Jump to content

Philipp A Hartmann

  • Content Count

  • Joined

  • Last visited

  • Days Won


Everything posted by Philipp A Hartmann

  1. In general: Yes, this implementation is sufficient to implement a TLM2 extension. Still, there is a more reliable pattern to implement the copy_from and clone methods by using the copy constructor and assignment operator of your extension type (which you may need to implement in some cases anyway and will be provided for free in your particular example): class reg_extension : public tlm::tlm_extension<reg_extension> { public: tlm::tlm_extension_base* clone() const { return new reg_extension(*this); } // use copy constructor void copy_from(tlm::tlm_extension_base const & that ) { *this = static_cast<const reg_extension&>(that); } // use assignment operator // ... }; This pattern works very well for all Copyable and CopyAssignable classes without having to enumerate the members in clone and copy_from. Hope that helps, Philipp
  2. Hi S., The biggest issue in your code is that you set up the tracing before you initialize the "payload_data" pointer. Nothing good can come out of this. Secondly, sc_trace requires stable memory locations to trace, as it stores a pointer internally. So you can't easily reallocate the extension over and over again (or even later than the sc_trace call). Instead, I would suggest to reuse the extension across transactions and move it to a plain member in the class (changes added below): Hope that helps, Philipp
  3. Side note: You can't use virtual inheritance with SystemC modules. See the following thread for a related discussion: Your example doesn't require this, so it should be fine to simply drop the virtual keyword in the sc_module inheritance. Hope that helps, Philipp
  4. Maybe we can just reject "--enable-pthreads" on this platform explicitly. Generally speaking, you usually want to avoid pthreads, also on Linux, unless you have specific requirements. Greetings from Duisburg, Philipp
  5. Can you explain, why you try to use Pthreads on the Msys/MinGW platform? This environment doesn't come with a Pthreads library by default, you would need to install a separate one. And Pthreads will certainly be less efficient than the WinFiber-based process implementation. Thanks, Philipp
  6. Thanks for the reproducer, Sumit! This is indeed a regression compared to SystemC 2.3.0. I'll forward the issue to the SystemC Language Working Group. Greetings from Duisburg, Philipp
  7. The instance specific extension is tied to a specific accessor object (usually a member in a module, e.g. an interconnect). So each accessor can have its own extension object and is responsible for cleaning it up later. For regular TLM extensions, yes, there can only be one extension object of each type in a transaction payload.
  8. You can have a look at the "nb2b_adapter" test in the SystemC regression test suite, located at tests/tlm/nb2b_adapter/nb2b_adapter.cpp. In this example, you find the usage of an instance-specific extension for exactly that purpose: struct route_extension: tlm_utils::instance_specific_extension<route_extension> { int id; }; // ... virtual tlm::tlm_sync_enum nb_transport_fw( int id, tlm::tlm_generic_payload& trans, tlm::tlm_phase& phase, sc_time& delay ) { route_extension* ext = 0; if (phase == tlm::BEGIN_REQ) { ext = new route_extension; ext->id = id; // <-- store ID in the extension accessor(trans).set_extension(ext); } // ... virtual tlm::tlm_sync_enum nb_transport_bw( int id, tlm::tlm_generic_payload& trans, tlm::tlm_phase& phase, sc_time& delay ) { route_extension* ext = 0; accessor(trans).get_extension(ext); sc_assert(ext); tlm::tlm_sync_enum status; status = targ_socket[ ext->id ]->nb_transport_bw( trans, phase, delay ); // use ID from the extension Hope that helps, Philipp
  9. In line 33 of your producer.h, it seems you accidentally try to copy an interface (something derived from sc_interface). In order to give more detailed feedback, please share the code around the offending line. Hope that helps, Philipp
  10. Hi Sumit, even if the sockets are optional, you still need to bind at least one socket callback, right? Can you please check, that you actually do this? Greetings from Duisburg, Philipp
  11. On modern GCC versions (starting with 5.x), you the binary interface changed between C++03 and C++11 mode. You need to build your SystemC library with the same compiler (settings) to get the C++11 version of the functions. Hope that helps, Philipp
  12. Hi Jonas, for now, this is a known leak (as stacks from statically created processes are currently deliberately not deleted) and allocations from sc_core::sc_cor_pkg_*::create should be suppressed. (Dynamically created threads are not affected, so it is not a "real" leak). Greetings from Duisburg, Philipp
  13. Hi Joshua, This could then indeed be a bug. I would expect that (sc_fix)(uint)-1 is a positive value? You say, that this is not the case? Thanks, Philipp
  14. Hi Joshua, thanks for the report also from my side. I think, a better fix for the regression tests in question would be to use one of the explicit conversion functions in sc_fxval(_fast), see e.g. IEEE 1666-2011, b[i] = (ushort)(b[i-1].to_ushort() * i * -1); The other question would be: Is the undefined behavior actually triggered? The quoted code looks like the sc_fxval values should indeed be positive as they have been assigned from a positive (ushort) value. The resulting double value then definitely holds a number that fits back into the unsigned short. Can you elaborate? Greetings from Duisburg, Philipp
  15. Hi Jean-Claude, in SystemC, there is currently no way to identify whether an event has been triggered during/before the current evaluation phase in the simulation. See e.g. http://forums.accellera.org/topic/4925-/(and other threads) for earlier discussions. Btw: I find allEvents a confusing name for an sc_event_or_list. To me, wait(allEvents) sounds like waiting for all events to be triggered, whereas an or-list will trigger if any event in the list is notified. Greetings from Duisburg, Philipp
  16. I'm afraid, we need a more detailed analysis / debugging from someone, who has access to such a platform. Otherwise, we can just disable QuickThreads on Cygwin again and select Pthreads consistently instead. @Ralph: Can you try hunting this down in a debugger? Thanks, Philipp
  17. Hi Ralph, I agree, that the change in 2.3.1 is in violation of IEEE 1666-2011. And this needs to be fixed for sure. For 2.3.2, it's currently most likely that we just restore the original behavior. On the question of how to simplify the usage of sc_bitref in boolean contexts, we might also look into "explicit conversion to bool", see for example http://stackoverflow.com/questions/6242768/is-the-safe-bool-idiom-obsolete-in-c11. This needs further exploration and discussions in order to avoid escapes as in 2.3.1. This could be added to either sc_bitref<sc_bv_base> (sc_bitref<sc_lv_base>) (sc_logic) For the first option, I see the least risk of breaking the 1666-2011 compliance. The other two options could have more unwanted side-effects. The IEEE 1666-2011 sc_bit_ref template argument looks like an oversight to me. Will add this to the errata for review in IEEE later. Thanks and Greetings from Duisburg, Philipp
  18. Hi csr18, Thanks for reaching out on this topic. Allowing a conversion from sc_bitref to bool has been added intentionally to 2.3.1 to address the use case discussed at http://forums.accellera.org/topic/1392-/. With "this", Ralph is referring to the issue of now failing comparisons (and other operators) with character and some other literals, which now silently changed their meaning. This has been an unfortunate oversight in SystemC 2.3.1: sc_dt::sc_bv<1> bv = "1"; sc_assert( bv[0] ); // works now, great! sc_assert( bv[0] != '0' ); // broken! We're currently investigating options to restore the behavior required by 1666-2011. Have you seen other misbehaviors? Thanks and Greetings from Duisburg, Philipp
  19. You need to add the preprocessor symbol SC_INCLUDE_DYNAMIC_PROCESSES to your build setup to enable sc_bind in SystemC, e.g. on the compilar command-line: -DSC_INCLUDE_DYNAMIC_PROCESSES. hth, Philipp
  20. You can use a custom "creator" to initialize elements of a vector with custom constructor parameters - here the inner vector. Something like this (assuming you have lambda support available): auto element_creator = [](const char* nm, size_t) // optional, depending on the "real" value type { return new sca_module(nm); }; size_t inner_size = 42; // adjust for your needs, could also be a vector of sizes element.init( outer_size, [&](const char* nm, size_t) { return new sc_vector<sca_module>( nm, inner_size, element_creator ); } ); If you don't have lambdas in your environment, you need to put the functionality in a custom function, e.g. static sc_vector<sca_module>* element_vector_creator(size_t size, const char* name, size_t) { return new sc_vector<sca_module(name, size); } // using sc_bind to pass in the size - placeholders needed for actual call element.init( outer_size, sc_bind(element_vector_creator, inner_size, sc_unnamed::_1, sc_unnamed::_2) ); Hope that helps, Philipp
  21. Which version of SystemC/TLM are you using? I'm afraid, there is a known issue in the 2.3.1 circular_buffer implementation. When the fifo is destroyed while there are still elements in it, the loop over the "to-be-cleared" items does not start at the right index. Instead, you need the following change: circular_buffer<T>::clear() { for( int i=0; i < used(); i++ ) { - buf_clear( m_buf, i ); + buf_clear( m_buf, (m_ri + i) % m_size ); } // ... In 2.3, there was another indexing bug related to resize, see http://forums.accellera.org/topic/1443-/. Hope that helps, Philipp
  22. Hi VanTeo, to me, the following line looks suspicious: You said, SCI_func inherits from sc_module, but you pass in a string constant to its constructor from the derived SCI class. This will disrupt the SystemC object hierarchy, as the corresponding sc_module_name argument will be destroyed too early. If you want to use the SC_CTOR macro, you should add a protected default constructor to the SCI_func base class, allowing to take the name from the current top of the naming stack. See IEEE 1666-2011, section 5.3.3 for more details about passing sc_module_name arguments through the inheritance hierarchy. Hope that helps, Philipp
  23. There is a recommendation in IEEE 1666-2011 (8.3.1) on the format of the message type: The message types used in the SystemC proof-of-concept implementation currently do not follow this scheme in many cases, as the reporting mechanism predates the definition of the IEEE standard for SystemC. In order to maintain backwards-compatibility, the existing message type strings have not been adapted. There needs to be a distinction between different message types in order to allow custom actions, severities etc. for individual message categories. In order to adapt the SystemC kernel to the above scheme, we would essentially need the same "IDs" as we have right now, and would just add some "/../..." prefix to them. But this would require changes to all existing models working with the existing IDs today. Hope that helps, Philipp
  24. To me, an Internet forum doesn't look like the best place to ask for legal advice. You might want to consult a lawyer instead. Greetings from Duisburg, Philipp
  • Create New...