nnicolosi/Raptor
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

reduce spatial compile-times in convolutions using a scf.for instead of materializing a huge number of instructions
Some checks failed
Validate Operations / validate-operations (push) Has been cancelled
Details

Browse Source
This commit is contained in:
NiccoloN
2026-04-10 18:50:25 +02:00
parent f3a36e9d43
commit f054e66ed0
18 changed files with 623 additions and 241 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
src/PIM/Pass/Pim/ConstantFolding/Patterns/Constant.cpp
View File

@@ -120,7 +120,15 @@ struct FoldConstantCoreMapPattern final : OpRewritePattern<linalg::MapOp> {
auto getGlobalOp = memref::GetGlobalOp::create(rewriter, mapOp.getLoc(), initType, globalOp.getName());
rewriter.setInsertionPoint(mapOp);
rewriter.replaceAllUsesExcept(mapOp.getInit(), getGlobalOp.getResult(), mapOp);
auto sizeInBytes = initType.getNumElements() * initType.getElementTypeBitWidth() / 8;
pim::PimMemCopyOp::create(rewriter,
mapOp.getLoc(),
initType,
mapOp.getInit(),
getGlobalOp.getResult(),
rewriter.getI32IntegerAttr(0),
rewriter.getI32IntegerAttr(0),
rewriter.getI32IntegerAttr(sizeInBytes));
rewriter.eraseOp(mapOp);
return success();
}
@@ -416,6 +424,9 @@ struct FoldConstantMemCpPattern final : OpRewritePattern<pim::PimMemCopyOp> {
return failure();
if (llvm::any_of(srcSubview->strides, [](int64_t stride) { return stride != 1; }))
return failure();
auto staticOffsets = getStaticSubviewOffsets(*srcSubview);
if (failed(staticOffsets))
return failure();
auto resultTensorType = RankedTensorType::get(allocType.getShape(), allocType.getElementType());
const int64_t numResultElements = resultTensorType.getNumElements();
@@ -428,7 +439,7 @@ struct FoldConstantMemCpPattern final : OpRewritePattern<pim::PimMemCopyOp> {
auto resultIndices = delinearizeIndex(i, resultTensorType.getShape(), resultStrides);
SmallVector<int64_t> sourceIndices;
sourceIndices.reserve(resultIndices.size());
for (auto [off, idx] : llvm::zip_equal(srcSubview->offsets, resultIndices))
for (auto [off, idx] : llvm::zip_equal(*staticOffsets, resultIndices))
sourceIndices.push_back(off + idx);
int64_t srcLinear = linearizeIndex(sourceIndices, sourceStrides);
resultValues[i] = sourceValues[srcLinear];