Matrix and vector operations need to be updated using Float32x4 about three.dart HOT 15 CLOSED

@unicomp21 Take a look at https://github.com/nelsonsilva/three.dart/blob/simd/lib/src/core/SIMDVector.dart
I've been benchmarking the webgl_performance example but still no performance improvement ...

from three.dart.

Is there a link for the example? My guess is that it's GPU intense, as
opposed to Float32x4, which is CPU intense. Am I right?
On May 4, 2013 9:42 AM, "Nelson Silva" [email protected] wrote:

@unicomp21 https://github.com/unicomp21 Take a look at
https://github.com/nelsonsilva/three.dart/blob/simd/lib/src/core/SIMDVector.dart
I've been benchmarking the webgl_performance example but still no
performance improvement ...

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/60#issuecomment-17434247
.

from three.dart.

Nelson, I missed the link earlier, nice job! You the man!

On Sat, May 4, 2013 at 9:42 AM, Nelson Silva [email protected]:

@unicomp21 https://github.com/unicomp21 Take a look at
https://github.com/nelsonsilva/three.dart/blob/simd/lib/src/core/SIMDVector.dart
I've been benchmarking the webgl_performance example but still no
performance improvement ...

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/60#issuecomment-17434247
.

from three.dart.

For what it's worth, this implementation opens the doors wide for
procedural geometry/texture generation on the CPU. (ie ray tracing of
static textures where the instruction count would be too high for the gpu,
procedural geometry generation using isosurfaces, the list goes on ...)

Another exciting application is offline rendering for Hollywood style
animation. Using your vector implementation, and a few thousand web
browsers collaborating, some very cool stuff could be done.

On Sun, May 5, 2013 at 6:06 AM, John Davis [email protected] wrote:

Nelson, I missed the link earlier, nice job! You the man!

On Sat, May 4, 2013 at 9:42 AM, Nelson Silva [email protected]:

@unicomp21 https://github.com/unicomp21 Take a look at
https://github.com/nelsonsilva/three.dart/blob/simd/lib/src/core/SIMDVector.dart
I've been benchmarking the webgl_performance example but still no
performance improvement ...

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/60#issuecomment-17434247
.

from three.dart.

Are there any unit tests yet?

from three.dart.

@unicomp21 we started experimenting with DRT to get at the least the samples running on our CI setup but never managed to get it working properly.

from three.dart.

DRT, CL ... I'm a bit lost.
On May 7, 2013 4:04 AM, "Nelson Silva" [email protected] wrote:

@unicomp21 https://github.com/unicomp21 we started experimenting with
DRT to get at the least the samples running on our CI setup but never
managed to get it working properly.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/60#issuecomment-17531207
.

from three.dart.

DRT <=> DumpRenderTree (see http://trac.webkit.org/wiki/Writing%20Layout%20Tests%20for%20DumpRenderTree)

CI <=> Continuous Integration (we're using drone.io)

from three.dart.

This is in reference to the new SIMDVector implementation?

On Tue, May 7, 2013 at 6:00 AM, Nelson Silva [email protected]:

DRT <=> DumpRenderTree (see
http://trac.webkit.org/wiki/Writing%20Layout%20Tests%20for%20DumpRenderTree
)

CI <=> Continuous Integration (we're using drone.io)

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/60#issuecomment-17535748
.

from three.dart.

@unicomp21 "Are there any unit tests yet?"
@nelsonsilva "... DRT...CI..."

from three.dart.

Nelson, I think it would be worthwhile to give your SIMDVector it's own git
repo. As far as I know, nobody has created one yet, and it has great
utility for many applications. We need something like Sylvester (vector
math for javascript), but rather for dart w/ SIMD. Heaven knows when
mccutchan will get around to updating the vector_math module, I think he's
actually painted himself into a bit of a corner with the code-gen framework.

On Tue, May 7, 2013 at 12:27 PM, Nelson Silva [email protected]:

@unicomp21 https://github.com/unicomp21 "Are there any unit tests yet?"
@nelsonsilva https://github.com/nelsonsilva "... DRT...CI..."

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/60#issuecomment-17556897
.

from three.dart.

Starting to realize there's more to the problem. The trick is keeping the
pipeline loaded, which makes things a bit more complex. To speed up
something like cross product there needs to be large blocks/arrays (not
larger than l1 cache) and then multiply/accumulate operations are done at
the block level (which keeps the pipeline loaded).

On Sat, May 4, 2013 at 9:42 AM, Nelson Silva [email protected]:

@unicomp21 https://github.com/unicomp21 Take a look at
https://github.com/nelsonsilva/three.dart/blob/simd/lib/src/core/SIMDVector.dart
I've been benchmarking the webgl_performance example but still no
performance improvement ...

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/60#issuecomment-17434247
.

from three.dart.

And to do what I just mentioned, we need swizzle support, which John M
hasn't added yet.

TODO(johnmccutchan): Add all 256 possible combinations.

typed_data.dart

On Sat, Jun 1, 2013 at 8:03 AM, John Davis [email protected] wrote:

Starting to realize there's more to the problem. The trick is keeping the
pipeline loaded, which makes things a bit more complex. To speed up
something like cross product there needs to be large blocks/arrays (not
larger than l1 cache) and then multiply/accumulate operations are done at
the block level (which keeps the pipeline loaded).

On Sat, May 4, 2013 at 9:42 AM, Nelson Silva [email protected]:

@unicomp21 https://github.com/unicomp21 Take a look at
https://github.com/nelsonsilva/three.dart/blob/simd/lib/src/core/SIMDVector.dart
I've been benchmarking the webgl_performance example but still no
performance improvement ...

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/60#issuecomment-17434247
.

from three.dart.

Yep. I read some papers by Intel and most of the gain comes from batching
ops. So you won't have vectors as float32x4 but use lists each with one of
the coordinates of a couple of vectors then you can really take advantage
of simd.
No dia 01/06/2013 14:03, "John Davis" [email protected] escreveu:

Starting to realize there's more to the problem. The trick is keeping the
pipeline loaded, which makes things a bit more complex. To speed up
something like cross product there needs to be large blocks/arrays (not
larger than l1 cache) and then multiply/accumulate operations are done at
the block level (which keeps the pipeline loaded).

On Sat, May 4, 2013 at 9:42 AM, Nelson Silva [email protected]:

@unicomp21 https://github.com/unicomp21 Take a look at

https://github.com/nelsonsilva/three.dart/blob/simd/lib/src/core/SIMDVector.dart
I've been benchmarking the webgl_performance example but still no
performance improvement ...

—
Reply to this email directly or view it on GitHub<
https://github.com/threeDart/three.dart/issues/60#issuecomment-17434247>
.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/60#issuecomment-18788739
.

from three.dart.

FYI, asked for them to fix/add support for swizzle/reorder, which will allow us to make fast cross products, etc.

https://code.google.com/p/dart/issues/detail?id=11021

from three.dart.