Skip to content

Conversation

@martinschwinzerl
Copy link
Contributor

@martinschwinzerl martinschwinzerl commented Nov 25, 2020

Features / Main Improvements:

Note:
The minimum required CMake version has been bumped to version 3.11. This is due to the use of the 'FetchContent` feature to streamline the retrieval of external dependencies.

Due to a search & replace error, 'sixtrack::(anonymous)::st_size_t` is defined as an alias to `::st_size_t` which is not necessarily the same size as `unsigned long long`. This seems to cause the problems described in SixTrack#133
NOTE: cmake 3.11 introduces the FetchContent extension which is much better suited than
      the currently used ExternalProject facility (or manually downloading stuff via git)
- uses FetchContent to download OpenCL C headers, OpenCL C++ headers
  or the OpenCL icd loader if needed

- Adapts to the new OpenCL 3.x header structure
…inzerl/sixtracklib into feature/pr137_sixtracklib_070
- Deprecate the "use legacy c++ header" -> this is now discovery based,
  the user should not have to use this setting. It is kept for the sake
  of compability
- simplifies the the detection of the headers in case they are downloaded ->
  this no longer yields random results depending on the order but always
  a predictible outcome, sorted by most desireable headers first

- supports the "always downloads headers" option from Settings.cmake

- verifies that the OpenCL ICD loaded can successfully be built
- Fixes copy & paste error in 10th and 11th factorial
  switch-case statement

- Removes recursion which does not work ootb on OpenCL 1.2
- removes recursive limitations -> this fixes issues on OpenCL
- provide a "binary" like implementation with minimal temporary / local
  storage requirements optimised for GPUs
- adds helper functions for 1D grids to get global and local ids and sizes
- adds searching for largest elements in a (shared memory) array bound to
  wavefront / warps
- Decrease SIXTRL_CERRF_ALG680_MIN_POW_2H_N to 2.22507385850720e-307
  which is one order of magnitude above the smallest possible
  64 Bit double precision number. This reduces the chance of hitting
  the limit close to the boundaries between the compute regions

- consequently, the SIXTRL_CERRF_ALG680_MIN_TWO_H_VALUE has to be
  decreased to allow for SIXTRL_CERRF_ALG680_MIN_TWO_H_VALUE ^ ( N_S0 + N_S1 )
  to still be >= than SIXTRL_CERRF_ALG680_MIN_POW_2H_N
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant