d783d999de
minor
2021-12-05 01:49:09 -05:00
6ea4a343fd
Fix upstream build
2021-12-01 14:51:48 -05:00
a1224cc23f
Add a polygeist-opt tool
...
This commit adds a `polygeist-opt` command line tool. This allows us to drive each of the polygeist passes individually and as such lays the foundation for adding tests of the individual passes.
In doing so, a fair bit of restructuring/modifications of the various namespaces, function names etc. employed in the passes have been done. This was mostly due to inconsistensies in symbol names and namespaces which weren't being caught as errors due to missing inclusion of pass registration.
2021-11-29 11:14:19 -05:00
f579f6ccc2
Fix lowering
2021-11-16 13:03:41 -06:00
e90f4525d3
Add LLVM recursive lowering
2021-11-16 13:03:41 -06:00
06d2e89a7d
Fix build on clang
2021-11-16 13:03:41 -06:00
4c2323a4c0
fix clang-format
2021-11-12 11:53:04 +01:00
56c878e72d
Update llvm
2021-11-11 10:40:08 -05:00
02bd12f11a
Bump layout api change
2021-11-11 10:40:08 -05:00
90c964a30f
Cleanup
2021-10-17 10:05:47 -04:00
b715377690
Handle return
2021-10-17 10:05:47 -04:00
d832458d9b
Ensure packed
2021-10-17 10:05:47 -04:00
fc19797b2a
Fix capture mem2reg
2021-10-17 10:05:47 -04:00
26bcbec66a
Bump LLVM
2021-10-17 10:05:47 -04:00
5e918f7e52
Enable nested constructors, and introduce pointer2memref ( #91 )
...
* Pointer 2 Memref handling
* Add tests
* Fix mem2reg
2021-10-15 13:32:16 -04:00
1b373bbdb5
cleanup
2021-10-12 23:49:21 -04:00
b8d3f38614
Handle undef initialization
2021-10-12 23:49:21 -04:00
1bd3675da5
Handle undef initialization
2021-10-09 13:16:04 -04:00
cff309032c
Fix format
2021-09-28 00:21:32 -04:00
972591e103
Correct parallel lower and related codegen
2021-09-28 00:21:32 -04:00
1f07d68b2a
no region
2021-09-28 00:21:32 -04:00
7a37ea7f00
Don't inline empty region
2021-09-28 00:21:32 -04:00
0778c08843
Correct cudaMemcpy
2021-09-28 00:21:32 -04:00
106ed7a513
[CMake] Add MLIRNVVMIR to LINK_LIBS for MLIRPolygeistTransforms.
...
This dependency is necessary, e.g. when using BUILD_SHARED_LIBS.
2021-09-24 10:13:39 -04:00
18f319c57e
Silence warnings (NFC) ( #80 )
2021-09-23 15:41:54 +02:00
9619d130ee
Cleanup code
2021-09-16 10:58:17 -04:00
199c1e7d1e
Fix tests
2021-09-16 10:58:17 -04:00
ec4212e126
Fix use-after-free in LoopRestructure
...
An op was being recreated (i.e., deleted and a new op created) in the
loop iterating its successors. Store the new op to keep visiting
successors.
2021-09-16 10:58:17 -04:00
4de0646d4d
Fixes to barrier continuation
2021-09-16 10:58:17 -04:00
acc2bd3f83
make continuation-style barrier removal support nested parallel
...
Only the innermost scf.parallel is allowed to contain barrier though.
Otherwise, we need a mechanism to recursively create local CFGs and
inline them into parent regions.
2021-09-16 10:58:17 -04:00
b570c7b115
skip over function declarations in continuation barrier removal
2021-09-16 10:58:17 -04:00
09fa8837bc
add files
2021-09-16 10:58:17 -04:00
336bc9b81c
Canonicalize for
2021-09-16 10:58:17 -04:00
6bc6937fe8
select of subindex
2021-09-16 10:58:17 -04:00
530597cfb7
Add cast of subindex canonicalization
2021-09-16 10:58:17 -04:00
eade58498b
Fix affine subindex bug
2021-09-16 10:58:17 -04:00
2f73a8622e
Only consider immediately-nested barriers in hasNestedBarrier
...
Barriers are active on the first ancestor ParallelOp. When looking for
nested barriers, ignore those that are nested inside ParallelOps, which
are in turn nested in the operation of interest as they are active on
different loops.
2021-09-16 10:58:17 -04:00
4123d514ea
Adding missing file
2021-09-16 10:58:17 -04:00
1a1a1955b6
Fixup lower
2021-09-16 10:58:17 -04:00
8b95da86a9
Fix parallel loop distribute
2021-09-16 10:58:17 -04:00
5799e8f124
Symbol linkage fixes
2021-09-16 10:58:17 -04:00
d6c7fe0d55
Simplify loop distribute
2021-09-16 10:58:17 -04:00
4f59718e85
Custom cuda lowering
2021-09-16 10:58:17 -04:00
ba28812f58
Bump LLVM and GPU fixes
2021-09-16 10:58:17 -04:00
afcba4add7
Add some extra LINK_LIBS that were missing.
...
These were uncovered during a development build with
BUILD_SHARED_LIBS=ON.
2021-08-24 14:57:54 -04:00
70b9d5e8a7
cleanup
2021-08-19 11:32:05 -04:00
9b9e755f39
Fix tuple constructs
2021-08-19 11:32:05 -04:00
df77398f80
Handle switch in mem2reg
2021-08-19 11:32:05 -04:00
8fb9ac6429
Bump LLVM 8c50b5fbfef113157d35f74b95c959c20eb37fab
2021-08-16 18:46:26 -04:00
e9b551c1e7
Use LowerTo pragma to generate a linalg.copyOp
...
- Refactor pragma lower to handling
2021-07-20 09:20:31 +02:00
William S. Moses