Dev dispatch #6

imrn99 · 2023-11-09T14:05:45Z

First part of the CPU dispatch code

TODO:

~~Check for a better kernel type using Box<dyn KernelArgs> as argument instead of an enum.~~ Since Kokkos uses templating, the static enum seems like the most coherent choice.
Change view implem to make it possible to write immutable kernels using thread-safe data type & methods
Add std::threads support
Add feature documentation

For the next part

Some benches
MDRange & tiling support
~~Team-based policies~~ For now, the focus will be on tiling.

should upcasting ever be stabilized, this version may compile

kernels are written in a Box, using an enum as argument this comes at a ~20% perf cost according to the mdrange_populate bench

need to write a bench with a simple parallel for to make sure the feature works

these warnings are technically erroneous since the auto fix creates errors

there is indeed a 2* speedup when using `--features rayon`

perf significantly worsen past the size of my L1 caches

kernel dispatch using raw pointers produces sigsegv :)

impl trait in type aliases have yet to be stabilized, might switch to nightly

its just bloat at this point

imrn99 added 30 commits October 31, 2023 10:31

modified functor traits to handle backedn specific closures

03b49ae

should upcasting ever be stabilized, this version may compile

Serial code compiles again

dcd0f2a

kernels are written in a Box, using an enum as argument this comes at a ~20% perf cost according to the mdrange_populate bench

rayon features compiles

9a74062

need to write a bench with a simple parallel for to make sure the feature works

moved imports inside test functions to fix warnings

34a3412

these warnings are technically erroneous since the auto fix creates errors

added feature bench to check beckend usage

ea58c51

there is indeed a 2* speedup when using `--features rayon`

docs

be5401c

docs again

70aa882

update readme

5ff96cd

new PartialEq impl for view data

11d452a

added a raw_val method to views use for test assertions

529f3bc

moved flat index computation to an inline function

b1b2179

added feature-specific writing interface to views

77d558f

update mirroring function

1543e5c

finished universal writing interface

99e3de7

updated feature bench; automatic kernel typing now works!

6da57c0

small doc update

6c7de73

modified layout bench to operate of a range of data sizes

320266f

perf significantly worsen past the size of my L1 caches

new ForKernel type for std::thread usage

1091fa9

better kernel type definition

bc555a1

added cfg_if in dispatch code

8cac955

added basic code for thread backend; currently doesn't work

a7eb044

indices dispatch ok; need to figure out how to share the kernel

3cbe4f9

kernel dispatch ok; need to fix lifetime issues

588503b

no more static lifetime using std threads;

74f4cf9

kernel dispatch using raw pointers produces sigsegv :)

small fixes

47759e0

cloning kernel works with Box<impl ...>;

6648595

impl trait in type aliases have yet to be stabilized, might switch to nightly

gated gpu impl behind a specific feature

6042d3b

removed DEPTH const generics from parallel_for method

aac6360

its just bloat at this point

parallel_for works with all features!

e2397c3

feature docs

6fe1342

imrn99 merged commit 7544eb1 into master Nov 17, 2023
4 checks passed

imrn99 deleted the dev_dispatch branch November 17, 2023 09:29

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Dev dispatch #6

Dev dispatch #6

imrn99 commented Nov 9, 2023 •

edited

Loading

Dev dispatch #6

Dev dispatch #6

Conversation

imrn99 commented Nov 9, 2023 • edited Loading

imrn99 commented Nov 9, 2023 •

edited

Loading