From Newsgroup: sci.physics.relativity

Hi,

Tablets and phone are more annoying to
use with WebGPU. The usual browsers don't
have a Chrome DevTools panel integrated,

so that one could do JavaScript Debugging
directly on the device. Instead one has to
use a desktop machine, and connect the

device via UBS-C , and start a Chrome
Browser there . And then start a Chrome
DevTools panel alone, that is pair with

the device, via UBS-C cable. So this way
I already see where it crashes on the
tablets and phone:

await output.mapAsync(GPUMapMode.READ)
Unhandled Promise Rejection: OperationError

The above is the error that one can re-produce
already here with this test:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Not sure what exactly happens. Maybe
a form of timeout or device lost, that the
primitive HTML / JavaScript doesn't handle

gracefully yet. Maybe redimensioning the
test, so that it consumes less time would
help. Who knows? Will see. For production

use of a GPU integration I have to anyway
provide work slicing it seems.

Bye

Mild Shock schrieb:

Hi,

Why does this Lama have a red pyjama.
Oh, its a baby Lama. Its still in the cradle
and needs some training:

RedPajama-Data-v2
https://github.com/togethercomputer/RedPajama-Data

But then Andrej Karpathy recently showed
GPT-2 training on rented GPUs for less
than 100 USD in less then 2 hours.

So where do these grown up Lamas go.
Well Georgi Gerganov prefered C++/C
when he shouted Llama Llama Red Pyjama.

But you also find WebLLM, wrapping the
underlying C++/C GPU interface via the
W3C standard WebGPU / WGSL, with JavaScript:

In-Browser LLM Inference Engine
https://webllm.mlc.ai/

My experience with WebLLM 6 months
ago on an iPad Pro 2024, still a little early
stage performance and robustness.

But hey hardware of AI mobile iGPUs is
still evolving, and AI laptop, AI smartphones
and AI tablets, will soon feature Chinese

hardware such some new Kirin AI in 2027.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Hi,

Looking at the floor plan of a NPU:

Getting peak TOPS on a Ryzen AI 7 350 NPU https://destevez.net/2026/05/getting-peak-tops-on-a-ryzen-ai-7-350-npu/

It seems to me comms between tiles takes
at least Manhattan Distance or L1 Norm time,
if there is no comms congestion

But how does a packet travel? This way:

+----E
|
|
S

Or this way, from start S to end E:

+-E
+
+
S

And what does the chip do if there is
traffic congestion? Some papers are
here, possibly an old problem giving

that processor "cubes" are nothing new.
But a "cube" would be 3D and not 2D.
This paper is old from 2007 or so:

Routing Algorithms for 2D NoC Architectures http://cva.stanford.edu/classes/ee382c/research/2DRouting.pdf

Bye

Mild Shock schrieb:

Hi,

Tablets and phone are more annoying to
use with WebGPU. The usual browsers don't
have a Chrome DevTools panel integrated,

so that one could do JavaScript Debugging
directly on the device. Instead one has to
use a desktop machine, and connect the

device via UBS-C , and start a Chrome
Browser there . And then start a Chrome
DevTools panel alone, that is pair with

the device, via UBS-C cable. So this way
I already see where it crashes on the
tablets and phone:

await output.mapAsync(GPUMapMode.READ)
Unhandled Promise Rejection: OperationError

The above is the error that one can re-produce
already here with this test:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Not sure what exactly happens. Maybe
a form of timeout or device lost, that the
primitive HTML / JavaScript doesn't handle

gracefully yet. Maybe redimensioning the
test, so that it consumes less time would
help. Who knows? Will see. For production

use of a GPU integration I have to anyway
provide work slicing it seems.

Bye

Mild Shock schrieb:

Hi,

Why does this Lama have a red pyjama.
Oh, its a baby Lama. Its still in the cradle
and needs some training:

RedPajama-Data-v2
https://github.com/togethercomputer/RedPajama-Data

But then Andrej Karpathy recently showed
GPT-2 training on rented GPUs for less
than 100 USD in less then 2 hours.

So where do these grown up Lamas go.
Well Georgi Gerganov prefered C++/C
when he shouted Llama Llama Red Pyjama.

But you also find WebLLM, wrapping the
underlying C++/C GPU interface via the
W3C standard WebGPU / WGSL, with JavaScript:

In-Browser LLM Inference Engine
https://webllm.mlc.ai/

My experience with WebLLM 6 months
ago on an iPad Pro 2024, still a little early
stage performance and robustness.

But hey hardware of AI mobile iGPUs is
still evolving, and AI laptop, AI smartphones
and AI tablets, will soon feature Chinese

hardware such some new Kirin AI in 2027.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

98 more questions to go, don't give up!

Bye

Mild Shock schrieb:

Hi,

Tablets and phone are more annoying to
use with WebGPU. The usual browsers don't
have a Chrome DevTools panel integrated,

so that one could do JavaScript Debugging
directly on the device. Instead one has to
use a desktop machine, and connect the

device via UBS-C , and start a Chrome
Browser there . And then start a Chrome
DevTools panel alone, that is pair with

the device, via UBS-C cable. So this way
I already see where it crashes on the
tablets and phone:

await output.mapAsync(GPUMapMode.READ)
Unhandled Promise Rejection: OperationError

The above is the error that one can re-produce
already here with this test:

11.4 Giga Lips with a Budget Laptop https://github.com/Jean-Luc-Picard-2021/gigabudget

Not sure what exactly happens. Maybe
a form of timeout or device lost, that the
primitive HTML / JavaScript doesn't handle

gracefully yet. Maybe redimensioning the
test, so that it consumes less time would
help. Who knows? Will see. For production

use of a GPU integration I have to anyway
provide work slicing it seems.

Bye

Mild Shock schrieb:

Hi,

Why does this Lama have a red pyjama.
Oh, its a baby Lama. Its still in the cradle
and needs some training:

RedPajama-Data-v2
https://github.com/togethercomputer/RedPajama-Data

But then Andrej Karpathy recently showed
GPT-2 training on rented GPUs for less
than 100 USD in less then 2 hours.

So where do these grown up Lamas go.
Well Georgi Gerganov prefered C++/C
when he shouted Llama Llama Red Pyjama.

But you also find WebLLM, wrapping the
underlying C++/C GPU interface via the
W3C standard WebGPU / WGSL, with JavaScript:

In-Browser LLM Inference Engine
https://webllm.mlc.ai/

My experience with WebLLM 6 months
ago on an iPad Pro 2024, still a little early
stage performance and robustness.

But hey hardware of AI mobile iGPUs is
still evolving, and AI laptop, AI smartphones
and AI tablets, will soon feature Chinese

hardware such some new Kirin AI in 2027.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Hi,

Ok, following the instructions here:

npm install webgpu
https://github.com/dawn-gpu/node-webgpu

I can now run webgpu also from CLI:

node.exe dogelog.mjs

Dogelog Spieler 2.2.5, Node, JavaScript 26.4.0
(c) 1985-2026, XLOG Technologies AG, Schweiz

?- ensure_loaded(library(edge/furryhaze)).
true.

?- between(1,3,_), time(expedite((between(1,100,_),
between(1,100,_), between(1,100,_)), [size(4096)])), fail.
% Zeit 1037.994 ms, GC 0.000 ms, Lips 111 k
% Zeit 1091.131 ms, GC 0.000 ms, Lips 106 k
% Zeit 1045.274 ms, GC 0.000 ms, Lips 110 k
fail.

Same benchmark result as in the browser.
Now I can rent a bigger GPU by the hour
and do some easy CLI testing.

LoL

Bye

Mild Shock schrieb:

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

98 more questions to go, don't give up!

Bye

Mild Shock schrieb:

Hi,

Tablets and phone are more annoying to
use with WebGPU. The usual browsers don't
have a Chrome DevTools panel integrated,

so that one could do JavaScript Debugging
directly on the device. Instead one has to
use a desktop machine, and connect the

device via UBS-C , and start a Chrome
Browser there . And then start a Chrome
DevTools panel alone, that is pair with

the device, via UBS-C cable. So this way
I already see where it crashes on the
tablets and phone:

await output.mapAsync(GPUMapMode.READ)
Unhandled Promise Rejection: OperationError

The above is the error that one can re-produce
already here with this test:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Not sure what exactly happens. Maybe
a form of timeout or device lost, that the
primitive HTML / JavaScript doesn't handle

gracefully yet. Maybe redimensioning the
test, so that it consumes less time would
help. Who knows? Will see. For production

use of a GPU integration I have to anyway
provide work slicing it seems.

Bye

Mild Shock schrieb:

Hi,

Why does this Lama have a red pyjama.
Oh, its a baby Lama. Its still in the cradle
and needs some training:

RedPajama-Data-v2
https://github.com/togethercomputer/RedPajama-Data

But then Andrej Karpathy recently showed
GPT-2 training on rented GPUs for less
than 100 USD in less then 2 hours.

So where do these grown up Lamas go.
Well Georgi Gerganov prefered C++/C
when he shouted Llama Llama Red Pyjama.

But you also find WebLLM, wrapping the
underlying C++/C GPU interface via the
W3C standard WebGPU / WGSL, with JavaScript:

In-Browser LLM Inference Engine
https://webllm.mlc.ai/

My experience with WebLLM 6 months
ago on an iPad Pro 2024, still a little early
stage performance and robustness.

But hey hardware of AI mobile iGPUs is
still evolving, and AI laptop, AI smartphones
and AI tablets, will soon feature Chinese

hardware such some new Kirin AI in 2027.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

On 02/08/2026 8:47 AM, Mild Shock wrote:

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

Don't worry about it. There are several regulars here
who
don't
understand

that programming can be done for fun.

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

98 more questions to go, don't give up!

Here in sci.math, as everyone knows, I'm gearing up for
/linear algebra/ for fun. Still waiting for DVDs because
I'm not in a hurry. The book /Linear Algebra Done Right/
is interesting, and I've yet to go through the other rec-
commendations.[1]

I'm curious if you've ever thought of doing OpenGL with Prolog?

Does that even work?


[1] I have no idea how this word is supposed to be hyphenated,
I just do it anyway, because I'm not an LLM.

Bye

Take care!
--
Johann | email: invalid -> com | http://www.myrkraverk.com/blog/
I'm not from the Internet, I just work there. | via Easynews.com
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

On 08/02/2026 09:34 AM, Johann 'Myrkraverk' Oskarsson wrote:

On 02/08/2026 8:47 AM, Mild Shock wrote:

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

Don't worry about it. There are several regulars here
who
don't
understand

that programming can be done for fun.

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

98 more questions to go, don't give up!

Here in sci.math, as everyone knows, I'm gearing up for
/linear algebra/ for fun. Still waiting for DVDs because
I'm not in a hurry. The book /Linear Algebra Done Right/
is interesting, and I've yet to go through the other rec-
commendations.[1]

I'm curious if you've ever thought of doing OpenGL with Prolog?

Does that even work?

[1] I have no idea how this word is supposed to be hyphenated,
I just do it anyway, because I'm not an LLM.

Bye

Take care!

You might have good luck looking up reputable university programs
and seeing what textbooks they require, these days.

Or, you know, just buy old ones when the library retires
the old good ones.

How about Householder's "The Theory of Matrices in Numerical Analysis".

Linear independence and linear spaces inevitably
get associated with vector spaces. There are much
simpler accounts though of reflections and rotations
about the determinantal and the singular and the decompositions
and the forms and the echelon forms and reduction with regards
to things like cumulants and orthogonants and the matroids,
vis-a-vis usual closed categories and so on.

The cumulants and orthogonants and so on are lesser-served
accounts of the earlier 20'th century, and determinantal analysis, while
the matroids are the a bit more obscure accounts of geometrizations with regards to matrices.

What "linear" even is is usually enough "linear is linear".
Generally considered "ordinary" if through substitution.


I'm an anti-reductionist, yet though reduction is one
of the most usual results in closed categories, the
methods and techniques, point being closed categories
aren't allowed to close themselves, only being found so.





--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

On 08/02/2026 10:41 AM, Ross Finlayson wrote:

On 08/02/2026 09:34 AM, Johann 'Myrkraverk' Oskarsson wrote:

On 02/08/2026 8:47 AM, Mild Shock wrote:

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

Don't worry about it. There are several regulars here
who
don't
understand

that programming can be done for fun.

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

98 more questions to go, don't give up!

Here in sci.math, as everyone knows, I'm gearing up for
/linear algebra/ for fun. Still waiting for DVDs because
I'm not in a hurry. The book /Linear Algebra Done Right/
is interesting, and I've yet to go through the other rec-
commendations.[1]

I'm curious if you've ever thought of doing OpenGL with Prolog?

Does that even work?


[1] I have no idea how this word is supposed to be hyphenated,
I just do it anyway, because I'm not an LLM.

Bye

Take care!

You might have good luck looking up reputable university programs
and seeing what textbooks they require, these days.

Or, you know, just buy old ones when the library retires
the old good ones.

How about Householder's "The Theory of Matrices in Numerical Analysis".

Linear independence and linear spaces inevitably
get associated with vector spaces. There are much
simpler accounts though of reflections and rotations
about the determinantal and the singular and the decompositions
and the forms and the echelon forms and reduction with regards
to things like cumulants and orthogonants and the matroids,
vis-a-vis usual closed categories and so on.

The cumulants and orthogonants and so on are lesser-served
accounts of the earlier 20'th century, and determinantal analysis, while
the matroids are the a bit more obscure accounts of geometrizations with regards to matrices.

What "linear" even is is usually enough "linear is linear".
Generally considered "ordinary" if through substitution.

I'm an anti-reductionist, yet though reduction is one
of the most usual results in closed categories, the
methods and techniques, point being closed categories
aren't allowed to close themselves, only being found so.

Sometimes "linear independence" is better read as "linear dependence",
this is because words like "abstract" and "general" and "closed" and
"regular" and "ordinary" have inverses, matters of perspective and
projection, then about the difference from the "non", the "super",
for example the "classical".

"Truth is regular. Geometry is motion."


--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

On 8/1/2026 5:47 PM, Mild Shock wrote:

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

98 more questions to go, don't give up!

[...]

Fwiw, I have several compute shaders that do what I want. Mainly
building vector fields, etc.... And yes I use textures for some input
and output, uniforms mainly for the settings, etc. Just, make sure to
code things up to a point where your compute shader never needs to wait
for something... Think of striving for wait-free algorithms.

For instance, this is 100% wait free.

void add_hit(ct_plane2d plane, vec2 p, vec3 weight)
{
vec2 uv = ct_plane2d_unproject(plane, p);
ivec2 px = ivec2(uv * u_resolution);

if (px.x >= 0 && px.x < int(u_resolution.x) &&
px.y >= 0 && px.y < int(u_resolution.y))
{
imageAtomicAdd(accum_r, px, weight.r);
imageAtomicAdd(accum_g, px, weight.g);
imageAtomicAdd(accum_b, px, weight.b);
imageAtomicAdd(accum_hits, px, 1.0f);
}
}


Notice how I separated my accumulation buffer into different textures?

layout(binding = 0, r32f) uniform coherent image2D accum_r;
layout(binding = 1, r32f) uniform coherent image2D accum_g;
layout(binding = 2, r32f) uniform coherent image2D accum_b;
layout(binding = 3, r32f) uniform coherent image2D accum_hits; // alpha
/ hit counter

Works great and runs really fast.
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Hi,

I assure you I have like 3-4 times already
communicated to you that my requirements are
bounded queues. And not the ideally unbounded queues
that you are using, i.e. imageAtomicAdd.

Just check the postings in this forum. I have
like 3-4 times already specified that I need
bounded queues.

Works great and runs really fast.

You repeating yourself. Whats the motivation
of this spamming. I mean I can officially acknowledge
here that I have seen your imageAtomicAdd code

already. I also responded back then that I
have a Queue prototype that exactly uses that.
But it doesn't work for my purpose because I need:

- bounded queues that can block
- sizes are typically like 4-32 elements
- blocking is not done in GPU
- blocking is done in Hack
- Hack can do work stealing etc..

Because Hack can do a lot of tricks, you shouldn't
worry at all. Also spinning with backoff etc..
could be part of the picture, just check out:

Parallel Programming, Spring 2019, Lecture 16+1:
Spinlocks, Deadlocks, Semaphores https://spcl.inf.ethz.ch/Teaching/2020-pp/lectures/PP-l17-BeyondLocks.pdf

So just let me do my research, and refrain from
spamming me with always the same nonsense. Better
listen. I assure you I have like 3-4 times already

communicated to you that my requirements are
bounded queues. And not the ideally unbounded queues
that you are using., i.e. imageAtomicAdd.

Just check the postings in this forum. I have
like 3-4 times already specified that I need
bounded queues.

Bye


Chris M. Thomasson schrieb:

On 8/1/2026 5:47 PM, Mild Shock wrote:

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

98 more questions to go, don't give up!

[...]

Fwiw, I have several compute shaders that do what I want. Mainly
building vector fields, etc.... And yes I use textures for some input
and output, uniforms mainly for the settings, etc. Just, make sure to
code things up to a point where your compute shader never needs to wait
for something... Think of striving for wait-free algorithms.

For instance, this is 100% wait free.

void add_hit(ct_plane2d plane, vec2 p, vec3 weight)
{
-a-a-a vec2 uv = ct_plane2d_unproject(plane, p);
-a-a-a ivec2 px = ivec2(uv * u_resolution);

-a-a-a if (px.x >= 0 && px.x < int(u_resolution.x) &&
-a-a-a-a-a-a-a px.y >= 0 && px.y < int(u_resolution.y))
-a-a-a {
-a-a-a-a-a-a-a imageAtomicAdd(accum_r,-a-a-a px, weight.r);
-a-a-a-a-a-a-a imageAtomicAdd(accum_g,-a-a-a px, weight.g);
-a-a-a-a-a-a-a imageAtomicAdd(accum_b,-a-a-a px, weight.b);
-a-a-a-a-a-a-a imageAtomicAdd(accum_hits, px, 1.0f);
-a-a-a }
}

Notice how I separated my accumulation buffer into different textures?

layout(binding = 0, r32f) uniform coherent image2D accum_r;
layout(binding = 1, r32f) uniform coherent image2D accum_g;
layout(binding = 2, r32f) uniform coherent image2D accum_b;
layout(binding = 3, r32f) uniform coherent image2D accum_hits;-a // alpha
/ hit counter

Works great and runs really fast.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

You seem to need queues, why not "imageAtomicAdd"

I don't need ideally unbouded queues from WebGL

But the "imageAtomicAdd" are wait-free

I don't need wait-free queues, my queues should block

Hard to swallow, isn't it? Not my problem, its yours!

97 more questions to go, don't give up!

Bye

Mild Shock schrieb:

Hi,

I assure you I have like 3-4 times already
communicated to you that my requirements are
bounded queues. And not the ideally unbounded queues
that you are using, i.e. imageAtomicAdd.

Just check the postings in this forum. I have
like 3-4 times already specified that I need
bounded queues.

Works great and runs really fast.

You repeating yourself. Whats the motivation
of this spamming. I mean I can officially acknowledge
here that I have seen your imageAtomicAdd code

already. I also responded back then that I
have a Queue prototype that exactly uses that.
But it doesn't work for my purpose because I need:

- bounded queues that can block
- sizes are typically like 4-32 elements
- blocking is not done in GPU
- blocking is done in Hack
- Hack can do work stealing etc..

Because Hack can do a lot of tricks, you shouldn't
worry at all. Also spinning with backoff etc..
could be part of the picture, just check out:

Parallel Programming, Spring 2019, Lecture 16+1:
Spinlocks, Deadlocks, Semaphores https://spcl.inf.ethz.ch/Teaching/2020-pp/lectures/PP-l17-BeyondLocks.pdf

So just let me do my research, and refrain from
spamming me with always the same nonsense. Better
listen. I assure you I have like 3-4 times already

communicated to you that my requirements are
bounded queues. And not the ideally unbounded queues
that you are using., i.e. imageAtomicAdd.

Just check the postings in this forum. I have
like 3-4 times already specified that I need
bounded queues.

Bye

Chris M. Thomasson schrieb:

On 8/1/2026 5:47 PM, Mild Shock wrote:

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

98 more questions to go, don't give up!

[...]

Fwiw, I have several compute shaders that do what I want. Mainly
building vector fields, etc.... And yes I use textures for some input
and output, uniforms mainly for the settings, etc. Just, make sure to
code things up to a point where your compute shader never needs to
wait for something... Think of striving for wait-free algorithms.

For instance, this is 100% wait free.

void add_hit(ct_plane2d plane, vec2 p, vec3 weight)
{
-a-a-a-a vec2 uv = ct_plane2d_unproject(plane, p);
-a-a-a-a ivec2 px = ivec2(uv * u_resolution);

-a-a-a-a if (px.x >= 0 && px.x < int(u_resolution.x) &&
-a-a-a-a-a-a-a-a px.y >= 0 && px.y < int(u_resolution.y))
-a-a-a-a {
-a-a-a-a-a-a-a-a imageAtomicAdd(accum_r,-a-a-a px, weight.r);
-a-a-a-a-a-a-a-a imageAtomicAdd(accum_g,-a-a-a px, weight.g);
-a-a-a-a-a-a-a-a imageAtomicAdd(accum_b,-a-a-a px, weight.b);
-a-a-a-a-a-a-a-a imageAtomicAdd(accum_hits, px, 1.0f);
-a-a-a-a }
}


Notice how I separated my accumulation buffer into different textures?

layout(binding = 0, r32f) uniform coherent image2D accum_r;
layout(binding = 1, r32f) uniform coherent image2D accum_g;
layout(binding = 2, r32f) uniform coherent image2D accum_b;
layout(binding = 3, r32f) uniform coherent image2D accum_hits;-a //
alpha / hit counter

Works great and runs really fast.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

You seem to need queues, why not "imageAtomicAdd"

I don't need ideally unbouded queues from WebGL

But the "imageAtomicAdd" are wait-free

I don't need wait-free queues, my queues should block

Hard to swallow, isn't it? Not my problem, its yours!

96 more questions to go, don't give up!

Bye

Mild Shock schrieb:

Hi,

I assure you I have like 3-4 times already
communicated to you that my requirements are
bounded queues. And not the ideally unbounded queues
that you are using, i.e. imageAtomicAdd.

Just check the postings in this forum. I have
like 3-4 times already specified that I need
bounded queues.

Works great and runs really fast.

You repeating yourself. Whats the motivation
of this spamming. I mean I can officially acknowledge
here that I have seen your imageAtomicAdd code

already. I also responded back then that I
have a Queue prototype that exactly uses that.
But it doesn't work for my purpose because I need:

- bounded queues that can block
- sizes are typically like 4-32 elements
- blocking is not done in GPU
- blocking is done in Hack
- Hack can do work stealing etc..

Because Hack can do a lot of tricks, you shouldn't
worry at all. Also spinning with backoff etc..
could be part of the picture, just check out:

Parallel Programming, Spring 2019, Lecture 16+1:
Spinlocks, Deadlocks, Semaphores https://spcl.inf.ethz.ch/Teaching/2020-pp/lectures/PP-l17-BeyondLocks.pdf

So just let me do my research, and refrain from
spamming me with always the same nonsense. Better
listen. I assure you I have like 3-4 times already

communicated to you that my requirements are
bounded queues. And not the ideally unbounded queues
that you are using., i.e. imageAtomicAdd.

Just check the postings in this forum. I have
like 3-4 times already specified that I need
bounded queues.

Bye

Chris M. Thomasson schrieb:

On 8/1/2026 5:47 PM, Mild Shock wrote:

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

98 more questions to go, don't give up!

[...]

Fwiw, I have several compute shaders that do what I want. Mainly
building vector fields, etc.... And yes I use textures for some input
and output, uniforms mainly for the settings, etc. Just, make sure to
code things up to a point where your compute shader never needs to
wait for something... Think of striving for wait-free algorithms.

For instance, this is 100% wait free.

void add_hit(ct_plane2d plane, vec2 p, vec3 weight)
{
-a-a-a-a vec2 uv = ct_plane2d_unproject(plane, p);
-a-a-a-a ivec2 px = ivec2(uv * u_resolution);

-a-a-a-a if (px.x >= 0 && px.x < int(u_resolution.x) &&
-a-a-a-a-a-a-a-a px.y >= 0 && px.y < int(u_resolution.y))
-a-a-a-a {
-a-a-a-a-a-a-a-a imageAtomicAdd(accum_r,-a-a-a px, weight.r);
-a-a-a-a-a-a-a-a imageAtomicAdd(accum_g,-a-a-a px, weight.g);
-a-a-a-a-a-a-a-a imageAtomicAdd(accum_b,-a-a-a px, weight.b);
-a-a-a-a-a-a-a-a imageAtomicAdd(accum_hits, px, 1.0f);
-a-a-a-a }
}


Notice how I separated my accumulation buffer into different textures?

layout(binding = 0, r32f) uniform coherent image2D accum_r;
layout(binding = 1, r32f) uniform coherent image2D accum_g;
layout(binding = 2, r32f) uniform coherent image2D accum_b;
layout(binding = 3, r32f) uniform coherent image2D accum_hits;-a //
alpha / hit counter

Works great and runs really fast.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Hi,

Chris M. Thomasson schrieb:

Strive to never make a compute shader wait
on something, like an empty condition of a queue, stack.

You are such a moron. GPU elasticity was
already invented in 2008 with CUDA. I posted
this quote already:

"CUDArao TEChNOLOGY UNLOCkS ThE
POWER OF TESLA MANY-CORE PROCESSORS

The CUDA C compiler simplifies many-core
programming by enabling code development
in a high-level language and optimizing code
to run on systems without knowledge of how

many cores are in the hardware. CUDA
applications automatically take advantage of
more cores or fewer cores in a system, so
they can scale from entry-level notebook

GPUs to high end GPUs in technical workstations
and further into racks of GPUs in data centers.
This allows developers to rCLcode oncerCY and
deploy on a range of systems, as well as scale

forward in time as future GPUs deliver more
performance per watt and more cores per
processor. The benefit for software users is
the opportunity to boost computing performance

simply by adding GPUs or using their existing GPUs in new ways" https://www.nvidia.com/docs/IO/43395/NV_DS_Tesla_S1070_US_Jun08_NV_LR_Final.pdf

Today elasticity is on logical thread aka task level,
not only on "core" level or something. Don't know
exactly what CUDA did back them, maybe only

a submit elasticity, like a time sharing system. Today
you have quite some run elasticity on modern machines,
for your logical threads. Even in budget laptops

like a Ryzen AI 7 350 /w Radeon 850M.

Bye

P.S.: I can demostrate the elasticity, but I didn't
write the medium.com article yet.

Mild Shock schrieb:

Hi,

Ok, following the instructions here:

npm install webgpu
https://github.com/dawn-gpu/node-webgpu

I can now run webgpu also from CLI:

node.exe dogelog.mjs

Dogelog Spieler 2.2.5, Node, JavaScript 26.4.0
(c) 1985-2026, XLOG Technologies AG, Schweiz

?- ensure_loaded(library(edge/furryhaze)).
true.

?- between(1,3,_), time(expedite((between(1,100,_),
between(1,100,_), between(1,100,_)), [size(4096)])), fail.
% Zeit 1037.994 ms, GC 0.000 ms, Lips 111 k
% Zeit 1091.131 ms, GC 0.000 ms, Lips 106 k
% Zeit 1045.274 ms, GC 0.000 ms, Lips 110 k
fail.

Same benchmark result as in the browser.
Now I can rent a bigger GPU by the hour
and do some easy CLI testing.

LoL

Bye

Mild Shock schrieb:

Hi,

Chris M. Thomasson can ask 100 more questions.
I will happily answer them. But maybe I should
make a Wiki to explain the ever same things:

But, I still don't know what you main goal is?

The goal is "Prolog inferencing"

It has textures to work with in the pipeline.

I don't need textures for "Prolog inferencing"

98 more questions to go, don't give up!

Bye

Mild Shock schrieb:

Hi,

Tablets and phone are more annoying to
use with WebGPU. The usual browsers don't
have a Chrome DevTools panel integrated,

so that one could do JavaScript Debugging
directly on the device. Instead one has to
use a desktop machine, and connect the

device via UBS-C , and start a Chrome
Browser there . And then start a Chrome
DevTools panel alone, that is pair with

the device, via UBS-C cable. So this way
I already see where it crashes on the
tablets and phone:

await output.mapAsync(GPUMapMode.READ)
Unhandled Promise Rejection: OperationError

The above is the error that one can re-produce
already here with this test:

11.4 Giga Lips with a Budget Laptop
https://github.com/Jean-Luc-Picard-2021/gigabudget

Not sure what exactly happens. Maybe
a form of timeout or device lost, that the
primitive HTML / JavaScript doesn't handle

gracefully yet. Maybe redimensioning the
test, so that it consumes less time would
help. Who knows? Will see. For production

use of a GPU integration I have to anyway
provide work slicing it seems.

Bye

Mild Shock schrieb:

Hi,

Why does this Lama have a red pyjama.
Oh, its a baby Lama. Its still in the cradle
and needs some training:

RedPajama-Data-v2
https://github.com/togethercomputer/RedPajama-Data

But then Andrej Karpathy recently showed
GPT-2 training on rented GPUs for less
than 100 USD in less then 2 hours.

So where do these grown up Lamas go.
Well Georgi Gerganov prefered C++/C
when he shouted Llama Llama Red Pyjama.

But you also find WebLLM, wrapping the
underlying C++/C GPU interface via the
W3C standard WebGPU / WGSL, with JavaScript:

In-Browser LLM Inference Engine
https://webllm.mlc.ai/

My experience with WebLLM 6 months
ago on an iPad Pro 2024, still a little early
stage performance and robustness.

But hey hardware of AI mobile iGPUs is
still evolving, and AI laptop, AI smartphones
and AI tablets, will soon feature Chinese

hardware such some new Kirin AI in 2027.

Bye

Mild Shock schrieb:

Hi,

Remember when first all local AI was Python
and PyTorch APIs. And then suddently people strated
using bare metal C/C++ Code. Here is the story:

How it started:

GPT-J or GPT-J-6B is an open-source large
language model (LLM) developed by EleutherAI
in 2021. As the name suggests, it is a
generative pre-trained transformer model
designed to produce human-like text that
continues from a prompt.
https://www.eleuther.ai/

How it was going [Georgi Gerganov]:

So a few days later comes out the LLaMA, I do
some calculations and I figure out rCLOkay, 65
billion parameters. You probably need about
40 gigs of RAM, with 4-bit quantization. So
this can run on a MacBook. Why not do it?rCY

Why I was able to do it so quickly - basically,
for all that I saw itrCOs pretty much GPT-J architecture
with some modifications, like some extra memorization
layers. ItrCOs minor changes. Basically, again, the
existing code for the GPT-J, I just simply
modified it there, it happened pretty quickly.
https://changelog.com/podcast/532

Georgi Gerganov, Bulgarian, now with Hugging
Face, ggml-cann also running on Chinese AI chips.
ggml Manifesto https://github.com/ggml-org/ggml

Bye

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

On 8/2/2026 2:09 PM, Mild Shock wrote:
[...]

Good bye.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Hi,

You give up, the rat your are.

Liar and spammer.

Bye

Chris M. Thomasson schrieb:

On 8/2/2026 2:09 PM, Mild Shock wrote:
[...]

Good bye.

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: sci.physics.relativity

Mild Shock wrote:

Hi,

You give up, the rat your are.

Liar and spammer.

you are so stupid you cant even fit a curve mathematically
--- Synchronet 3.22a-Linux NewsLink 1.2