Skip to content

[NFC] Add debug logging to Materialize Block Pointer pass #2348

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
alexbaden merged 1 commit into from
Sep 26, 2024
Merged

[NFC] Add debug logging to Materialize Block Pointer pass #2348

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 16 additions & 0 deletions third_party/intel/lib/TritonIntelGPUTransforms/MaterializeBlockPointer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,11 @@
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/IR/Visitors.h"
#include "triton/Analysis/Utility.h"
#include "llvm/Support/Debug.h"

#define DEBUG_TYPE "tritonintelgpu-materialize-block-pointer"
#define DBGS() (llvm::dbgs() << "[" DEBUG_TYPE "]: ")
#define LDBG(X) LLVM_DEBUG(DBGS() << X << "\n")

using namespace mlir;
namespace tt = mlir::triton;
Expand Down Expand Up @@ -33,6 +38,8 @@ struct TritonIntelGPUMaterializeBlockPointerPass

MLIRContext *context = &getContext();
mod.walk([context](tt::LoadOp loadOp) {
LDBG("Considering op: " << loadOp);

Value ptr = loadOp.getPtr();
if (!tt::isTensorPointerType(ptr.getType()))
return;
Expand All @@ -41,26 +48,35 @@ struct TritonIntelGPUMaterializeBlockPointerPass
"Expected 'loadOp' to load a tensor value.");

tt::MakeTensorPtrOp makeTensorPtrOp = getMakeTensorPtrOp(ptr);
LDBG("Found make tensor ptr op: " << makeTensorPtrOp);
auto ptrType = cast<tt::PointerType>(makeTensorPtrOp.getType());
auto tensorType = cast<RankedTensorType>(ptrType.getPointeeType());
ArrayRef<int32_t> order = makeTensorPtrOp.getOrder();
unsigned rank = order.size();
LDBG("Rank: " << rank);
if (rank == 1)
return;

unsigned fastChangeDim = order[0];
LDBG("Fast change dim: " << fastChangeDim);
if (fastChangeDim >= (rank - 2)) {
Operation::operand_range strides = makeTensorPtrOp.getStrides();

// HW 2D block read instruction only supports contiguous access.
Value fastChangeStride = strides[fastChangeDim];
LLVM_DEBUG({
DBGS() << "fastChangeStride: ";
fastChangeStride.print(llvm::dbgs());
llvm::dbgs() << "\n";
});
if (!mlir::triton::gpu::intel::isConstant(fastChangeStride, 1))
return;

// Across Intel platforms, the strictest pitch restriction is to be a
// multiple of OWord(128 bits).
Value pitch =
strides[(fastChangeDim == rank - 1) ? rank - 2 : rank - 1];
LDBG("Pitch: " << pitch);
if (!ttgi::isDivisible(pitch, 64 / tensorType.getElementTypeBitWidth()))
return;

Expand Down