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

[OPTIMIZER] layout simplification: ignore non-tensor iter arguments in for loop rematerialization (#749)

Browse Source
This commit is contained in:
Philippe Tillet
2022-10-07 21:52:29 -07:00
committed by GitHub
parent e843257295
commit 498c685b46
1 changed files with 2 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
lib/Dialect/TritonGPU/Transforms/Combine.cpp
View File

@@ -306,6 +306,8 @@ public:
auto fwdCvtIt = std::find_if(opIt, fwdEndIt, isCvt); auto fwdCvtIt = std::find_if(opIt, fwdEndIt, isCvt);
auto bwdCvtIt = std::find_if(bwdBeginIt, opIt, isCvt); auto bwdCvtIt = std::find_if(bwdBeginIt, opIt, isCvt);
if (!iterArg.value().getType().isa<RankedTensorType>())
continue;
if (fwdCvtIt != fwdEndIt) { if (fwdCvtIt != fwdEndIt) {
auto newFor = tryConvertIterArg(forOp, rewriter, iterArg.index(), auto newFor = tryConvertIterArg(forOp, rewriter, iterArg.index(),
(*fwdCvtIt)->getResult(0).getType()); (*fwdCvtIt)->getResult(0).getType());