[Mlir-commits] [mlir] [mlir][amdgpu] Shared memory access optimization pass (PR #75627)

[llvmlistbot at llvm.org]

================
@@ -0,0 +1,48 @@
+//===- Utils.cpp - Transform utilities ------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/AMDGPU/Transforms/Utils.h"
+
+#include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/Dialect/Vector/IR/VectorOps.h"
+
+using namespace mlir;
+using namespace mlir::amdgpu;
+
+Operation::operand_range amdgpu::getIndices(Operation *op) {
+  if (auto loadOp = dyn_cast<memref::LoadOp>(op))
+    return loadOp.getIndices();
+  if (auto storeOp = dyn_cast<memref::StoreOp>(op))
+    return storeOp.getIndices();
+  if (auto vectorReadOp = dyn_cast<vector::LoadOp>(op))
+    return vectorReadOp.getIndices();
+  if (auto vectorStoreOp = dyn_cast<vector::StoreOp>(op))
+    return vectorStoreOp.getIndices();
+  if (auto transferReadOp = dyn_cast<vector::TransferReadOp>(op))
+    return transferReadOp.getIndices();
+  if (auto transferWriteOp = dyn_cast<vector::TransferWriteOp>(op))
+    return transferWriteOp.getIndices();
+  llvm_unreachable("unsupported op type");
+}
+
+void amdgpu::setIndices(Operation *op, ArrayRef<Value> indices) {
+  if (auto loadOp = dyn_cast<memref::LoadOp>(op))
+    return loadOp.getIndicesMutable().assign(indices);
+  if (auto storeOp = dyn_cast<memref::StoreOp>(op))
+    return storeOp.getIndicesMutable().assign(indices);
+  if (auto vectorReadOp = dyn_cast<vector::LoadOp>(op))
+    return vectorReadOp.getIndicesMutable().assign(indices);
+  if (auto vectorStoreOp = dyn_cast<vector::StoreOp>(op))
+    return vectorStoreOp.getIndicesMutable().assign(indices);
+  if (auto transferReadOp = dyn_cast<vector::TransferReadOp>(op))
+    return transferReadOp.getIndicesMutable().assign(indices);
+  if (auto transferWriteOp = dyn_cast<vector::TransferWriteOp>(op))
+    return transferWriteOp.getIndicesMutable().assign(indices);
+  llvm_unreachable("unsupported op type");
+}
----------------
erman-gurses wrote:

Thanks for the comments above. 

I did my homework for that. Like finding this interface here below and tried to design for implementation. 
https://github.com/llvm/llvm-project/blob/main/mlir/lib/Dialect/MemRef/Transforms/RuntimeOpVerification.cpp#L351-L353

I did not figure out how to apply this example to our case (The interface already knows what concrete type is passed but in our case, we pass a generic _Operation_ type then figure out what concrete op type is in the function). Basically, it is not obvious how to use it. 

Do you have any pointers on that? If not, can we leave it like this?

https://github.com/llvm/llvm-project/pull/75627