WIP: Enhancing VPR lookahead map #1351

acomodi · 2020-06-08T16:58:51Z

Signed-off-by: Alessandro Comodi [email protected]

Description

This PR is an initial Draft with the effort of getting the lookahead map work well with the Series 7 architectures.

Related Issue

The issue tracking what needs to be done and all suggestions is here: #1325

Motivation and Context

Currently, the SymbiFlow VTR fork implements the connection box lookahead map which has been tested to work well with Series 7 devices.
The idea is to get the lookahead map code produce similar results, so that eventually, there can be only one lookahead code being used.

In case the two different approaches cannot be merged, an alternative would be to have a common utility library for the different lookaheads.

How Has This Been Tested?

Types of changes

Bug fix (change which fixes an issue)
New feature (change which adds functionality)
Breaking change (fix or feature that would cause existing functionality to change)

Checklist:

My change requires a change to the documentation
I have updated the documentation accordingly
I have added tests to cover my changes
All new and existing tests passed

Signed-off-by: Alessandro Comodi <[email protected]>

Signed-off-by: Keith Rothman <[email protected]>

- New sampling method for connection box lookahead map. - New spiral hole filling algorithm - Use a lightweight version of PQ_Entry (PQ_Entry_Lite) - Use maps in run_dijkstra() - Move context of router_lookahead_map_utils inside of util namespace. Signed-off-by: Dusty DeWeese <[email protected]> Signed-off-by: Keith Rothman <[email protected]>

Signed-off-by: Dusty DeWeese <[email protected]>

Expand base_cost too. The delay and base cost matricies are independent, so should be expanded and filled independently. Signed-off-by: Keith Rothman <[email protected]> Signed-off-by: Dusty DeWeese <[email protected]>

Signed-off-by: Dusty DeWeese <[email protected]>

Signed-off-by: Dustin DeWeese <[email protected]>

Signed-off-by: Keith Rothman <[email protected]>

Signed-off-by: Dusty DeWeese <[email protected]> Signed-off-by: Keith Rothman <[email protected]>

Before on the A200T: > Computing connection box lookahead map took 31769.29 seconds (max_rss 45671.6 MiB, delta_rss +14486.3 MiB) After on the A200T: > Computing connection box lookahead map took 16791.79 seconds (max_rss 40113.6 MiB, delta_rss +8895.9 MiB) Signed-off-by: Keith Rothman <[email protected]>

Signed-off-by: Keith Rothman <[email protected]>

Signed-off-by: Alessandro Comodi <[email protected]>

This reverts commit 8c4519c8723b7adf2c4571700202f06690d04278.

Signed-off-by: Alessandro Comodi <[email protected]>

acomodi · 2020-06-15T14:52:02Z

@litghost, @kmurray, @HackerFoo FYI.

I have completed the merge of the connection box lookahead map and the upstream lookahead map.

From the connection box lookahead, the following has been taken:

Sampling method: this has been extracted in a separate file
Paths expansion and lookahead filling
Lookahead map format (different capnp schema)

From the upstream lookahead map:

OPIN --> CHAN information
CHAN --> IPIN information
Delta locations calculations
Map addressability: e.g. cost_map[CHAN][SEG_ID][DELTA_X][DELTA_Y]

There are two assertions that needed to be skipped as they were triggered during the place delay matrix calculations:

vtr-verilog-to-routing/vpr/src/place/timing_place.h

Line 196 in 1b90651

    
           VTR_ASSERT_MSG(num_connections == 0 || num_level_before_leaves < num_connections, "Level before should have fewer nodes than connections (to ensure using the smallest binary tree)");

https://github.com/SymbiFlow/vtr-verilog-to-routing/blob/8980e46218542888fac879961b13aa7b0fba8432/vpr/src/route/connection_box_lookahead_map.cpp#L488

The last one is due to the fact that there are some non valid values for the expected cost which is set to infinite.

I also had to temporarily remove the lookahead generation test under vpr/test which still needs to be adapted to the new lookahead generation mechanism.

With this implementation, the routing step has the following results:

238.30 seconds
18.796 ns (CPD)
Works on HW

With the connection box lookahead map:

146.26 seconds
17.381 ns (CPD)
Works on HW

The test design is the baselitex test from SymbiFlow

There is an increase in the CPD which I still need to analyse, as well as an increased run-time. Both of them are probably due to mispredictions, which I still need to analyse in details.

This is the route log out of this PR's implementation: route.log

litghost · 2020-06-15T16:19:51Z

There are two assertions that needed to be skipped as they were triggered during the place delay matrix calculations:

Both of these assertions should be in place, and violating them likely indicates a bug?

litghost · 2020-06-15T16:21:06Z

libs/libvtrcapnproto/CMakeLists.txt

@@ -19,6 +19,7 @@ endif()
 # Each schema used should appear here.
 set(CAPNP_DEFS
    place_delay_model.capnp
+    connection_map.capnp


connection_map.capnp shouldn't be required, because it isn't used?

Or if we are going to use this, it should move to map_lookahead.capnp

litghost · 2020-06-15T16:21:43Z

vpr/src/route/route_profiling.cpp

@@ -29,6 +30,10 @@ void time_on_fanout_analysis() {}

 void profiling_initialization(unsigned /*max_net_fanout*/) {}

+void conn_start() {}


The changes to route_profiling.cpp/.h should likely be in another PR.

litghost · 2020-06-15T16:22:26Z

vpr/src/route/route_timing.cpp

@@ -1096,6 +1096,8 @@ bool timing_driven_route_net(ConnectionRouter& router,
            conn_delay_budget.short_path_criticality = budgeting_inf.get_crit_short_path(net_id, target_pin);
        }

+        profiling::conn_start();


Should move out of this PR with the route_profiling.h/.cpp changes

litghost · 2020-06-15T16:23:41Z

vpr/test/test_map_lookahead_serdes.cpp

@@ -1,62 +0,0 @@
-#include "catch.hpp"


Why did this test get deleted?

This was based on the previous router_lookahead_map code, I would need to readapt the test with the code merged from the connection box lookahead map and the upstream lookahead map. It still is in the TODO list.

litghost · 2020-06-15T16:24:17Z

vpr/src/route/rr_graph_storage.h

@@ -48,7 +48,7 @@
 *       This field is valid only for IPINs and OPINs and should be ignored  *
 *       otherwise.                                                          */
 struct alignas(16) t_rr_node_data {
-    int8_t cost_index_ = -1;
+    int16_t cost_index_ = -1;


This should be put in another PR, to examine changing this independently of the other changes.

litghost · 2020-06-15T16:25:17Z

vpr/src/route/router_lookahead_map.h

@@ -1,48 +1,84 @@
-#pragma once
+#ifndef CONNECTION_BOX_LOOKAHEAD_H_
+#define CONNECTION_BOX_LOOKAHEAD_H_


Either leave this as #pragma once, or fix the #define

litghost · 2020-06-15T16:28:10Z

This PR likely combines too many changes at once. Things unrelated to the lookahead change should go in their own PR, and the initial change to the map lookahead should be extensions that do not change behavior significantly. For example, the CHAN to IPIN PR should be separated from the sampling / parallelism PR.

acomodi · 2020-06-15T16:35:18Z

@litghost All right, I'll split this into multiple ones in the following order to have an incremental merge of the changes:

Route profiling (this should be independent from the other PRs)
Sampling method: WIP: New Lookahead sampling method #1367
Lookahead generation
OPIN --> CHAN and CHAN --> IPIN

kmurray · 2020-06-16T13:35:07Z

I'll split this into multiple ones in the following order to have an incremental merge of the changes

I agree this seems like the right approach.

acomodi · 2020-08-04T16:27:00Z

Closing, superseded by #1449

probot-autolabeler bot added lang-cpp C/C++ code VPR VPR FPGA Placement & Routing Tool labels Jun 8, 2020

acomodi marked this pull request as draft June 8, 2020 16:59

acomodi force-pushed the enhance-lookahead-map branch from 4b3302a to 08656d1 Compare June 9, 2020 12:02

acomodi and others added 26 commits June 15, 2020 16:24

Add connection box lookahead.

8885050

Signed-off-by: Alessandro Comodi <[email protected]>

Integrate upstream uxsdcxx parser with connection box lookahead

4f49a02

Signed-off-by: Alessandro Comodi <[email protected]>

Complete connection box lookahead to be aware of connection box.

b71a487

Signed-off-by: Keith Rothman <[email protected]>

Mark file location function as static.

596c10f

Signed-off-by: Keith Rothman <[email protected]>

Add bounds checking when search grid.

15b676b

Signed-off-by: Keith Rothman <[email protected]>

add documentation

b351679

Signed-off-by: Dusty DeWeese <[email protected]>

un-shadow a variable

3c1a190

Signed-off-by: Dusty DeWeese <[email protected]>

remove method from set_cost_map

5d19549

Signed-off-by: Dusty DeWeese <[email protected]>

replace std::list with std::vector

39af170

Signed-off-by: Dusty DeWeese <[email protected]>

parallelize connection box lookahead

aec3721

Signed-off-by: Dusty DeWeese <[email protected]>

performance improvements

31b2f47

Signed-off-by: Dusty DeWeese <[email protected]>

suggested changes

9d1aaa9

Signed-off-by: Dusty DeWeese <[email protected]>

changes suggested in PR

9c9437b

Signed-off-by: Dusty DeWeese <[email protected]>

add penalty to Cost_Map::find_cost for points outside of the map

8abd79f

Signed-off-by: Dusty DeWeese <[email protected]>

store penalties

e5ddec7

Signed-off-by: Dusty DeWeese <[email protected]>

printf -> VTR_LOG

f00073f

Signed-off-by: Dusty DeWeese <[email protected]>

change sampling method

56a28ce

Signed-off-by: Dusty DeWeese <[email protected]>

remove unnecessary lookups

ff57a66

Signed-off-by: Dusty DeWeese <[email protected]>

remove SimpleCache, use vtr::hash_combine, add timing stats

afc36aa

Signed-off-by: Dusty DeWeese <[email protected]>

integrate lighost's patches

2446c03

Signed-off-by: Dusty DeWeese <[email protected]>

independently minimize delay and base cost (congestion)

d654bd4

Expand delay and base cost independently

24eeb80

Expand base_cost too. The delay and base cost matricies are independent, so should be expanded and filled independently. Signed-off-by: Keith Rothman <[email protected]> Signed-off-by: Dusty DeWeese <[email protected]>

site_pin_delay shouldn't affect congestion

578d084

Signed-off-by: Dusty DeWeese <[email protected]>

use histogram to choose sample sizes, set minimum penalty

bb9943e

Signed-off-by: Dustin DeWeese <[email protected]>

Use get_rr_cong_cost to compute base_cost's.

0eda42b

Signed-off-by: Keith Rothman <[email protected]>

HackerFoo and others added 10 commits June 15, 2020 16:24

Remove cached segment map from connection map lookahead file.

669d580

Signed-off-by: Dusty DeWeese <[email protected]> Signed-off-by: Keith Rothman <[email protected]>

Skip empty windows.

7144db3

Signed-off-by: Keith Rothman <[email protected]>

vpr: lookahead map enhanced with CHAN -> IPIN delay information

c900a78

Signed-off-by: Alessandro Comodi <[email protected]>

route: extract lookahead sampling to external file

feeab99

Signed-off-by: Alessandro Comodi <[email protected]>

route: avoid using connection box information

880e40b

Signed-off-by: Alessandro Comodi <[email protected]>

make format

b51ccec

Signed-off-by: Alessandro Comodi <[email protected]>

lookahead: move connection box to router lookahead map

52014e7

Signed-off-by: Alessandro Comodi <[email protected]>

Revert "Integrate upstream uxsdcxx parser with connection box lookahead"

309a96d

This reverts commit 8c4519c8723b7adf2c4571700202f06690d04278.

WIP: Fixes to have Series7 place-delay and lookahead being computed

1334f91

Signed-off-by: Alessandro Comodi <[email protected]>

acomodi force-pushed the enhance-lookahead-map branch from 08656d1 to 1334f91 Compare June 15, 2020 14:27

probot-autolabeler bot added build Build system lang-make CMake/Make code tests labels Jun 15, 2020

litghost reviewed Jun 15, 2020

View reviewed changes

acomodi mentioned this pull request Jun 15, 2020

route: profiling: add debugging prints. #1363

Merged

7 tasks

This was referenced Jun 16, 2020

WIP: New Lookahead sampling method #1367

Closed

New Lookahead map status SymbiFlow/vtr-verilog-to-routing#517

Open

acomodi closed this Aug 4, 2020

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

WIP: Enhancing VPR lookahead map #1351

WIP: Enhancing VPR lookahead map #1351

Uh oh!

acomodi commented Jun 8, 2020

Uh oh!

acomodi commented Jun 15, 2020 •

edited

Loading

Uh oh!

litghost commented Jun 15, 2020

Uh oh!

litghost Jun 15, 2020

Uh oh!

litghost Jun 15, 2020

Uh oh!

litghost Jun 15, 2020 •

edited

Loading

Uh oh!

litghost Jun 15, 2020

Uh oh!

litghost Jun 15, 2020

Uh oh!

acomodi Jun 15, 2020

Uh oh!

litghost Jun 15, 2020

Uh oh!

litghost Jun 15, 2020

Uh oh!

litghost commented Jun 15, 2020

Uh oh!

acomodi commented Jun 15, 2020 •

edited

Loading

Uh oh!

kmurray commented Jun 16, 2020

Uh oh!

acomodi commented Aug 4, 2020

Uh oh!

Uh oh!

		@@ -29,6 +30,10 @@ void time_on_fanout_analysis() {}

		void profiling_initialization(unsigned /max_net_fanout/) {}

		void conn_start() {}

WIP: Enhancing VPR lookahead map #1351

WIP: Enhancing VPR lookahead map #1351

Uh oh!

Conversation

acomodi commented Jun 8, 2020

Description

Related Issue

Motivation and Context

How Has This Been Tested?

Types of changes

Checklist:

Uh oh!

acomodi commented Jun 15, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

litghost commented Jun 15, 2020

Uh oh!

litghost Jun 15, 2020

Choose a reason for hiding this comment

Uh oh!

litghost Jun 15, 2020

Choose a reason for hiding this comment

Uh oh!

litghost Jun 15, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

litghost Jun 15, 2020

Choose a reason for hiding this comment

Uh oh!

litghost Jun 15, 2020

Choose a reason for hiding this comment

Uh oh!

acomodi Jun 15, 2020

Choose a reason for hiding this comment

Uh oh!

litghost Jun 15, 2020

Choose a reason for hiding this comment

Uh oh!

litghost Jun 15, 2020

Choose a reason for hiding this comment

Uh oh!

litghost commented Jun 15, 2020

Uh oh!

acomodi commented Jun 15, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kmurray commented Jun 16, 2020

Uh oh!

acomodi commented Aug 4, 2020

Uh oh!

Uh oh!

acomodi commented Jun 15, 2020 •

edited

Loading

litghost Jun 15, 2020 •

edited

Loading

acomodi commented Jun 15, 2020 •

edited

Loading