Splom open items

[etpinard]

Follow-up from #2372, #2505 and #2527

Things that were left out

• Splom traces placed on overlaid axes are broken, see Introducing splom traces #2505 (comment)
• Multi-selection across subplot is disallowed. To do so, we would need to share selectedpoints info across multiple splom-generated subplots, see Introducing splom traces #2505 (comment)
• Relayout'ing grid style on hasOnlyLargeSploms: true with scattergl trace(s) fails to redraw scattergl trace(s), see Implement better (more maintainable) way to redraw gl canvases on incremental updates #2562 FIXED in Add redrawReglTraces subroutine #3067

Things we could do, but would require some thought

• Add sub-trace on the diagonal (e.g. with diagonal.mode: 'scattergl' || 'histogram' || 'box' || 'violin', see Introducing splom traces #2505 (comment)
• Add attribute in dimensions that could default the splom generated axes' style (e.g. dimensions[i].showgrid: false would remove grid lines from all axes corresponding to dimensions[i]). Edit: PR Multi-axis-type sploms #2899 added dimensions[i].axis.type, it would be easy to add more attribute under that container.
• Make one single cartesian <g .draglayer> covering the whole plot area instead of multiple (12 I think) per subplot. This would speed up initial rendering (especially for splom, but also for all graphs with many subplots), make cross-subplot selections a little easier to manage, see Introducing splom traces #2505 (comment)
• optimize regl-line2d for grid data
• implement matching axes Matching axes #1549
• Remove remaining O(n^2) SVG items. In addition to g.draglayer mentioned above, there are two categories I can see:
  • Subplot backgrounds rect.bg - These could be omitted entirely when plot_bgcolor matches paper_bgcolor (and both are fully opaque - which is true by default), otherwise perhaps we could move them into the WebGL canvas though @etpinard points out this may impact pan/zoom performance. UPDATE: optimized when plot and paper bgcolor match in Aggressive splom perf #3057, possible future improvements in Aggressive splom perf #3057 (comment)
  • Subplot groups .subplot.x<N>y<M> - normally only the ones with tick labels and titles have any visible content in them, but currently they all have some empty groups, an associated clipPath#clip<UID>x<N>y<M>plot, and a few <path d="M0,0"> elements, we should be able to avoid even creating the unused ones.

Related issues that would benefit splom performance:

• faster axis autorange relayout: Faster axis autorange relayout #2546
• per-axis 'axrange' relayout: Per axis axrange edit type #2547
• stash pts selections for pan: Stash SVG points selections for faster pan #2548
• speed up cartesian axis defaults: Speed up cartesian axes supplyDefaults #2549
• replacing getBoundingClientRect calls in Axes.doTicks Replace getBoundingClientRect calls in axes.js #1988 - done in: Axes.draw w/o getBoundingClientRect + many axis automargin fixes #4165

[etpinard]

Pasting benchmarks from #2505 (comment) out for future reference:

---

Here are some first-render benchmarks with showupperhalf: false and diagonal.visible: false:

# of dimensions	# of values per dimension	regl-line2d grid lines?	time [ms]
10	20k	no	400-500
20	5k	yes	600-700
30	2k	yes	900-1000
40	1k	yes	1300-1400
50	800	yes	2000-2300

computed using:

var Nvars = /* fill this in */
var Nrows = /* fill this in */
var dims = []

for(var i = 0; i < Nvars; i++) {
  dims.push({values: []})
  
  for(var j = 0; j < Nrows; j++) {
     dims[i].values.push(Math.random())
  }
}

Plotly.purge(gd);

console.time('splom')
Plotly.newPlot(gd, [{
  type: 'splom',
  dimensions: dims,
  showupperhalf: false,
  diagonal: {visible: false},
}], {width: 2000, height: 2000})
console.timeEnd('splom')

[etpinard]

From @alexcjohnson 's #3057 (comment), about speeding up lsInner:

overlappingDomain might itself be a heavy lift for a large number of subplots (am I reading this right that for NxN subplots this is O(N^4), ie O(N^2) tests per subplot???), all for the fairly unusual case of inset plots, or otherwise partially occluding subplots. Perhaps it would be worthwhile doing a faster check for whether overlappingDomain is even necessary to call, by looping over all x axes to see if any of their domains overlap, then the same over all y axes? That would only be O(N^2) to possibly (usually!) avoid O(N^4).

[sebastian-raubach]

Hi, I was wondering if there has been any progress on this issue since late 2018? Especially on the "Add sub-trace on the diagonal" functionality.
Thanks

[AniWar]

Hello , any updates for diagonal sub trace ? Would love to have this feature .

[InCogNiTo124]

Bump for histograms on the diagonals