Connections better use typedefed types Segment, Synapse, ... #331

breznak · 2019-03-18T15:42:02Z

Connections (and related SP, TM) properly use typedefed types for Segment, SynapseIdx etc
- allowing us to really change them as needed
minor cleanups

Implements size_t for TM, SP, Connections #357

use const loop

performance critical method, make some variables const

remove bounds check for Permanence range, callers must ensure it's within [minPermanence, maxPermanence]

This reverts commit a5f4b13.

allow to be templated, not hardcoded to UInt

instead of UInt, This allows later changing easily the datatype of said typedef

This reverts commit 4df1210.

This reverts commit 09b6f51.

for Synapse, Segment, ... instead of UInt only

breznak

Please have a look and check if I correctly replace with the new types.

breznak · 2019-03-18T15:43:05Z

src/nupic/algorithms/Connections.cpp

@@ -303,10 +306,9 @@ void Connections::updateSynapsePermanence(Synapse synapse,
      potentialPreseg.push_back( segment );
    }

-    for (auto h : eventHandlers_) {
+    for (auto h : eventHandlers_) { //TODO handle callbacks in performance-critical method only in Debug?


this method is a hotspot I found in profiling, do we really need the callbacks here, can be left out only for debug,.. ?

Yea I have no idea why these exist but their part of the public API so I didn't remove them.
Maybe try: for( const auto & : eventHandlers_) for speed?

I'll profile w/o the loop, imo this would be a candidate to break the API, if shows performance gain.

breznak · 2019-03-18T15:44:45Z

src/nupic/algorithms/Connections.hpp

-typedef UInt16 SynapseIdx; /** Index of synapse in segment. */
-typedef UInt32 Segment;    /** Index of segment's data. */
-typedef UInt32 Synapse;    /** Index of synapse's data. */
+typedef UInt32 CellIdx;  //TODO instead of typedefs, use templates for proper type-checking


typedefs have the disadvantage of only renaming the type for the user(=programmer), for proper type-checking (so that compiler checks difference for SegmentIdx, SynapseIdx) we'd need to use templates instead.
I'll consider that in subsequent PR.

src/nupic/algorithms/Connections.hpp

breznak · 2019-03-18T15:48:00Z

src/nupic/algorithms/Connections.hpp

+typedef unsigned char SegmentIdx; /** Index of segment in cell. */
+typedef unsigned char SynapseIdx; /** Index of synapse in segment. */
+typedef UInt16 Segment;    /** Index of segment's data. */
+typedef UInt16 Synapse;    /** Index of synapse's data. */


note, a WIP commit changes these to much lower values.
I'll need to asses the impact, suprisingly, the lower values (compared to UInt) do not gain a significant performance gain (possibly some forgotten type-casting happens?)

Unit test SP:: ExactOutput "breaks" as stays stuck, does not finish, nor fail. I don;t know why

The size of this data type controls the maximum number of synapses & segments.
So with an 8-bit integer, it can only have a 256 synapses, maximum.
16-bits as shown here allows a maximum of 65,536 synapses.

I will give this a more thorough review within a day or two.

So with an 8-bit integer, it can only have a 256 synapses, maximum.
16-bits as shown here allows a maximum of 65,536 synapses.

yes. Are you suggesting the 16b should match the 8-bit? As SynapseIdx is "synapses per segment", and Synapse is a flat array/index to all existing synapses (their data). But there can be a number of segments in cells, so Synapses >> SynapseIdx.

Are you suggesting

No i just wanted to clarify.
What u have here looks good, as long as it doesnt overflow.

And if it does overflow then we can easily fix it by changing these typedefs which after this PR should be used much more consistently. :)

src/nupic/algorithms/Connections.cpp

src/test/unit/algorithms/SpatialPoolerTest.cpp

format. This does not break API anymore, as we broke serialization compatibility anyway.

in createSynapse, createSegment

which used Connections, and not really TP (aka Cells4)

ctrl-z-9000-times

This looks good to me.

rather that typedef

used in SDR_sparse_t, SDR_dense_t containers (vectors) ElemSparse must match with connections::CellIdx ElemDense : allow changing the type in the future

as these must be the same for some comparisons in SP, TM

breznak

One more review, please.

I've merged master
implemented "size_t" for Connections,TM,SP which was needed for interchanging the connections' typedefs
binded SDR sparse elem type & Connections CellIdx

breznak · 2019-04-04T19:15:19Z

src/nupic/algorithms/Connections.hpp

-typedef UInt32 Synapse;    /** Index of synapse's data. */
-typedef Real32 Permanence;
+//TODO instead of typedefs, use templates for proper type-checking?
+using CellIdx   = nupic::sdr::ElemSparse; // CellIdx must match with sdr::ElemSparse, to change, change it there


FYI I had to bind connections::CellIdx and sdr::ElemSparse, where ElemSparse is the element type of SDR_sparse_t (currently UInt32);
Both have to be the same for some methods in SP, TM, so I binded them together. It could be untangeled but then I'd have to do type-conversion of the vectors in said methods.

breznak · 2019-04-04T19:15:52Z

src/nupic/algorithms/Connections.hpp

@@ -333,7 +335,7 @@ class Connections : public Serializable
   *
   * @retval A vector length
   */
-  UInt32 segmentFlatListLength() const;
+  size_t segmentFlatListLength() const { return segments_.size(); };


implemented size_t for Connections, TM, and SP.

breznak · 2019-04-04T19:18:05Z

src/nupic/types/Sdr.hpp

-typedef std::vector<std::vector<UInt>>  SDR_coordinate_t;
-typedef std::function<void()>           SDR_callback_t;
+using ElemDense        = Byte; //TODO allow changing this
+using ElemSparse       = UInt32; //must match with connections::CellIdx 


ElemSparse & ElemDense represent data type of elements in SDR_sparse/dense_t;

ElemSparse is used to match in Connections with CellIdx

ElemDense is in preparation for changing this (at runtime with teplate?)

breznak · 2019-04-05T00:21:50Z

@dkeeney @ctrl-z-9000-times please have a look when you have time, I'd like to continue tomorrow and it'd help to have this already in.

ctrl-z-9000-times

Here are some tips for getting PRs merged.

Make smaller PRs. The size_t changes probably could have had their own PR, and then most of this work would have been merged a day ago.
Don't leave so many TODO notes in the code. Either do them, or talk about them on this github forum. Every change in the code is fair game for us to argue over, and it can really slow things down. If it's a comment of the github page then we can discuss it without blocking the PR from merging.

ctrl-z-9000-times · 2019-04-05T14:11:13Z

src/nupic/algorithms/Connections.hpp

-                  std::vector<SynapseIdx> &numActivePotentialSynapsesForSegment,
-                  const CellIdx activePresynapticCell,
-                  const Permanence connectedPermanence) const;
+  void computeActivity(std::vector<SynapseIdx> &numActiveConnectedSynapsesForSegment,  //TODO remove the 2 arg version, use only the 3 arg (with 1 being optional)


No can do, the two different versions do different things. One computes less than have as much as the other, making it significantly faster. Removing it would likely double the spatial poolers runtime.

One computes less than have as much as the other, making it significantly faster.

I'm aware, if we use {} the loop will skip, or some empheral value, as in TM(?)

ctrl-z-9000-times · 2019-04-05T14:22:41Z

src/nupic/algorithms/Connections.hpp


  /**
   * Gets the number of segments.
   *
   * @retval Number of segments.
   */
-  Segment numSegments() const;
+  size_t numSegments() const { 
+	  NTA_ASSERT(segments_.size() >= destroyedSegments_.size()); //TODO remove destroyedSegments_? 


This assert is not needed.
destroyedSegments_ is very much needed, otherwise segments could never be destroyed.

can I keep the assert?, I guess someone can make a working code-change where #destroyed > #segments_

One computes less than have as much as the other, making it significantly faster.

can we just rm the SegmentIdx from the list of segments?

can we just rm the SegmentIdx from the list of segments?

No, because then the other segment indexes would change. We could go through the work of updating these references since they're mostly contained within the connections class, but I'm not sure it would be any faster.

ctrl-z-9000-times · 2019-04-05T14:23:28Z

src/nupic/algorithms/Connections.hpp


  /**
   * Gets the number of synapses.
   *
   * @retval Number of synapses.
   */
-  Synapse numSynapses() const;
+  size_t numSynapses() const {
+    NTA_ASSERT(synapses_.size() >= destroyedSynapses_.size());


nit: assert not needed.

ctrl-z-9000-times · 2019-04-05T14:28:00Z

src/nupic/algorithms/Connections.hpp

-  bool operator==(const Connections &other) const;
-  bool operator!=(const Connections &other) const;
+  virtual bool operator==(const Connections &other) const;
+  inline bool operator!=(const Connections &other) const { return !operator==(other); }


virtual is not needed, since this class should never be subclassed. It's not supported, this is the only virtual method here. I expect if it were subclassed that a lot of things would break.

My understanding is that all modern compilers ignore inline and figuring out what to inline on its own.

I'll remove the virtual, you're right about the inline, but it does not hurt either.

inline is a nitpick, you can ignore me here.

ctrl-z-9000-times · 2019-04-05T14:37:36Z

src/nupic/algorithms/TemporalMemory.cpp

-    connections.createSynapse(segment, candidates[i], initialPermanence);
+  for (size_t c = 0; c < nActualWithMax; c++) {
+    const auto i = rng.getUInt32(static_cast<UInt32>(candidates.size()));
+    connections.createSynapse(segment, candidates[i], initialPermanence); //TODO createSynapse create a vector of new synapses at once


I disagree with this TODO note. I don't see any advantage to making a list of synapses and then making them all at once. It's more work & complexity to make, populate, and then iterate over an extra list. I think the performance of this issue is likely insignificant either way.

The next line however is almost certainly a performance issue:
candidates.erase(candidates.begin() + i);

I don't see any advantage to making a list of synapses and then making them all at once.

I tend to prefer passing data in vectors as long as possible, it avoids nested loops and is better for spotting hot-spots. Also SIMD (auto)vectorization can do a better job then.

The next line however is almost certainly a performance issue:
candidates.erase(candidates.begin() + i);

good point!

Anyway, out of scope of the PR, updated the comments and I'll open a separate issue.

src/nupic/algorithms/TemporalMemory.cpp

ctrl-z-9000-times · 2019-04-05T14:45:17Z

src/nupic/regions/TMRegion.cpp

-    VectorHelpers::unionOfVectors<UInt>(sdr.getSparse(), active, predictive);
-    sdr.setSparse(sdr.getSparse()); // to update the cache in SDR.
+    vector<CellIdx> uni;
+    VectorHelpers::unionOfVectors<CellIdx>(uni, active, predictive);


Before this was an in-place update and now it makes a new vector and swaps it in...

Good spot!
This was for mismatch in SDR_sparse_t, and vector //you can change CellIdx;
but got fixed later by binding connections::CellIdx and sdr::ElemSparse

TL;DR reverted back

breznak · 2019-04-05T15:26:45Z

ake smaller PRs. The size_t changes probably could have had their own PR, and then most of this work would have been merged a day ago.

you're right, this PR grew big, justification for adding size_t is that it's needed for the typedef changes to reliably work
- is it OK to review as is, or do I need to split?

Don't leave so many TODO notes in the code. Either do them, or talk about them on this github forum. Every change in the code is fair game for us to argue over, and it can really slow things down. If it's a comment of the github page then we can discuss it without blocking the PR from merging.

this is true, I've abused the code for discussion here, I'll rm the PRs or make them to issues.

ctrl-z-9000-times · 2019-04-05T15:31:24Z

is it OK to review as is, or do I need to split?

This is fine to review as it is, especially since I've already done the review. I just meant in general. Thanks!

breznak · 2019-04-05T22:18:23Z

@ctrl-z-9000-times I hope to have addressed your feedback, one more review, please

ctrl-z-9000-times

This looks great to me!

breznak · 2019-04-06T01:07:35Z

This looks great to me!

thanks for bearning with me on this PR!

breznak added 11 commits March 17, 2019 00:52

SP: remove round5_() as not needed

3244350

SP: globa inh: optimization

c6b3696

use const loop

Connections: updateSynapsePermanence optimization

a5f4b13

performance critical method, make some variables const

Connections::updateSynapsePermanence optimization

09b6f51

remove bounds check for Permanence range, callers must ensure it's within [minPermanence, maxPermanence]

Revert "Connections: updateSynapsePermanence optimization"

4df1210

This reverts commit a5f4b13.

VectorHelpers: union, sparse_cellsToColumns use template T

ca6cee2

allow to be templated, not hardcoded to UInt

Connections: use typedefs Synapse, Segment

6be5eb3

instead of UInt, This allows later changing easily the datatype of said typedef

Revert "Revert "Connections: updateSynapsePermanence optimization""

9535b56

This reverts commit 4df1210.

Revert "Connections::updateSynapsePermanence optimization"

7b17674

This reverts commit 09b6f51.

Connections, SP, TM: use typedefs from Connections

b8199bd

for Synapse, Segment, ... instead of UInt only

Connections: use smaller datatypes WIP

def0898

breznak added the code code enhancement, optimization, cleanup..programmer stuff label Mar 18, 2019

breznak self-assigned this Mar 18, 2019

breznak commented Mar 18, 2019

View reviewed changes

ctrl-z-9000-times reviewed Mar 19, 2019

View reviewed changes

src/nupic/algorithms/Connections.cpp Outdated Show resolved Hide resolved

ctrl-z-9000-times reviewed Mar 19, 2019

View reviewed changes

src/test/unit/algorithms/SpatialPoolerTest.cpp Show resolved Hide resolved

breznak added 14 commits March 19, 2019 18:29

Connections::load fix max value

858df3b

Connections::load simplify, support only version 2

aa84db2

format. This does not break API anymore, as we broke serialization compatibility anyway.

Connections: add check for data-type overflow

27511f0

in createSynapse, createSegment

Merge branch 'master_community' into connections_opt

f88416e

add debug info

bf56929

Connections: fix check comparison

ac9bfb4

TM: replace isSortedWithoutDuplicates with is_sorted

57e4e27

TM: sorted SDR indices for activateCells

c45e85f

ConnectionsPerformanceTest: cleanup and use SDRs

ad9cbea

ConnectionsPerformanceTest: remove testTP

3b59e36

which used Connections, and not really TP (aka Cells4)

Connections: fixing data-type overflow checks

5537fa2

ConnectionsPerformanceTest: use default SP params

3a389ad

Connections: fix type for removeSynapseFromPreSynapticMap_

99a20ff

TM: use typedefs from Connections

3a89a7c

Connections: improve operator==

e87166e

breznak mentioned this pull request Apr 4, 2019

Anomaly removal, cleanup #347

Closed

6 tasks

ctrl-z-9000-times previously approved these changes Apr 4, 2019

View reviewed changes

breznak added 7 commits April 4, 2019 18:30

TM: adapt to size_t changes from Connections

c1499f7

Merge branch 'master_community' into connections_opt

132d706

TM small fix

d27d9ce

SDR: use using for typedef

ecb5584

rather that typedef

SDR: add typedefs for ElemSparse, ElemDense

fa9e70d

used in SDR_sparse_t, SDR_dense_t containers (vectors) ElemSparse must match with connections::CellIdx ElemDense : allow changing the type in the future

COnnections: bind CellIdx with sdr::ElemSparse

d37ede8

as these must be the same for some comparisons in SP, TM

SdrTools: fixes for change-able SDR::ElemDense

069ba3c

breznak dismissed ctrl-z-9000-times’s stale review via 069ba3c April 4, 2019 19:09

breznak commented Apr 4, 2019

View reviewed changes

breznak requested a review from ctrl-z-9000-times April 4, 2019 19:27

breznak added 2 commits April 4, 2019 21:48

fix typo in Debug assert

db49a59

Merge branch 'master_community' into connections_opt

7c5921b

ctrl-z-9000-times requested changes Apr 5, 2019

View reviewed changes

breznak added 2 commits April 5, 2019 21:11

Merge branch 'master_community' into connections_opt

90169f7

review feedback

c46e4c4

breznak mentioned this pull request Apr 5, 2019

Connections: computeActivity() consider using only 3 arg version #371

Open

feedback 2

7cbe5b5

breznak mentioned this pull request Apr 5, 2019

TM:growSynapses improve methods #373

Open

3 tasks

breznak requested a review from ctrl-z-9000-times April 5, 2019 22:18

ctrl-z-9000-times approved these changes Apr 6, 2019

View reviewed changes

breznak merged commit fcfe07d into master Apr 6, 2019

breznak deleted the connections_opt branch April 6, 2019 01:07

Connections better use typedefed types Segment, Synapse, ... #331

Connections better use typedefed types Segment, Synapse, ... #331

Conversation

breznak commented Mar 18, 2019 • edited Loading

breznak left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ctrl-z-9000-times Mar 18, 2019 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ctrl-z-9000-times left a comment

Choose a reason for hiding this comment

breznak left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

breznak commented Apr 5, 2019

ctrl-z-9000-times left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

breznak commented Apr 5, 2019

ctrl-z-9000-times commented Apr 5, 2019

breznak commented Apr 5, 2019

ctrl-z-9000-times left a comment

Choose a reason for hiding this comment

breznak commented Apr 6, 2019

breznak commented Mar 18, 2019 •

edited

Loading

ctrl-z-9000-times Mar 18, 2019 •

edited

Loading