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

[Analysis/Allocation] Allocation passes now assumes that slices always alias (#108)

Browse Source 
This code in this branch assumes the `src` operand in
`insert_slice_async` always aliases the result, which shouldn't hold for
generally cases but is just a workaround to make the pipeline pass work.

I'm also working on the complete analysis in another
[branch](https://github.com/openai/triton-mlir/tree/keren/analyze-slice).
This commit is contained in:
Keren Zhou
2022-09-09 12:03:41 -07:00
committed by GitHub
parent 9bd5a3dcd2
commit 16aed94ff5
14 changed files with 299 additions and 195 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
lib/Analysis/Alias.cpp
View File

@@ -22,25 +22,24 @@ ChangeResult SharedMemoryAliasAnalysis::visitOperation(
AliasInfo aliasInfo;
bool pessimistic = true;
if (maybeSharedAllocationOp(op)) {
// These ops will allocate a new shared memory buffer.
// These ops may allocate a new shared memory buffer.
auto result = op->getResult(0);
if (isSharedEncoding(result)) {
aliasInfo.insert(result);
// FIXME(Keren): extract and insert are always alias for now
if (auto extractSliceOp = dyn_cast<triton::gpu::ExtractSliceOp>(op)) {
// extract_slice %src, %index
aliasInfo = AliasInfo(operands[0]->getValue());
} else if (auto insertSliceOp =
dyn_cast<triton::gpu::InsertSliceAsyncOp>(op)) {
// insert_slice_async %src, %dst, %index
aliasInfo = AliasInfo(operands[1]->getValue());
} else {
aliasInfo.insert(result);
}
pessimistic = false;
} else {
llvm::errs() << "op: " << op->getName() << "\n";
}
}
// XXX(Keren): triton ops don't support aliasing yet.
// else if (auto viewOp = dyn_cast<triton::ViewOp>(op) ||
// dyn_cast<triton::ExpandDimsOp>(op)) {
// // These ops will reate a new view of the same shared memory buffer.
// auto result = op->getResult(0);
// if (isSharedEncoding(result)) {
// aliasInfo = AliasInfo(operands[0]->getValue());
// pessimistic = false;
// }
//}
if (pessimistic) {
return markAllPessimisticFixpoint(op->getResults());
}

12
lib/Analysis/Allocation.cpp
View File

@@ -39,11 +39,13 @@ private:
/// Initializes explicitly defined shared memory values for a given operation.
void getExplicitValueSize(Operation *op) {
/// Values returned from scf.yield will not be allocated even though they
/// have the shared encoding.
/// For example: %a = scf.if -> yield
/// %a must be allocated elsewhere by other operations.
if (!maybeSharedAllocationOp(op)) {
// Values returned from scf.yield will not be allocated even though they
// have the shared encoding.
// For example: %a = scf.if -> yield
// %a must be allocated elsewhere by other operations.
// FIXME(Keren): extract and insert are always alias for now
if (!maybeSharedAllocationOp(op) || isa<triton::gpu::ExtractSliceOp>(op) ||
isa<triton::gpu::InsertSliceAsyncOp>(op)) {
return;
}

16
lib/Analysis/Membar.cpp
View File

@@ -45,19 +45,25 @@ void MembarAnalysis::transfer(Operation *op, RegionInfo *regionInfo,
if (op->getNumResults() < 1)
return;
if (dyn_cast<scf::ForOp>(op) || dyn_cast<scf::IfOp>(op) ||
dyn_cast<scf::YieldOp>(op)) {
// Do not insert barriers before control flow operations.
if (isa<scf::ForOp>(op) || isa<scf::IfOp>(op) || isa<scf::YieldOp>(op) ||
isa<triton::gpu::ExtractSliceOp>(op) ||
isa<triton::gpu::InsertSliceAsyncOp>(op) ||
isa<triton::gpu::AllocTensorOp>(op)) {
// Do not insert barriers before control flow operations and
// alloc/extract/insert
// alloc is an allocation op without memory write.
// In contrast, arith.constant is an allocation op with memory write.
// FIXME(Keren): extract and insert are always alias for now
return;
}
if (dyn_cast<gpu::BarrierOp>(op)) {
if (isa<gpu::BarrierOp>(op)) {
// If the current op is a barrier, we sync previous reads and writes
regionInfo->sync();
return;
}
if (dyn_cast<triton::gpu::AsyncWaitOp>(op)) {
if (isa<triton::gpu::AsyncWaitOp>(op)) {
// If the current op is an async wait, we insert a barrier op and sync
// previous reads and writes.
OpBuilder::InsertionGuard g(*builder);