[Visualpython-users] Does Glowscript have numpy-like performance?

[K.-Michael Aye]

Dear all,

I was wondering if the good old VPython gas simulation that uses numpy
so efficiently will run as nicely in Glowscript? Does Glowscript offer
efficient arrays?

Best regards,
Michael

[Bruce Sherwood]

The quick answer to "numpy in GlowScript" is "no". GlowScript
(glowscript.org) is necessarily written in JavaScript, and GlowScript
user programs are necessarily written in JavaScript (or CoffeeScript,
which compiles to JavaScript). JavaScript is slower than C++ but
considerably faster than Python. To my knowledge, there is no
possibility for something like numpy, which runs at the speed of C.

However, here is the stars program in GlowScript, a program which is
very similar to the gas program:

http://www.glowscript.org/#/user/GlowScriptDemos/folder/Examples/program/Stars

(You need to enable WebGL in your browser; see the Help at glowscript.org.)

As you see, it runs adequately fast. However, if you examine the code,
you'll see that David Scherer pulled some unintuitive tricks to get
the speed to be adequate. In fairness, you might say that the use of
numpy in the VPython demo programs stars.py and gas.py is also pretty
unintuitive.

Another speed issue to consider is that because GlowScript uses WebGL,
which requires the use of modern GPU-type graphics cards, many kinds
of graphics can be significantly faster than in VPython, which
currently uses GPUs only optionally, for textures. I've been dreaming
of a VPython+ that would also require the use of a modern graphics
card.

I just released GlowScript 1.0 which includes some major new
capabilities. A particularly interesting example is a pulse running
along a rug:

http://www.glowscript.org/#/user/GlowScriptDemos/folder/Examples/program/Rug

For another capability that goes beyond VPython, here is transparency
done right:

http://www.glowscript.org/#/user/GlowScriptDemos/folder/Examples/program/Transparency

In VPython, transparent objects are sorted by their pos attribute,
with the result that intersecting transparent objects cannot be
displayed correctly. I implemented "depth peeling" in GlowScript which
uses the GPUs to do pixel-level sorting. Here are the details:

http://www.khronos.org/message_boards/viewtopic.php?f=44&t=4905

Bruce Sherwood

Post by K.-Michael Aye
Dear all,
I was wondering if the good old VPython gas simulation that uses numpy
so efficiently will run as nicely in Glowscript? Does Glowscript offer
efficient arrays?
Best regards,
Michael

[Bruce Sherwood]

I should add for completeness that JavaScript does offer typed arrays
such as Float32Array, in which all elements are 32-bit floats. A major
use of these typed arrays is for communication between CPU and GPUs on
the graphics card.

Bruce Sherwood

Post by K.-Michael Aye
Dear all,
I was wondering if the good old VPython gas simulation that uses numpy
so efficiently will run as nicely in Glowscript? Does Glowscript offer
efficient arrays?
Best regards,
Michael

[Bruce Sherwood]

I've now converted the VPython demo program gas.py to GlowScript:

http://www.glowscript.org/#/user/GlowScriptDemos/folder/Examples/program/HardSphereGas

To run the program, you need to have a WebGL-enabled browser and a
modern graphics card. For details, see the Help at glowscript.org.

As for speed, this is what I found:

VPython gas.py with 100 atoms, dt = 1e-5, no graphing: 100 iterations
= 1.1 seconds. This similar Glowsript program with same conditions:
100 iterations = 11.7 seconds. Presumably the big difference is the
Python numpy library, which runs at C speed, and which eliminates some
explicit loops.

Bruce Sherwood

[Bruce Sherwood]

I was wrong. Firefox on my Windows 7 machine is also a 32-bit
application, but it runs the gas program about twice as fast as
Chrome. I also see the higher speed in Safari on a Macbook Pro of
comparable speed.

Bruce Sherwood

Correction: I made my measurements with Chrome on a 64-bit Windows
machine, unaware that Chrome runs slowly on 64-bit machines because it
is a 32-bit application. Here are improved measurements, copied from
On a 32-bit Windows computer running Chrome: VPython gas.py with 100
atoms, dt = 1e-5, rate(100), no graphing: 100 iterations = 1 second.
The similar Glowsript program with same conditions: 100 iterations =
4.5 seconds.
/With rate(300), VPython is 0.35 s, GlowScript remains 4.5 s.
Presumably the big difference is the Python numpy library, which runs
at C speed, and which eliminates some explicit loops.
Note that on a 64-bit Windows computer Chrome runs this program half
as fast as on a 32-bit computer, because Chrome is a 32-bit
application. On 64-bit Windows Firefox is a 64-bit application and
runs the program twice as fast as Chrome.
Bruce Sherwood

Post by Bruce Sherwood
http://www.glowscript.org/#/user/GlowScriptDemos/folder/Examples/program/HardSphereGas
To run the program, you need to have a WebGL-enabled browser and a
modern graphics card. For details, see the Help at glowscript.org.
VPython gas.py with 100 atoms, dt = 1e-5, no graphing: 100 iterations
100 iterations = 11.7 seconds. Presumably the big difference is the
Python numpy library, which runs at C speed, and which eliminates some
explicit loops.
Bruce Sherwood

[Bruce Sherwood]

After more study of GlowScript vs. VPython versions of the VPython
gas.py program:

VPython gas.py with 100 atoms, dt = 1e-5, rate 500, no graphing: 100
iterations = 0.4 seconds.

If gas.py is modified to not use numpy, the time is 3.6 s.

The similar Glowsript program with same conditions: 100 iterations =
3.2 seconds on Firefox.

Oddly, on at least one 64-bit Windows machine, the time is 5.3 seconds
on Chrome.

On a 32-bit Windows laptop, 4.2 seconds on Firefox, 2.5 seconds on
Chrome, the reverse of the ordering on the other Windows machine.

There is a very major difference between Chrome and Firefox. With
Firefox, the animation periodically freezes, probably due to
JavaScript garbage collection. These freeze-ups do not occur with
Chrome or Safari.

-----------------------

You can run this program (if you have a WebGL-enabled browser; see the
Help at glowscript.org) at

http://www.glowscript.org/#/user/GlowScriptDemos/folder/Examples/program/HardSphereGas

A summary statement is that the factor of 10 speed advantage of
VPython over GlowScript comes completely from the parallel C-speed
numpy arrays. Without numpy, the VPython and GlowScript (Python and
JavaScript) execution speeds are about the same. I had mistakenly
thought that JavaScript was much faster than Python, but that seems
not to be the case.

Bruce Sherwood

[Bruce Sherwood]

The plot thickens. David Scherer rewrote a small piece of the
GlowScript gas example, and the time on both Chrome and Firefox
dropped to 1 second, not so different from the 0.4 second seen with
the numpy version. (The change has to do with disabling vector
overloading in the GlowScript handling of JavaScript.)

Bruce Sherwood

Post by Bruce Sherwood
After more study of GlowScript vs. VPython versions of the VPython
VPython gas.py with 100 atoms, dt = 1e-5, rate 500, no graphing: 100
iterations = 0.4 seconds.
If gas.py is modified to not use numpy, the time is 3.6 s.
The similar Glowsript program with same conditions: 100 iterations =
3.2 seconds on Firefox.
Oddly, on at least one 64-bit Windows machine, the time is 5.3 seconds
on Chrome.
On a 32-bit Windows laptop, 4.2 seconds on Firefox, 2.5 seconds on
Chrome, the reverse of the ordering on the other Windows machine.
There is a very major difference between Chrome and Firefox. With
Firefox, the animation periodically freezes, probably due to
JavaScript garbage collection. These freeze-ups do not occur with
Chrome or Safari.
-----------------------
You can run this program (if you have a WebGL-enabled browser; see the
Help at glowscript.org) at
http://www.glowscript.org/#/user/GlowScriptDemos/folder/Examples/program/HardSphereGas
A summary statement is that the factor of 10 speed advantage of
VPython over GlowScript comes completely from the parallel C-speed
numpy arrays. Without numpy, the VPython and GlowScript (Python and
JavaScript) execution speeds are about the same. I had mistakenly
thought that JavaScript was much faster than Python, but  that seems
not to be the case.
Bruce Sherwood