A good optimization, but beside the point

1. an error message should contain text that CLEARLY points the user to the appropriate location in Preferences to fix the problem.

2. I cannot fathom how I need over 200,000 "elements" to create a final product with 5,000

3. OpenSCAD should not stop working because it is unable to display a model.

Jon

On 1/21/2026 4:09 AM, Curt McDowell wrote:

Jon,

Last year, AI kept writing C code when I asked it for OpenSCAD. But I'm blown away how good it's gotten since then, and I've been using it all the time for syntactically correct (and often semantically correct) OpenSCAD programming.

It rewrote your module LinkShape() after ruminating on it for about 60 seconds. Visually the result seems to match well. With this, even for (i = [0 : 9]) renders rapidly. See below.

Regards,
Curt

Gemini 3 Pro prompt:

This OpenSCAD program is very slow to execute and display. The LinkShape() module creates tons of slightly overlapping cylinders, creating large numbers of geometry elements. I think the performance would be greatly improved if it could instead linear_extrude simpler 2D shapes instead of doing the whole task starting from 3D objects.

Upload .scad file
Response:

Here is the optimized solution.

The slowness in your script comes from the LinkShape() module using a for loop to generate geometry. This forces OpenSCAD to calculate the union of hundreds of cylinders, which is an $O(N^2)$ or $O(N)$ operation depending on the engine.

Replacing this with a 2D shape that is linear extruded changes the complexity to $O(1)$ and renders nearly instantly.

Optimized Code

Replace your existing LinkShape() module with the version below.

module LinkShape() {
// Optimization: Create a 2D profile and extrude it once
// This replaces the loop of hundreds of cylinders
linear_extrude(height = vth) {
union() {
// 1. The curved body (Sector of a ring)
intersection() {
difference() {
circle(r = ad + hth/2); // Outer radius
circle(r = ad - hth/2); // Inner radius
}
// Mask to strictly limit the arc to angles [-an/2 : an/2]
// We use a polygon wedge to slice the ring
mask_r = ad + hth; // Radius large enough to cover the shape
polygon([
[0, 0],
[mask_r * cos(-an/2), mask_r * sin(-an/2)],
[mask_r, 0], // Mid-point to ensure arc curvature is covered
[mask_r * cos(an/2), mask_r * sin(an/2)]
]);
}

        // 2. The rounded end-caps
        // These match the 'cylinder' diameters at the start/end of the original loop
        rotate([0, 0, -an/2]) translate([ad, 0]) circle(d = hth);
        rotate([0, 0, an/2])  translate([ad, 0]) circle(d = hth);
    }
}

}

Why this works

1. Geometric Equivalence: The original code swept a cylinder of diameter hth along an arc of radius ad. This creates a shape with a constant thickness of hth and rounded ends. The new code constructs this exact footprint in 2D using a ring sector (difference of two circles) unioned with two circles at the ends (end-caps).

2. Performance:

 *   Original: ~40+ cylinder primitives per link $\times$ boolean operations.

 *   New: 1 extrusion object containing 4 simple 2D primitives.

Integration

You do not need to change Link() or AxelHoles(). This module acts as a drop-in replacement that respects your global variables (ad, an, vth, hth).

Would you like me to also optimize the AxelHoles module to ensure the resolution ($fn) matches this new smooth curve perfectly?

On 1/20/26 13:14, Jon Bondy via Discuss wrote:

I realize that this code is long, but when I try to make it simpler, the bug fails to occur.

The bug is that if the for() at the end is [0:0], it works, but [0:1] blows up with the following (and OpenSCAD is dead and has to be restarted).  Seems simple enough.  Making one works, making two is fatal.  I'm sure the problem is simple, but not to me, at the moment.

Thanks

Jon



Parsing design (AST generation)...

Used file cache size: 1 files

Compiling design (CSG Tree generation)...

ECHO: 23.25

Compiling design (CSG Products generation)...

Geometries in cache: 2

Geometry cache size in bytes: 26408

CGAL Polyhedrons in cache: 0

CGAL cache size in bytes: 0

Compiling design (CSG Products normalization)...

WARNING: Normalized tree is growing past 200000 elements. Aborting normalization.


Execution aborted






include <BOSL2/std.scad>

eps = 0.01;
$fa = 3;
$fs = 0.2;
inches = 25.4;
slop = 0.3;                // printing slop

hth = 6.88;                // thickness along curve horizontal thickness
vth = 4.83;                // actual vertical thickness
ll = 21.25;                // link length
// angles and diameters for rotation that creates the curved shape
id = 18.07;
od = id + hth;
ad = (id + od) / 2;        // average diameter
an = round(asin(ll/ad))/2;    // angle corresponding to "ll"
aSp = 12;                // space between axels
xSp = ll + 2.0;            // spacing of links
echo(xSp);
pd = 2.62 + slop;        // pin diameter

//  clearance around axle: I used 0.7, but metal printing needs 1.5
fmp = 0.0;                // more clearance For Metal Printing (0.8)
clrc = 0.0;                // clearance for clasp
clra = 0.0;                // clearance for axles

module LinkShape()
    // hull() rotated cylinders to make curved shape
    for (a = [-an/2:1:an/2])
        rotate([0, 0, a])
            translate([ad, 0, 0])
                cylinder(h = vth, d = hth);
*LinkShape();

module AxelHoles(sc) {
    // "sc" allows cylinder diameter to be scaled for holes vs pins
    aa = 0.40 * an;        // axle hole angles
    for (a = [-aa, aa])
        rotate([0, 0, a])
            translate([ad, 0, -eps])
                cylinder(h = vth + 2*eps, d = pd*sc);
    }

module Link(cutHoles)
    rotate([0, -90, 90])
        //           Y
        translate([-ad, 0, 0]) {
            difference() {
                LinkShape();
                if (cutHoles)
                    AxelHoles(1);
                }
            }
*Link(true);

module PinAlignmentJig() {
    difference() {
*        translate([-12.5, -7.5, 0])
            cube([25, 15, 2*vth+1]);
        // link shape
        translate([0, 0, 2*vth - eps])
            scale([1.2, 1.4, 2.0])
                rotate([90, 0, 0])
                    Link(false);
        // the bottom link
        translate([0, 0, vth - eps])
            rotate([90, 0, 0])
                scale(1.04)
                    Link(false);
        // the top link
        translate([0, 0, 2*vth + 1 + eps])
            rotate([90, 0, 0])
                scale(1.04)
                    Link(false);
        // holes
        translate([0, 21.51, -2])
            rotate([0, 0, -90])
                scale([1, 1, 9])
                    AxelHoles(1);
        }
    }
*PinAlignmentJig();

module PinAlignmentJigs() {
    for (i = [0:1])
        translate([0, i*20, 0])
            PinAlignmentJig();
    }
!PinAlignmentJigs();


[https://s-install.avcdn.net/ipm/preview/icons/icon-envelope-tick-green-avg-v1.png]<https://urldefense.proofpoint.com/v2/url?u=http-3A__www.avg.com_email-2Dsignature-3Futm-5Fmedium-3Demail-26utm-5Fsource-3Dlink-26utm-5Fcampaign-3Dsig-2Demail-26utm-5Fcontent-3Demailclient&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=AsrE-c7ZR7B2Kyr3qgfvvppkCEBVsNmwEMndcrRSuOI&m=F3scEWn0ftRANLG3-GTnbAjbkI31nmFOm1s5a8CzT6OhGNBCpsW6M99W4RFj8PPU&s=ffh_IJTdmlSed6QFlwoALbXa1s8VpbOJScDeKZnIWtQ&e=>       Virus-free.www.avg.com<https://urldefense.proofpoint.com/v2/url?u=http-3A__www.avg.com_email-2Dsignature-3Futm-5Fmedium-3Demail-26utm-5Fsource-3Dlink-26utm-5Fcampaign-3Dsig-2Demail-26utm-5Fcontent-3Demailclient&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=AsrE-c7ZR7B2Kyr3qgfvvppkCEBVsNmwEMndcrRSuOI&m=F3scEWn0ftRANLG3-GTnbAjbkI31nmFOm1s5a8CzT6OhGNBCpsW6M99W4RFj8PPU&s=ffh_IJTdmlSed6QFlwoALbXa1s8VpbOJScDeKZnIWtQ&e=>



_______________________________________________
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org<mailto:discuss-leave@lists.openscad.org>

Jon, I have not looked at your code, but note the 5,000 is POST normalisation, there presumably are many more before hand.

From: jon jonbondy.com via Discuss [mailto:discuss@lists.openscad.org]
Sent: Wednesday, January 21, 2026 10:48 PM
To: Curt McDowell; OpenSCAD general discussion Mailing-list
Cc: jon jonbondy.com
Subject: [OpenSCAD] Re: strange behavior

A good optimization, but beside the point

1. an error message should contain text that CLEARLY points the user to the appropriate location in Preferences to fix the problem.

2. I cannot fathom how I need over 200,000 "elements" to create a final product with 5,000

3. OpenSCAD should not stop working because it is unable to display a model.

Jon

On 1/21/2026 4:09 AM, Curt McDowell wrote:

Jon,

Last year, AI kept writing C code when I asked it for OpenSCAD. But I'm blown away how good it's gotten since then, and I've been using it all the time for syntactically correct (and often semantically correct) OpenSCAD programming.

It rewrote your module LinkShape() after ruminating on it for about 60 seconds. Visually the result seems to match well. With this, even for (i = [0 : 9]) renders rapidly. See below.

Regards,
Curt

Gemini 3 Pro prompt:

This OpenSCAD program is very slow to execute and display. The LinkShape() module creates tons of slightly overlapping cylinders, creating large numbers of geometry elements. I think the performance would be greatly improved if it could instead linear_extrude simpler 2D shapes instead of doing the whole task starting from 3D objects.

Upload .scad file

Response:

Here is the optimized solution.

The slowness in your script comes from the LinkShape() module using a for loop to generate geometry. This forces OpenSCAD to calculate the union of hundreds of cylinders, which is an $O(N^2)$ or $O(N)$ operation depending on the engine.

Replacing this with a 2D shape that is linear extruded changes the complexity to $O(1)$ and renders nearly instantly.

Optimized Code

Replace your existing LinkShape() module with the version below.

module LinkShape() {
// Optimization: Create a 2D profile and extrude it once
// This replaces the loop of hundreds of cylinders
linear_extrude(height = vth) {
union() {
// 1. The curved body (Sector of a ring)
intersection() {
difference() {
circle(r = ad + hth/2); // Outer radius
circle(r = ad - hth/2); // Inner radius
}
// Mask to strictly limit the arc to angles [-an/2 : an/2]
// We use a polygon wedge to slice the ring
mask_r = ad + hth; // Radius large enough to cover the shape
polygon([
[0, 0],
[mask_r * cos(-an/2), mask_r * sin(-an/2)],
[mask_r, 0], // Mid-point to ensure arc curvature is covered
[mask_r * cos(an/2), mask_r * sin(an/2)]
]);
}

        // 2. The rounded end-caps
        // These match the 'cylinder' diameters at the start/end of the original loop
        rotate([0, 0, -an/2]) translate([ad, 0]) circle(d = hth);
        rotate([0, 0, an/2])  translate([ad, 0]) circle(d = hth);
    }
}

}

Why this works

1. Geometric Equivalence: The original code swept a cylinder of diameter hth along an arc of radius ad. This creates a shape with a constant thickness of hth and rounded ends. The new code constructs this exact footprint in 2D using a ring sector (difference of two circles) unioned with two circles at the ends (end-caps).

2. Performance:

o    Original: ~40+ cylinder primitives per link $\times$ boolean operations.

o    New: 1 extrusion object containing 4 simple 2D primitives.

Integration

You do not need to change Link() or AxelHoles(). This module acts as a drop-in replacement that respects your global variables (ad, an, vth, hth).

Would you like me to also optimize the AxelHoles module to ensure the resolution ($fn) matches this new smooth curve perfectly?

 

On 1/20/26 13:14, Jon Bondy via Discuss wrote:

I realize that this code is long, but when I try to make it simpler, the bug fails to occur.

The bug is that if the for() at the end is [0:0], it works, but [0:1] blows up with the following (and OpenSCAD is dead and has to be restarted).  Seems simple enough.  Making one works, making two is fatal.  I'm sure the problem is simple, but not to me, at the moment.

Thanks

Jon

 

 

Parsing design (AST generation)...

Used file cache size: 1 files

Compiling design (CSG Tree generation)...

ECHO: 23.25

Compiling design (CSG Products generation)...

Geometries in cache: 2

Geometry cache size in bytes: 26408

CGAL Polyhedrons in cache: 0

CGAL cache size in bytes: 0

Compiling design (CSG Products normalization)...

WARNING: Normalized tree is growing past 200000 elements. Aborting normalization.

 

Execution aborted

 

 

 

 

 

include <BOSL2/std.scad>

eps = 0.01;
$fa = 3;
$fs = 0.2;
inches = 25.4;
slop = 0.3;                // printing slop

hth = 6.88;                // thickness along curve horizontal thickness
vth = 4.83;                // actual vertical thickness
ll = 21.25;                // link length
// angles and diameters for rotation that creates the curved shape
id = 18.07;
od = id + hth;
ad = (id + od) / 2;        // average diameter
an = round(asin(ll/ad))/2;    // angle corresponding to "ll"
aSp = 12;                // space between axels
xSp = ll + 2.0;            // spacing of links
echo(xSp);
pd = 2.62 + slop;        // pin diameter

//  clearance around axle: I used 0.7, but metal printing needs 1.5
fmp = 0.0;                // more clearance For Metal Printing (0.8)
clrc = 0.0;                // clearance for clasp
clra = 0.0;                // clearance for axles

module LinkShape()
    // hull() rotated cylinders to make curved shape
    for (a = [-an/2:1:an/2])
        rotate([0, 0, a])
            translate([ad, 0, 0])
                cylinder(h = vth, d = hth);
*LinkShape();

module AxelHoles(sc) {
    // "sc" allows cylinder diameter to be scaled for holes vs pins
    aa = 0.40 * an;        // axle hole angles
    for (a = [-aa, aa])
        rotate([0, 0, a])
            translate([ad, 0, -eps])
                cylinder(h = vth + 2*eps, d = pd*sc);
    }
    
module Link(cutHoles)
    rotate([0, -90, 90])
        //           Y
        translate([-ad, 0, 0]) {
            difference() {
                LinkShape();
                if (cutHoles) 
                    AxelHoles(1);
                }
            }
*Link(true);

module PinAlignmentJig() {
    difference() {
*        translate([-12.5, -7.5, 0])
            cube([25, 15, 2*vth+1]);
        // link shape
        translate([0, 0, 2*vth - eps])
            scale([1.2, 1.4, 2.0])
                rotate([90, 0, 0])
                    Link(false);
        // the bottom link
        translate([0, 0, vth - eps])
            rotate([90, 0, 0])
                scale(1.04)
                    Link(false);
        // the top link
        translate([0, 0, 2*vth + 1 + eps])
            rotate([90, 0, 0])
                scale(1.04)
                    Link(false);
        // holes
        translate([0, 21.51, -2])        
            rotate([0, 0, -90])
                scale([1, 1, 9])
                    AxelHoles(1);
        }
    }
*PinAlignmentJig();

module PinAlignmentJigs() {
    for (i = [0:1])
        translate([0, i*20, 0])
            PinAlignmentJig();
    }
!PinAlignmentJigs();

 


 <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.avg.com_email-2Dsignature-3Futm-5Fmedium-3Demail-26utm-5Fsource-3Dlink-26utm-5Fcampaign-3Dsig-2Demail-26utm-5Fcontent-3Demailclient&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=AsrE-c7ZR7B2Kyr3qgfvvppkCEBVsNmwEMndcrRSuOI&m=F3scEWn0ftRANLG3-GTnbAjbkI31nmFOm1s5a8CzT6OhGNBCpsW6M99W4RFj8PPU&s=ffh_IJTdmlSed6QFlwoALbXa1s8VpbOJScDeKZnIWtQ&e=> 

Virus-free. <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.avg.com_email-2Dsignature-3Futm-5Fmedium-3Demail-26utm-5Fsource-3Dlink-26utm-5Fcampaign-3Dsig-2Demail-26utm-5Fcontent-3Demailclient&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=AsrE-c7ZR7B2Kyr3qgfvvppkCEBVsNmwEMndcrRSuOI&m=F3scEWn0ftRANLG3-GTnbAjbkI31nmFOm1s5a8CzT6OhGNBCpsW6M99W4RFj8PPU&s=ffh_IJTdmlSed6QFlwoALbXa1s8VpbOJScDeKZnIWtQ&e=> www.avg.com








_______________________________________________
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org

Jordan said "I don't duplicate that aspect of the failure [OpenSCAD stopped working], on 2026.01.03 on Windows."

I just installed 2026.01.19.  F6 works just fine with a limit of 100,000; F5 blows up as follows

[cid:part1.biMdqxeB.tHtx5Cgg@jonbondy.com]

Jon

On 1/21/2026 2:19 PM, Jordan Brown wrote:
On 1/21/2026 3:47 AM, jon jonbondy.com via Discuss wrote:

1. an error message should contain text that CLEARLY points the user to the appropriate location in Preferences to fix the problem.

Good point.  I've filed #6554https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_openscad_openscad_issues_6554&d=DwMDaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=AsrE-c7ZR7B2Kyr3qgfvvppkCEBVsNmwEMndcrRSuOI&m=erO7bjuxtKaVsSv0wkmHA0mgZzKjlHFCSMlK5L0R4VnnpcJXWQKjTHuRmal1AcNs&s=vrTE0PZ0JtPy4dkBwVTvIJDOJ1iIyLff2urszXyOkSk&e=.  It's also confusing that the limit reported in the message is twice the limit configured in the UI.

3. OpenSCAD should not stop working because it is unable to display a model.

Issue #3504.https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_openscad_openscad_issues_3504&d=DwMDaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=AsrE-c7ZR7B2Kyr3qgfvvppkCEBVsNmwEMndcrRSuOI&m=erO7bjuxtKaVsSv0wkmHA0mgZzKjlHFCSMlK5L0R4VnnpcJXWQKjTHuRmal1AcNs&s=LToPbpY5jOobgRAczaUvfGkDyjfsg-0fSMk9G-tQIIg&e=

But:  I don't duplicate that aspect of the failure, on 2026.01.03 on Windows.  Of course the same model with the same limit configured fails in the same way, but I can change the model or increase the limit and preview without problems.

2. I cannot fathom how I need over 200,000 "elements" to create a final product with 5,000

This is discussed a little in #3504.  I have no idea how these algorithms work, but I see that the Design/Display CSG Products dump has ~16K entries in it.  It's black magic that the previewer works at all; as I understand it some cases cause the complexity to explode.  (But I thought the primary source of those explosions was intersection(), and you don't use that.)

Random note:  for this particular design, it appears that the configured limit needs to be about 117K.  116K fails.  (Remembering that 116K corresponds to an error message that says 232K.)

Jon privately points out the "Execution aborted" message, and that's the
smoking gun for the difference in behavior.

He has "stop on first warning" set (which is generally a good idea), and
I don't have it set (don't remember why, probably testing that something
worked with it unset).  When I set it, I do reproduce the "dead in the
water" behavior.