Comments (6)
DrTimothyAldenDavis
commented on August 14, 2024
1
Yes, and it's made worse for the sparse case. For dense matrix
multiplication, it's easy to tile the matrix and get a better ratio of
flops per memory read/write (and so I would never try to get that loop in
the post to go fast for the dense case). That's harder to do for the
sparse case, particularly in a large complex library like GraphBLAS where I
have to support lots of operations on a consistent data structure.
The solution would be to break up my matrices into smaller submatrices ...
but then I have to make that data structure play nice and mix/match with
any other matrix in any other operation. It's on my distant TODO list.
…On Sat, Nov 21, 2020 at 2:46 PM Victor Stewart ***@***.***> wrote:
here's another link to benchmarks of SIMD-ing a bunch of linear algebra
operations, and showing only the slightest of benefits. (obviously i think
it's safe to assume for the same memory starvation reason).
https://github.com/rxwp5657/Clutch#benchmark-results-simd-disabled
so this doesn't seem like a road worth pursuing current hardware as it
stands. maybe in the coming years though?
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#21 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AEYIIONCET5YNNG7WEQPJNLSRARLZANCNFSM4T44QLFA>
.
from graphblas.
DrTimothyAldenDavis
commented on August 14, 2024
Yes, eventually. I already tried but wasn’t able to speed it up. There are
some AVX instructions in the drafts in GitHub but they are disabled.
On Fri, Nov 20, 2020 at 9:10 AM Victor Stewart ***@***.***> wrote:
just out of pure curiosity, do you have any plans to vectorize GraphBLAS?
(I saw at least GB_DOT_SIMD but couldn't locate any usages).
on that topic, avx512 will be closer to penalty-less come Ice Lake.
and of course there are inbound matrix instructions as well. (who knows
the chip wide downscaling penalties yet though).
[image: IMG_2285]
<https://user-images.githubusercontent.com/17534897/99815412-2b58ae00-2b42-11eb-992b-e74f14ba4573.jpeg>
^^ from this broadcast:
https://sc20.gallery.video/detail/videos/sc20/video/6208439527001/a-closer-look-at-ice-lake-architecture-overview
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#21>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AEYIIOP4YWYW5ZC3GK5HUJTSQ2BETANCNFSM4T44QLFA>
.
--
Sent from Gmail Mobile
from graphblas.
victorstewart
commented on August 14, 2024
i'll take a look. maybe a breakthrough will come to me.
i assume those ones you attacked were the least performant as scalar / would benefit the most by being vectorized?
if not maybe a brief list would be a helpful lay of the land.
from graphblas.
DrTimothyAldenDavis
commented on August 14, 2024
See this code, for C=A'*B when using the dot product method (assuming A and
B are both CSC; or C=A*B' if both CSR):
https://github.com/DrTimothyAldenDavis/GraphBLAS/blob/6d7fcea6093f264e16343918e61737eb4c66e3fa/Source/Template/GB_AxB_dot_cij.c#L399
The code is there in the stable v3.3.3 but disabled because it was no
faster than the conventional code, at this line:
https://github.com/DrTimothyAldenDavis/GraphBLAS/blob/6d7fcea6093f264e16343918e61737eb4c66e3fa/Source/Template/GB_AxB_dot_cij.c#L702
…On Sat, Nov 21, 2020 at 7:15 AM Victor Stewart ***@***.***> wrote:
i'll take a look. maybe a breakthrough will come to me.
i assume those ones you attacked were the least performant as scalar /
would benefit the most by being vectorized?
if not maybe a brief list would be a helpful lay of the land.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#21 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AEYIIOJJHCUODR2JPIVK5Y3SQ64O5ANCNFSM4T44QLFA>
.
from graphblas.
victorstewart
commented on August 14, 2024
this blog post is super informative, especially @travisdowns 's comment which i linked directly to here.
TLDR; it's basically a memory starvation problem.
from graphblas.
victorstewart
commented on August 14, 2024
here's another link to benchmarks of SIMD-ing a bunch of linear algebra operations, and showing only the slightest of benefits. (obviously i think it's safe to assume for the same memory starvation reason).
https://github.com/rxwp5657/Clutch#benchmark-results-simd-disabled
so this doesn't seem like a road worth pursuing current hardware as it stands. maybe in the coming years though?
from graphblas.
Related Issues (20)
- Consider adding COLEQ and ROWEQ IndexUnaryOp operators HOT 9
- New unary operators to calculate principal cube root of real (floating point) values HOT 5
- Matrix_extractElement_Structural HOT 6
- Pass a print function for UDTs HOT 4
- GrB_Descriptor_set is unable to set expected descriptor fields (does not match GxB_Desc_set) HOT 4
- Set name of UDT when serializing HOT 12
- Build broken with spaces in folder names HOT 8
- build fails on windows 11 using CMake with MinGW HOT 5
- ISEQ monoids HOT 7
- Fix bitwise operator monoid names HOT 4
- Remove va_arg HOT 5
- GrB_Vector_(de)serialize HOT 6
- Sparse Index Space HOT 2
- Removed symbols without soname bump HOT 12
- atomic*: Undefined symbol on armel and mipsel architectures HOT 33
- "ZEROB" Binary Operator HOT 3
- GxB sort with smaller (or larger) output objects
- cpu_features: Build error for MinGW HOT 13
- Size of Static Library HOT 16
- Link error with Intel igx and Ninja generator on Windows HOT 9
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from graphblas.