gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[OPTIMIZER] Made layout simplification pass efficient for fused attention kernels (#790)

Browse Source
This commit is contained in:
Philippe Tillet
2022-10-21 16:52:15 -07:00
committed by GitHub
parent c4726333bf
commit bb0f9235d1
26 changed files with 683 additions and 229 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
lib/Dialect/Triton/Transforms/Combine.td
View File

@@ -12,21 +12,21 @@ include "triton/Dialect/Triton/IR/TritonOps.td"
// AddIOp(d, DotOp(a, b, c)) and c==0 => DotOp(a, b, d)
// AddFOp(d, DotOp(a, b, c)) and c==0 => DotOp(a, b, d)
def CombineDotAddIPattern : Pat<
(Arith_AddIOp $d, (TT_DotOp:$res $a, $b, $c, $allowTF32)),
(TT_DotOp $a, $b, $d, $allowTF32),
(Arith_AddIOp $d, (TT_DotOp:$res $a, $b, $c, $allowTF32, $transA, $transB)),
(TT_DotOp $a, $b, $d, $allowTF32, $transA, $transB),
[(Constraint<CPred<"isZero($0)">> $c)]>;
def CombineDotAddFPattern : Pat<
(Arith_AddFOp $d, (TT_DotOp:$res $a, $b, $c, $allowTF32)),
(TT_DotOp $a, $b, $d, $allowTF32),
(Arith_AddFOp $d, (TT_DotOp:$res $a, $b, $c, $allowTF32, $transA, $transB)),
(TT_DotOp $a, $b, $d, $allowTF32, $transA, $transB),
[(Constraint<CPred<"isZero($0)">> $c)]>;
def CombineDotAddIRevPattern : Pat<
(Arith_AddIOp (TT_DotOp:$res $a, $b, $c, $allowTF32), $d),
(TT_DotOp $a, $b, $d, $allowTF32),
(Arith_AddIOp (TT_DotOp:$res $a, $b, $c, $allowTF32, $transA, $transB), $d),
(TT_DotOp $a, $b, $d, $allowTF32, $transA, $transB),
[(Constraint<CPred<"isZero($0)">> $c)]>;
def CombineDotAddFRevPattern : Pat<
(Arith_AddFOp (TT_DotOp:$res $a, $b, $c, $allowTF32), $d),
(TT_DotOp $a, $b, $d, $allowTF32),
(Arith_AddFOp (TT_DotOp:$res $a, $b, $c, $allowTF32, $transA, $transB), $d),
(TT_DotOp $a, $b, $d, $allowTF32, $transA, $transB),
[(Constraint<CPred<"isZero($0)">> $c)]>;