* evaluation was too pessimistic for many cores, leading to designs
  significantly slower than possible
* fixed by using more aggressive synthesis options
* more aggressive pnr options had no measurable effect on WNS
* using new options in both evaluation and compose

* alternatives are now checked and built, in case the alternatives
  dimension is activated in DSE

* reduced required instances to one
* minimal number of threads tested is now twice the number of physical
  processors (accounts for hyper-threading)

Jaco Hofmann authored Jun 07, 2017 and Jens Korinth committed Jun 07, 2017

* removed all interrupt controllers for VC709
* replaced by 132 MSI-X interrupts (128 PE, 4 other)
* replaced driver code

* slurm executables are highly unreliable, both sbatch and squeue often
  fail with socket errors and similar
* need a retry mechanism for sbatch, too
* will retry a number of times now, with 10secs in between, before failing

* since we are by now using the jar to execute, there is no more need to
  change the current directory in the tapasco/itapasco scripts

* synthesis can be accelerated by dumping an EDIF netlist during IP
  evaluation and only using the netlist in the IP-XACT core
* no need to re-synthesize IP, especially useful in case of many
  instances
* replaced old links to original zip files by new zip that contains only
  the component.xml and the netlist

* retiming is not activated by default, but helps with timing closure
* OOC is the new default for all IP cores in the top-level; since we are
  only running a single time we cannot benefit from the feature

* SLURM should ignore resource constraints
* tasks now start correctly in SLURM mode

* better control to restrict parallelism (~ memory consumption)
* global parameter --maxThreads [NUM] controls general.maxThreads in Vivado
* applies to both synthesis and implementation
* removed superfluous implicit in Composer

* stopped working due to rename of Vivado project

* only works for Bash shell; warning for others

* Jobs are executed sequentially by default, but there are cases in
  which it is useful to execute them in parallel instead (especially in
  combination with SLURM)
* implemented global --parallel switch to activate this mode
* added to all representations (parser, Json, prettyPrint)

* just had a number of runs killed, because their memory usage exceeded
  8 GiB (Xilinx claims 6 GiB, but apparently for PnR only)
* moved to generous 10 GiB

* fixed; existence of cores is now checked as in Compose