Input Array becomes zeros on iOS devices

[sukhsung]

What is wrong?

When I pass an array to kernel, all of entries becomes 0 on iOS devices (Chrome and Safari)

Where does it happen?

Confirmed on at least three devices (my iPad, iPhone and friend's iPhone)

How do we replicate the issue?

Here is minimum working example : https://jsfiddle.net/yzpaefuo/5/

How important is this (1-5)?

5 if I'm not messing something up

Expected behavior (i.e. solution)

I should get same results on both all devices.

Other Comments

I've tested this on android devices and it works fine, and it works on my Mac and PC as well.
It also seems to work properly with older version of gpu.js. Specifically one that worked was 1.10.1

[robertleeplummerjr]

Looking at this now.

[robertleeplummerjr]

Does using tactic: 'precision' resolve the issue?

[sukhsung]

No, I think it breaks during evaluation without error. Nothing is being returned.

[sukhsung]

Actually, I lied. It returns something, but it's little off
0,1,2.01562476158141,2.999999523162842,3.999999761581421
Can this be avoided? My inputs in my actual code are smaller numbers, so having better precision is preferred. I guess I can try premultiply my input to large number and divide in kernel

[robertleeplummerjr]

I'm still investigating.

[harshkhandeparkar]

Can I be of any help?

[robertleeplummerjr]

It seems to be a bug with packing/unpacking floats on ios. Related code is that of "dragons":

gpu.js/src/backend/web-gl/fragment-shader.js

Line 246 in 1955ad3

// Here be dragons!

[harshkhandeparkar]

Which floats? User end floats or backend?

[harshkhandeparkar]

Can a strongly typed function fix this error?

[robertleeplummerjr]

Which floats? User end floats or backend?

Its complicated, but it has to do with how a 32 bit float array is:

1. uploaded (Float32Array to UInt8Array)
2. reconstituted/unpacked (8 bit vec4 -> 32 bit float)
3. packed (32 bit float -> 8 bit float)
4. and then downloaded (UInt8Array to Float32Array)

iOS and OSX are notoriously difficult to work with. Their driver's are buggy and they leave things half implemented. Boo Apple!

[harshkhandeparkar]

But the bug occurs even if I don't take any input array. ```js gpu.createKernel(function() { return this.thread.x; }, {output: [25]}) ``` This ^^^ also returns a weird output.

…

On Tue, 24 Mar, 2020, 11:23 PM Robert Plummer, ***@***.***> wrote: Which floats? User end floats or backend? Its complicated, but it has to do with how a 32 bit float array is: 1. uploaded (Float32Array to UInt8Array) 2. reconstituted/unpacked (8 bit vec4 -> 32 bit float) 3. packed (32 bit float -> 8 bit float) 4. and then downloaded (UInt8Array to Float32Array) iOS and OSX are notoriously difficult to work with. Their driver's are buggy and they leave things half implemented. Boo Apple! — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#585 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIJI5H74K5EVMXMVTHX2Y6DRJDXR7ANCNFSM4LSNZKEQ> .

[robertleeplummerjr]

I don't get a bug, per-say, I get a low precision: [0, 1, 2.014709234237671, 2.999084234237671, 3.999084234237671], but that actually does help narrow things down a little.

[harshkhandeparkar]

glitch rather than a bug.

[harshkhandeparkar]

Can a strongly typed function fix this error?

?

[harshkhandeparkar]

Can we not work with integers instead of floats?

[robertleeplummerjr]

Can a strongly typed function fix this error?

They are already strong types, no.

Can we not work with integers instead of floats?

Uint8Array, it is working with both.

[harshkhandeparkar]

Why is the output float and not int?

[robertleeplummerjr]

Why is the output float and not int?

Lets discuss implementation details on gitter, and here focus on the findings we come up with when they are found.

[harshkhandeparkar]

Okay.

…

On Wed, 25 Mar, 2020, 12:02 AM Robert Plummer, ***@***.***> wrote: Why is the output float and not int? Lets discuss implementation details on gitter, and here focus on the findings we come up with when they are found. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#585 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIJI5HYRIIKKABNITENDTJDRJD4EBANCNFSM4LSNZKEQ> .

[robertleeplummerjr]

It seems that you must use tactic: 'precision' with both Android and iOS, we'll go ahead and make it part of the internals.

[robertleeplummerjr]

It returns something, but it's little off
0,1,2.01562476158141,2.999999523162842,3.999999761581421

Still investigating that as well.

[robertleeplummerjr]

Released https://github.com/gpujs/gpu.js/releases/tag/2.9.2 which addresses the inaccurate lookup. My suspicion for values like 0,1,2.01562476158141,2.999999523162842,3.999999761581421 is that there is little we can do about it in the confines of webgl.