Rewrite Index Tree Dialect #79

AK2000 · 2025-01-15T01:30:52Z

Over the past year, we undertook nearly a complete rewriting of the index tree dialect in order enable future analysis and transformation. These changes included:

Reversing the structure of the index "tree" so nodes higher in the conceptual tree also occur higher in the IR, and "leaf" nodes in the tree occur at the bottom. Not only does this more naturally match the intution of the tree, it also enables properties of the tree to be enforced by the IR. Specifically, each IntexNodeOp only accepts a single "parent" index

…pace transformation pass

…ring passes

Decouple semiring attribute from domain inference

Fix triangle counting example

…runner. Also updated README.md to reflect this change

…umbers

…ger type (i32,i64) instead of index, to allow the frontends choose how they want to represent them. #69

Also, fixed test to correctly reflect its name

…header to the files, reorganize testcases)

…ized

Fix masking in index tree rewrite

* [WIP] Lowering dense domain to parallel loop * [WIP] fixing errors when building with parallel op * [WIP] Limited parallelism working on spmm CSRxDense case * Trying to fix indexing into tensor slices * Removing parallelism for sparse output. Fixing problems with dense parallelism * Fixing issue with COO and parallel loops

AK2000 added 30 commits October 14, 2024 17:47

Begin working on IndexTree transformations

4dac7bc

Fixing some of the problems introduced on merge

20f916a

Resolved including of device mapping attribute

864c9f8

Fixed type inclusion, parsing and printing

0447561

V1 - Lower TA to new IndexTree ops, but removed everything else

cf505bd

Fixes to TA to change how file is included

3f4e849

Creating new block for index tree

d6bc7fd

Implement domain inference pass, fix to index ordering

48fe5fd

[WIP] Fragile version of index tree to SCF lowering

2d6ba90

Fix carrying tensors inside loop, refactor domain concretization

972ee7e

Adding TA to index tree patterns for elementwise operations

c29ee1f

[WIP] Trying to implement intersection op lowering

8d44653

[WIP] Got domain intersection working, but only with dense output

f25fb52

[WIP] Minor fix to ordering of reduce args

35cf1c5

[WIP] Beginning support for sparse output tensors with new index tree

04b31ba

[WIP] Inlined itree op, got hacky version of removing set op working

1c06a26

[WIP] Included lowering to LLVM, lowering print op does not work

00801b1

[WIP] Almost got print op lowering working

352da6a

[WIP] Fixed bufferization

9436939

[WIP] Generate symbolic pass for sparse tensor declarations

ac98adf

[WIP] Lots of changes for first try at symbolic domain pass and works…

935079d

…pace transformation pass

[WIP] Broke everythink trying to redo tensor conversion infrastructure

0c2d98d

Changing alot to create new sparse tensor types, and appropriate lowe…

bb32d70

…ring passes

Fixing some problems with tests, ad pure ops

ac27cfb

Fixed inconsistencies in test suite

52495d4

Fixing more of the test cases

27dd373

Fixed dense transpose and print elapsed time

8bc6187

Fixing errors in typing

3f777dd

Fixing errors in typing and set op

8cb0d2a

Adding back ttgt pass

dc0e445

AK2000 and others added 30 commits November 5, 2024 16:24

"fixed" triangle counting test :)

5b66b30

Merge pull request #72 from AK2000/fix-semiring-tests

6efa6e3

Decouple semiring attribute from domain inference

Merge pull request #73 from AK2000/fix-triangle-counting

e4a987a

Fix triangle counting example

[COMETPY] Fixed issue with COO formats

880c9cc

[COMETPY] Update tests to make use of pytest and removed custom test …

de6469d

…runner. Also updated README.md to reflect this change

[COMETPY] Added data directory to tests

5cfd397

[COMETPY] Added missing file

bfd5993

Added support for f32 data #69

f4fe43a

Updated numpy-scipy side to support f32 #69

836f8ce

Updated SparseTensor type to use the respective enum instead of raw n…

664a7df

…umbers

Sparse tensor indices crd(_tile),pos(_tile) now have an explicit inte…

878c711

…ger type (i32,i64) instead of index, to allow the frontends choose how they want to represent them. #69

[COMETPY][WIP] Enabling i32 indices for sparse matrices

89ee8c4

Fixed bugs in Index to Integer casts and vice-versa

001cf29

[COMETPY] Finished support for i32,i64 sparse matrix indices.

6aacaa5

Also, fixed test to correctly reflect its name

[WIP] reviewing new index tree dialect and related test cases (added …

978b77c

…header to the files, reorganize testcases)

blis_interface is updated - debug prints removed

ac744a9

Bug is fixed in the dense transpose optimization. Testcases are organ…

304f20e

…ized

Fixed path in CMakeLists.txt

4433b75

Changing how index tree to SCF conversion pass is structured

33621c7

[WIP] adding semiring zero

5557cb3

[WIP] Symbolic pass and workspace transform not working

4176e61

Fixed issues with workspace transformation and symbolic pass

814240f

Resolved conflicts with restructure SCF conversion

5112041

[COMETPY] Updated sparse tensor representation

a25befd

[WIP] Got masking partially working but broke other things

4253d33

[WIP] Fixed other problems, but still issues with Sandia_LL with masking

b93b044

Helped compilation for SandiaLL w Masking

d3551f4

Merge pull request #76 from AK2000/rewrite_masking

28f5cf7

Fix masking in index tree rewrite

Rearranging comet.cpp to emit it at different stage

b0fa0ae

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Rewrite Index Tree Dialect #79

Rewrite Index Tree Dialect #79

AK2000 commented Jan 15, 2025

Rewrite Index Tree Dialect #79

Are you sure you want to change the base?

Rewrite Index Tree Dialect #79

Conversation

AK2000 commented Jan 15, 2025