descript.txt

VPR布局布线源码

Version 5.00, March 1, 2008

- Andy Ye's modification to include support for the uni-directional and 
  single-drive routing architectures.

- Andy Ye's modification to include support for single-tile based
  architectures.

- Mark Wei Fang's modifications for mux balancing

- Jason Luu and Ted Campbell's XML architecture format

- Jason Luu and Ted Campbell's heterogeneous structures


Version 4.30, March 25, 2000

- Sandy Marquardt's timing-driven placement enhancements are in this
  version.  Previous versions of VPR had only routability-driven placement.

  I cleaned up various minor things:  removed an unused #define NEVER from 
  main.c, set the date in the title bar, and fixed a couple of mistakes in
  the command line processor's checking for invalid combinations of 
  options.  I regression tested this code fairly thoroughly against the
  last release.

  This version was released on the web.


Version 4.22, Jan. 26, 1999

- Changes "pr.h" to "vpr_types.h" and "ext.h" to "globals.h".
  The old names were around just for historical reasons; the new filenames
  are more clear.

- Changed the input netlist format slightly so that subblock output pins
  can be hooked directly to unused (open) CLB output pins.  When the netlist
  specifies this, it means that a CLB OPIN of that class must be used by
  that subblock, since the subblock is directly connected to an OPIN.  This
  correctly models Altera LABs, for example.  The router has been altered so
  that it ensures any "locally used CLB OPINs" specified by the netlist are
  properly reserved (i.e. are not used by other connections).

  This change lets me correctly model both types of CLB output pin logical
  equivalence.  One equivalence type corresponds to there being muxes between 
  a set of subblocks in a logic block and a set of CLB output pins.  In this
  case, if some of the subblock outputs don't have to go outside the CLB,
  the other subblocks can use more than one CLB OPIN if it helps routability
  (for routing a high-fanout net for example).  There is another type of 
  logically equivalent outputs, however.  If the subblock outputs are hooked
  directly to the CLB outputs, but a set of subblocks are all identical and
  muxes let you make any connection you want to their inputs, then all
  the CLB OPINs connected to these subblocks are logically equivalent.  In this
  case, however, a subblock whose output is only used locally (never goes 
  outside the CLB) still consumes the OPIN connected to it, so other 
  subblocks can not use this OPIN.  This is the case in Altera LABS, for
  example.  The changes I've made let the netlist specify either type of 
  logical equivalence (or even a mix of both types).

- Made the delay of a routing switch 1 second when doing a global routing
  (old value was 0).  A routing resource must have a delay > 0 for the 
  timing-driven router to work properly.


Version 4.21, Nov. 19, 1998

- Freed the button array in close_graphics (graphics.c) -- prevents a small
  memory leak that would otherwise occur if VPR started and shut down the
  graphics window several times.  Noticed by Paul Leventis.

- Changed place_cost_type from int to an enumerated type (just for
  cleanliness).

- Changed read_arch.c so purify doesn't complain about me printing
  out unset values in print_arch (the values were uninitialized only
  when they weren't relevant for the architecture used, so this is 
  just a cosmetic cleanup).


Version 4.20, Sept. 10, 1998

- Simplified equation used to count transistors in pass transistor muxes --
  Sandy noticed that the "approximate" equation I listed was actually exact,
  and simpler than the summation used in VPR.

- Changed all "class" variables to iclass so the code compiles with a C++ 
  compiler.  Paul Leventis wanted this.

- Changed draw.c to correctly draw architectures in which connected x-directed 
  (or y-directed) wire segments can overlap some in a channel.  Needed by 
  Steve Trimberger for some of the architectures he's looking at.


Version 4.19, July 7, 1998

- Put an fabs around one of the tests for perturbing an input switch
  pattern (in rr_graph.c).  The lack of an fabs meant that I was
  perturbing the switch patterns half the time when I didn't need to.
  That could have hurt routability a little bit, although probably not 
  a whole lot.

- Made a CLB output pin driving a global net only a warning when the 
  CLB pin isn't global.  This allows locally generated clocks to be
  generated by CLBs, and put on global resources.  Paul Leventis wanted
  this so he could run one of his new circuits (des) without changing the
  architecture file around.


Version 4.18, June 1, 1998

- Changed the rr_graph generator so that it perturbs the input pin switch 
  pattern whenever the Fc_input is a perfect multiple of Fc_output (so that
  the output and input switch patterns are prevented from lining up perfectly).

- Added -timing_analysis_only_with_net_delay <float> parameter.  This let's
  you just use VPR's timing analyzer from the command line, with the delay
  of each net set to a constant and other delays taken from the architecture
  file.  Sandy and Catherine Wong both wanted this.


Version 4.17, April 30, 1998

- Fixed a minor problem in the rr_graph generator -- the perturbed IPIN 
  connection block routine could create two switches from the last track 
  in a channel to one IPIN under certain (weird) conditions (basically very
  low track counts).  Fixed it by changing a min to a mod.

- Changed the information printed out about the critical path slightly -- now
  counts of normal nets and of global nets (i.e. the clock) on the critical
  path are computed and printed out separately.

- Cleaned up the code some by making chan_width_io, chan_x_dist and chan_y_dist
  part of the chan_width_dist structure and passing that structure around to
  the places that need it, rather than having them be global variables.
  These variables weren't used widely enough to justify their being global.

- Cleaned up the code a bit by moving a few functions that didn't really
  belong in place.c into place_and_route.c.  This file now contains the 
  overall control routines that start the placer and router, etc.  Place.c
  now just has the placer itself in it.


Version 4.16, April 20, 1998

- Fixed bug where architectures with both pass transistors and buffered
  segments, and paths from buffered to unbuffered segments and back again,
  could go into an infinite loop in the router.  Problem:  the cost is not
  monotonic in my directed-timing driven router under certain cases.  The 
  cost of an unbuffered node can be less after you go from it to a buffered
  segment and back, due to the reduction in upstream resistance.  To fix it,
  I remember the backward_path_cost back to the start of the routing 
  connection, and never allow a node to be re-expanded if it would make the
  backward_path_cost go down.  Now loops can never occur.


Version 4.15, April 15, 1998

- Added extra info to the timing graph so I can tell what kind of resource
  each timing node is.  This let's me print out and display the critical
  path.  Added code to print out the critical path and graphically display
  it.


Version 4.14, April 9, 1998

- Added the optional keyword "global" to CLB inpin statements in the .arch
  file.  This let's pins used only for global signals (e.g. clocks) be
  flagged as special pins that shouldn't connect into the general purpose
  routing.  This stops muxes, etc. from being built for them, so the area
  numbers are more accurate.  Also, it means they don't disrupt the input
  pin switch pattern -- since input pin switches are evenly distributed
  across the tracks, for low Fc's hooking the clock ipins into the normal
  routing meant some tracks connected to the clock pin and not to any of the
  normal ipins.

  The netlist checker has been beefed up to check that global signals only
  connect to global CLB pins and so on.

- Changed the switch pattern used for CLB input pins when Fc_input = Fc_output.
  The old VPR would make perfectly regular switch patterns, so when Fc_input
  = Fc_output, an output could only talk to certain inputs if Fc was low.
  This would result in an FPGA with very poor routeability (at least if
  the switch box was subset (planar) -- the Wilton switch box would probably
  work OK even with this IPIN switch pattern since it let's every track get
  to every other one).
   
  The change checks for the case Fc_input = Fc_output and perturbs the switch
  pattern for CLB input pins to make it different than the output pin pattern.
  Hence an output can talk to more input pins, and the FPGA is more 
  routable.  The perturbation is as small as I could think to make it, since
  I still wanted the switches nicely distributed over the tracks and logically
  equivalent inputs on the same side should try to hit different tracks.


Version 4.13, March 26, 1998

- Changed the router defaults (acc_fac, first_iter_pres_fac, pres_fac_mult) 
  to the values my experimentation determined were the best (from experiments
  on a buffered, unit-length wire architecture).  Also, the timing-driven 
  router now uses acc_fac = 0 for the first router iteration.

- Changed the area model slightly.  It now assumes that the buffers from tracks
  to ipin_cblocks are 4x minimum width, as are the buffers from the output 
  muxes to logic block input pins.

- Changed rr_graph_timing_params.c so it shares the pull-up, pull-down part
  of tri-state buffer switches in the routing.  This means the input capacitance
  of a buffer is added into a node's capacitance only once at a given (i,j)
  location for that node, since other switches at the same spot using buffers
  will share that buffer.


Version 4.12, Jan. 26, 1998

- Added code to allow several different base cost types; some appropriate
  for area-based routing, some appropriate for timing-driven routing.

- Wrote the timing-driven router.  It uses an A-star directed search algorithm
  and keeps track of the Elmore delay of each node in the partial route
  tree as it constructs it.  Everything is done except the dynamic base cost
  changing net by net.
  
- Split route.c into route_common.c, route_timing.c and route_breadth_first.c.

- Code cleanup.  Removed the net_block_pin_num array and made that data a
  part of the net structure (member blk_pin).  Changed the name of net.pins
  to net.blocks.  Removed the net.tempcost and net.ncost members (since they
  were only used by the placer) and added two static arrays to place.c instead.

- Changed rr_base_cost to a member of the rr_indexed_data structure.  Anything
  that is the same for all segments of a given type can be stored in this 
  structure to save data.  Put various timing values and such for quick 
  computation of expected costs to a target here for use by the timing-driven
  router.

- Changed segment_stats.c to use the length information stored in rr_indexed_
  data, so it doesn't need local static stuff anymore.

- Fixed a minor bug in read_arch.c -- I wasn't setting the loneline parameter
  properly when setting up an FPGA architecture for global routing only. 
  Thanks to Russ Tessier at MIT for finding this.


Version 4.11, Dec. 4, 1997

- Stopped storing the cost of an rr_node in the router, and now store only 
  acc_cost and pres_cost.  This enables dynamic costing of resources by only
  changing the rr_base_cost array.  The cost of a node is then computed as each
  node is expanded during routing.  This slows the router down by 7%.


Version 4.10, Dec. 3, 1997