Skip to content

McStas 3.0beta technology preview release notes

Jump to bottom
Peter Willendrup edited this page Mar 24, 2020 · 32 revisions

McStas and McXtrace logos

This page gives a status of the McStas 3.0beta technology preview to be released end of February 2020. The beta is very much result of a concerted effort done end of January 2020 by the whole McStas-McXtrace team and observers from RAMP.

Thanks:

  • Thanks to all members of the joint McStas-McXtrace team, you guys ROCK!
  • Thanks to Guido Juckeland (HZDR,DE) and Sebastian Alfthan (CSC,FI) who were behind the GPU Hackathons we participated in
  • Thanks to our NVIDIA mentors Vishal Metha, Christian Hundt and Alexey Romanenko

Installation:

  • As this is NOT a production release, we will only offer installation packages in the form of "manually installable" packages, see http://download.mcstas.org/mcstas-3.0beta/
  • Installing mcstas-3.0beta on Unix-oriented systems may replace your "system default McStas". If this was not intended, please run e.g.
    • sudo /usr/share/mcstas/2.6/bin/postinst set_mccode_default to get v2.6 back as default.

Bugfixes etc:

  • As we move along with the development, you may find patched versions of components in the Updates folder found next to the distribution binaries at http://download.mcstas.org/mcstas-3.0beta/
  • Unfortunately, a last-minute change to the PowderN component before mcstas-3.0beta was built introduced an error which make it produce 0 scattered intensity. Please find an improved version in http://download.mcstas.org/mcstas-3.0beta/Updates . Further bugfixes to the component are expected since the component produces scattering but has certain edge-case issues on GPU, especially at high ncount rates.
  1. New code-generation scheme based on functions instead of #defines, which brings
    • Much improved compilation-times, the code is better suited for modern compilers
    • In most cases a speed-up of order 20% on CPU
    • The neutron _particle is now represented by a struct
    • The component types and instances are also represented by structs
    • In the generic TRACE function of a given component type, the _comp var is short-hand for "whatever the component instance is"
    • New instrument section of USERVARS %{ double example_flag; %} which enriches the _particle struct
    • In component DECLARE blocks, assignments can no longer be done and all declarations must be listed independently, i.e double a; is OK, double a,b; is not. Variables in this scope are automatically so-called "OUTPUT PARAMETERS" (we may deprecate that keyword completely for the official McStas 3.0 release)
    • Components no longer support DEFINITION PARAMETERS, instead the SETTING PARAMETERS must be used, which now includes a vector and string type supplementing the (default) double/MCNUM and int types.
    • New macros have been added for
      • INSTRUMENT_GETPAR(parameter_name)
      • COMP_GETPAR(component, parameter_name) which is similar to the legacy MC_GETPAR
      • MC_GETPAR3(component_class, component_name,parameter_name)
      • FIXME: document these fully
    • Further, the new cogen implements support for Nvidia GPU's, for details see point 2 below.
    • Status on CPU (20200220) is that 134 out of 176 instruments from the mcstas-3.0 branch compile for CPU - and most run as expected.
  2. Limited, experimental support for OpenACC acceleration on NVIDIA GPU's
    • #pragma driven, inserted by the code-generation, but also implemented in (many, not all) libs and comps
    • Status on GPU (20200220) is that 91 out of 176 instruments from the mcstas-3.0 branch compile for GPU. These 91 instruments include in total 73 out of 213 components. Out of the functional instruments, around 45 produce meaningful data at this point.
    • Speedups measured using top-notch NVIDIA V100 datacenter cards are in the range of 10-600 with respect to a single-core CPU, see below figure which was generated for an "ideally" parallel instrument. Consult e.g. the graphical output page for the February 20th test run.
  3. Platform support / compiler configuration
    • Required compiler for GPU/OpenACC: PGI 19.4 or newer. Community edition works fine
    • Required GPU hardware: NVIDIA Tesla card + configured driver
    • Windows: At this point unsupported for GPU/OpenACC since the managed memory mode is not available for this platform. (Very simple instruments without monitors may work, at this point untested). The multi-threading should work with pgcc -ta:multicore -DOPENACC.
    • macOS: Only CPU-thread parallelisation is supported, since NVIDIA Cards at this point are unsupported on recent macOS versions. Parallelises similarly to MPI, but slightly slower.
    • Linux: Full acceleration support with GPU, and with multicore.
    • Install the compiler and put it on your system PATH. Install and configure Nvidia drivers for your card.
    • We hope that GCC will better support OpenACC in the near future.
  4. Tool support
    • On Linux and macOS mcrun is preconfigured so that mcrun -c --openacc compiles with:
    • Linux: pgcc -ta:tesla,managed,deepcopy -DOPENACC
    • macOS: pgcc -ta:multicore -DOPENACC
    • For both of the above, adding -Minfo:accel will output verbose information on parallelisation
    • In mcgui, the mcrun --openacc configuration can be selected via the preferences
  5. Interoperability with McStas 2.6
    • Rudimentary support for MCPL event interchange has been added through a set of MCPL_input_GPU and MCPL_output_GPU components.
  6. Known limitations
    • As mentioned above, not all components/instruments are ported to the GPU technology yet
    • We have not fully decided if our newly implemented random number algorithm is sufficiently robust/stable.
    • Monitor_nD user vars can not currently access the neutron USERVARS, but a solution is in the pipe.
    • Not all features of all components correspond to those from McStas 2.6, partly because not all modifications have been ported from the 2.6 tree to the 3.0 tree.
    • Especially the sample components may
      • Create GPU-side segmentation faults etc. with big datasets / large statistic, this is a targeted area of development before the actual release of 3.0. :-)
      • Give simulation results that are systematically 'off' with respect to CPU results.
    • mcdisplay-oriented --trace mode is NOT supported on GPU as this requires use of the file-descriptor stdout which is not accessible on the GPU. Recompile and run on CPU in this case.
    • On Linux with target multicore, one should add -Mnollvm, i.e. configure OACCFLAGS in mccode_config to be -ta:multicore -Mnollvm -DOPENACC
Clone this wiki locally