[Mlir-commits] [mlir] [MLIR] Fix VectorEmulateNarrowType constant op mask bug (PR #116064)

llvmlistbot at llvm.org llvmlistbot at llvm.org
Wed Nov 13 08:44:01 PST 2024


https://github.com/lialan updated https://github.com/llvm/llvm-project/pull/116064

>From 5be2464b76082602cc4d51af866136484a9de408 Mon Sep 17 00:00:00 2001
From: Alan Li <me at alanli.org>
Date: Tue, 12 Nov 2024 22:43:58 -0500
Subject: [PATCH] [MLIR] Fix VectorEmulateNarrowType constant op mask bug

This commit adds support for handling mask constants generated by the
`arith.constant` op in the `VectorEmulateNarrowType` pattern. Previously, this
pattern would not match due to the lack of mask constant handling in
`getCompressedMaskOp`.

The changes include:

1. Updating `getCompressedMaskOp` to recognize and handle `arith.constant` ops as
   mask value sources.

2. Handling cases where the mask is not aligned with the emulated load width.
   The compressed mask is adjusted to account for the offset.

Limitations:
- The arith.constant op can only have 1-dimensional constant values.

Resolves: #115742

Signed-off-by: Alan Li <me at alanli.org>
---
 .../Transforms/VectorEmulateNarrowType.cpp    | 46 ++++++++++++++++++-
 .../vector-emulate-narrow-type-unaligned.mlir | 38 +++++++++++++++
 .../Vector/vector-emulate-narrow-type.mlir    | 25 ++++++++++
 3 files changed, 107 insertions(+), 2 deletions(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
index eb4ce24548e603..f2e9ae18d3371c 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorEmulateNarrowType.cpp
@@ -70,7 +70,9 @@ static FailureOr<Operation *> getCompressedMaskOp(OpBuilder &rewriter,
   Operation *maskOp = mask.getDefiningOp();
   SmallVector<vector::ExtractOp, 2> extractOps;
   // Finding the mask creation operation.
-  while (maskOp && !isa<vector::CreateMaskOp, vector::ConstantMaskOp>(maskOp)) {
+  while (maskOp &&
+         !isa<arith::ConstantOp, vector::CreateMaskOp, vector::ConstantMaskOp>(
+             maskOp)) {
     if (auto extractOp = dyn_cast<vector::ExtractOp>(maskOp)) {
       maskOp = extractOp.getVector().getDefiningOp();
       extractOps.push_back(extractOp);
@@ -78,7 +80,8 @@ static FailureOr<Operation *> getCompressedMaskOp(OpBuilder &rewriter,
   }
   auto createMaskOp = dyn_cast_or_null<vector::CreateMaskOp>(maskOp);
   auto constantMaskOp = dyn_cast_or_null<vector::ConstantMaskOp>(maskOp);
-  if (!createMaskOp && !constantMaskOp)
+  auto constantOp = dyn_cast_or_null<arith::ConstantOp>(maskOp);
+  if (!createMaskOp && !constantMaskOp && !constantOp)
     return failure();
 
   // Computing the "compressed" mask. All the emulation logic (i.e. computing
@@ -129,6 +132,45 @@ static FailureOr<Operation *> getCompressedMaskOp(OpBuilder &rewriter,
       auto denseAttr = DenseElementsAttr::get(newMaskType, newMaskValues);
       newMask = rewriter.create<arith::ConstantOp>(loc, newMaskType, denseAttr);
     }
+  } else if (constantOp) {
+    assert(shape.size() == 1 && "expected 1-D mask");
+    // Rearrange the original mask values to cover the whole potential loading
+    // region. For example, in the case of using byte-size for emulation, given
+    // the following mask:
+    //
+    //   %mask = vector.constant_mask [0, 1, 0, 1, 0, 0] : vector<6xi2>
+    //
+    // with front offset of 1, the mask will be padded zeros in the front and
+    // back so that its length is multiple of `scale` (and the total coverage
+    // size is mulitiple of bytes):
+    //   %new_mask = vector.constant_mask [0, 0, 1, 0, 1, 0, 0, 0] :
+    //   vector<8xi2>
+    //
+    // The %new_mask is now aligned with the effective loading area and can now
+    // be compressed.
+    SmallVector<bool> maskValues(intraDataOffset, false);
+    if (auto denseAttr =
+            mlir::dyn_cast<DenseIntElementsAttr>(constantOp.getValue())) {
+      for (auto value : denseAttr.getValues<bool>()) {
+        maskValues.push_back(value);
+      }
+      while (maskValues.size() < numElements * scale) {
+        maskValues.push_back(false);
+      }
+    } else {
+      return failure();
+    }
+    // Compressing by combining every `scale` elements:
+    SmallVector<bool> compressedMaskValues;
+    for (size_t i = 0; i < maskValues.size(); i += scale) {
+      bool combinedValue = false;
+      for (int j = 0; j < scale; ++j) {
+        combinedValue |= maskValues[i + j];
+      }
+      compressedMaskValues.push_back(combinedValue);
+    }
+    newMask = rewriter.create<arith::ConstantOp>(
+        loc, DenseElementsAttr::get(newMaskType, compressedMaskValues));
   }
 
   while (!extractOps.empty()) {
diff --git a/mlir/test/Dialect/Vector/vector-emulate-narrow-type-unaligned.mlir b/mlir/test/Dialect/Vector/vector-emulate-narrow-type-unaligned.mlir
index 7ed75ff7f1579c..c61343a328d791 100644
--- a/mlir/test/Dialect/Vector/vector-emulate-narrow-type-unaligned.mlir
+++ b/mlir/test/Dialect/Vector/vector-emulate-narrow-type-unaligned.mlir
@@ -249,3 +249,41 @@ func.func @vector_maskedload_i2_dynamic_indexing_mixed(%passthru: vector<3xi2>,
 // CHECK: %[[IN8:.+]] = vector.insert %[[EX8]], %[[IN7]] [1] : i2 into vector<3xi2>
 // CHECK: %[[EX9:.+]] = vector.extract %[[SELECT]][%[[INCIDX2]]] : i2 from vector<8xi2>
 // CHECK: %[[IN9:.+]] = vector.insert %[[EX9]], %[[IN8]] [2] : i2 into vector<3xi2>
+
+// -----
+
+func.func @vector_maskedload_i4_constant_mask_unaligned(%passthru: vector<5xi2>) -> vector<5xi2> {
+  %0 = memref.alloc() : memref<3x5xi2>
+  %mask = arith.constant dense<[false, true, true, true, false]> : vector<5xi1>
+  %c0 = arith.constant 0 : index
+  %c1 = arith.constant 1 : index
+  %1 = vector.maskedload %0[%c1, %c0], %mask, %passthru :
+    memref<3x5xi2>, vector<5xi1>, vector<5xi2> into vector<5xi2>
+  return %1 : vector<5xi2>
+}
+
+// CHECK: func @vector_maskedload_i4_constant_mask_unaligned(
+// CHECK-SAME: %[[PTH:.+]]: vector<5xi2>) -> vector<5xi2>
+// CHECK: %[[ALLOC:.+]] = memref.alloc() : memref<4xi8>
+// CHECK: %[[MASK:.+]] = arith.constant dense<[false, true, true, true, false]> : vector<5xi1>
+
+// CHECK: %[[CST0:.+]] = arith.constant dense<true> : vector<2xi1>
+// CHECK: %[[CST1:.+]] = arith.constant dense<0> : vector<8xi2>
+// CHECK: %[[INSERT:.+]] = vector.insert_strided_slice %[[PTH]], %[[CST1]]
+// CHECK-SAME: {offsets = [1], strides = [1]} : vector<5xi2> into vector<8xi2>
+
+// Emulated masked load from alloc:
+// CHECK: %[[BCAST:.+]] = vector.bitcast %[[INSERT]] : vector<8xi2> to vector<2xi8>
+// CHECK: %[[C1:.+]] = arith.constant 1 : index
+// CHECK: %[[MASKLOAD:.+]] = vector.maskedload %[[ALLOC]][%[[C1]]], %[[CST0]], %[[BCAST]]
+// CHECK: %[[BCAST2:.+]] = vector.bitcast %[[MASKLOAD]] : vector<2xi8> to vector<8xi2>
+
+// Select from emulated loaded vector and passthru vector:
+// TODO: fold this part if possible.
+// CHECK: %[[CST2:.+]] = arith.constant dense<false> : vector<8xi1>
+// CHECK: %[[INSERT2:.+]] = vector.insert_strided_slice %[[MASK]], %[[CST2]]
+// CHECK-SAME: {offsets = [1], strides = [1]} : vector<5xi1> into vector<8xi1>
+// CHECK: %[[SELECT:.+]] = arith.select %[[INSERT2]], %[[BCAST2]], %[[INSERT]] : vector<8xi1>, vector<8xi2>
+// CHECK: %[[EXTRACT:.+]] = vector.extract_strided_slice %[[SELECT]]
+// CHECK-SAME: {offsets = [1], sizes = [5], strides = [1]} : vector<8xi2> to vector<5xi2>
+// CHECK: return %[[EXTRACT]] : vector<5xi2>
diff --git a/mlir/test/Dialect/Vector/vector-emulate-narrow-type.mlir b/mlir/test/Dialect/Vector/vector-emulate-narrow-type.mlir
index 034bd47f6163e6..1181605352a4a9 100644
--- a/mlir/test/Dialect/Vector/vector-emulate-narrow-type.mlir
+++ b/mlir/test/Dialect/Vector/vector-emulate-narrow-type.mlir
@@ -624,3 +624,28 @@ func.func @vector_maskedstore_i4_constant_mask(
 // CHECK32:           %[[SELECT:.+]] = arith.select %[[ORIG_MASK]], %[[VAL_TO_STORE]], %[[BITCAST]] : vector<8xi1>, vector<8xi4>
 // CHECK32:           %[[NEW_VAL:.+]] = vector.bitcast %[[SELECT]] : vector<8xi4> to vector<1xi32>
 // CHECK32:           vector.maskedstore %[[ALLOC]]{{\[}}%[[LIDX]]], %[[NEW_MASK]], %[[NEW_VAL]] : memref<3xi32>, vector<1xi1>, vector<1xi32>
+
+// -----
+
+func.func @vector_maskedload_i4_arith_constant(%passthru: vector<8xi4>) -> vector<8xi4> {
+  %0 = memref.alloc() : memref<3x8xi4>
+  %cst = arith.constant dense<0> : vector<8xi4>
+  %mask = arith.constant dense<[false, true, true, true, true, false, false, false]> : vector<8xi1>
+  %c0 = arith.constant 0 : index
+  %1 = vector.maskedload %0[%c0, %c0], %mask, %passthru :
+    memref<3x8xi4>, vector<8xi1>, vector<8xi4> into vector<8xi4>
+  return %1 : vector<8xi4>
+}
+
+// CHECK: func @vector_maskedload_i4_arith_constant(
+// CHECK-SAME:   %[[PASSTHRU:[a-zA-Z0-9]+]]: vector<8xi4>) -> vector<8xi4> {
+// CHECK: %[[ALLOC:.+]] = memref.alloc() : memref<12xi8>
+// CHECK: %[[MASK:.+]] = arith.constant dense<[false, true, true, true, true, false, false, false]> : vector<8xi1>
+// CHECK: %[[MASK2:.+]] = arith.constant dense<[true, true, true, false]> : vector<4xi1>
+// CHECK: %[[BITCAST:.+]] = vector.bitcast %[[PASSTHRU]] : vector<8xi4> to vector<4xi8>
+// CHECK: %[[C0:.+]] = arith.constant 0 : index
+// CHECK: %[[LOAD:.+]] = vector.maskedload %[[ALLOC]][%[[C0]]], %[[MASK2]], %[[BITCAST]]
+// CHECK-SAME : memref<12xi8>, vector<4xi1>, vector<4xi8> into vector<4xi8>
+// CHECK: %[[BITCAST2:.+]] = vector.bitcast %[[LOAD]] : vector<4xi8> to vector<8xi4>
+// CHECK: %[[SELECT:.+]] = arith.select %[[MASK]], %[[BITCAST2]], %[[PASSTHRU]] : vector<8xi1>, vector<8xi4>
+// CHECK: return %[[SELECT]] : vector<8xi4>



More information about the Mlir-commits mailing list