handicap while converting conventional locking
systems to the TM model? (I am having some trouble while dealing with
benchmarks)

> 
> Don't confuse transaction descriptor and transaction identifier.
> I think it is better to use always implicit transaction descriptor it 
> avoids the burden to manage it.
> 

> > --- I tried to get transaction handler which does not ---have explicit
> > ---descriptor where another function needs transaction ---descriptor as
> > ---a parameter.
> 
> Hummm... which one has explicit tx descriptor?
> All ITM functions are using implicit transaction descriptor as far as I 
> remember, otherwise it is a bit odd.
>>> I have mixed tx_descriptor because of IntelTM ABI's specification.
Yes there is just the following in the libitm:

>>>typedef uint64_t _ITM_transactionId_t;/* Transaction identifier */

>>>#define _ITM_noTransactionId 1 /* Id for non-transactional code. */

>>>extern _ITM_transactionId_t _ITM_getTransactionId(void)
>>>ITM_REGPARM;  

>>>BTW It may seem a general question but we are trying to figure out
>>>several issues related with following question . It would be great If
>>>you can share your opinion: 

>>>what are the main challenges right now w.r.t. transmem in C (we do
>>>not care much about C++ at this point), and what are his expectations
>>>and plans about exploitation of OpenMP with TM.
> 
>>>THANKS A LOT

>>> Ismail KURU
> Patrick.