Add mask to attention op tiling and decomposition

Signed-off-by: stanley-nod <[email protected]>

Author: raikonenfnu

compiler/src/iree/compiler/Dialect/LinalgExt/IR/LinalgExtOps.cpp

@@ -1329,9 +1329,9 @@ LogicalResult AttentionOp::verify() {
   int numInputs = getNumDpsInputs();
   int numOutputs = getNumDpsInits();
 
-  if (numInputs != 4) {
+  if (numInputs < 4 || numInputs > 5) {
     return op->emitOpError(
-        "expected 4 input operands: Query, Key, Value and Scale");
+        "expected 4 or 5 input operands: Query, Key, Value, Scale, and Mask");
   }
 
   if (numOutputs != 1 && numOutputs != 3) {
@@ -1340,8 +1340,9 @@ LogicalResult AttentionOp::verify() {
   }
 
   bool isTiled = numOutputs == 3;
-
-  if (!llvm::all_of(llvm::drop_end(getDpsInputs()), [](Value input) {
+  SmallVector<Value> dpsInputs = getDpsInputs();
+  ArrayRef<Value> qkvValues(dpsInputs.begin(), dpsInputs.begin() + 3);
+  if (!llvm::all_of(qkvValues, [](Value input) {
         return isa<ShapedType>(input.getType());
       })) {
     return op->emitOpError(
@@ -1462,10 +1463,17 @@ SmallVector<AffineMap> AttentionOp::getIndexingMapsArray() {
         AffineMap::get(/*dimCount=*/5, /*symbolCount=*/0, {batch, k2, n}, ctx);
   }
 
+  SmallVector<AffineMap> results = {qMap, kMap, vMap};
+
+  if (getMask()) {
+    AffineMap maskMap =
+        AffineMap::get(/*dimCount=*/5, /*symbolCount=*/0, {batch, m, k2}, ctx);
+    results.push_back(maskMap);
+  }
+
   AffineMap resMap =
       AffineMap::get(/*dimCount=*/5, /*symbolCount=*/0, {batch, m, n}, ctx);
-
-  SmallVector<AffineMap> results = {qMap, kMap, vMap, resMap};
+  results.push_back(resMap);
 
   if (getMax()) {
     AffineMap maxMap =

compiler/src/iree/compiler/Dialect/LinalgExt/IR/LinalgExtOps.td

@@ -522,11 +522,15 @@ def IREELinalgExt_AttentionOp : IREELinalgExt_Op<"attention",
   let description = [{
     Computes the scaled dot product attention function:
 
-    attention(Q, K, V, scale) = softmax(Q @ K.T * scale) @ V
+    attention(Q, K, V, scale, mask) = softmax(mask(Q @ K.T * scale)) @ V
 
     Here Q, K, V are given tensors and scale is a scalar value specifying
     the scale to use.
 
+    `mask` is an optional boolean tensor that specifies which relations
+    in attn_weight that should be ignored. This is useful for
+    causal attention, padded attention, and some special SD use cases.
+
     For self-attention, all inputs and the result have the same shape BxNxd
     where B is the batch dimension, N is the sequence length and d is head
     dimension. Typically N >>> d. Usually, this operator also performs
@@ -539,10 +543,10 @@ def IREELinalgExt_AttentionOp : IREELinalgExt_Op<"attention",
     FlashAttention 2 and optionally results in the current `max` and `sum`
     statistics while processing the current tile.
 
-    If transpose_v is speciifed, the V tensor passed as input is assumed to
+    If transpose_v is specifed, the V tensor passed as input is assumed to
     be transposed:
 
-    attention(Q, K, V, scale) = softmax(Q @ K.T * scale) @ V.T
+    attention(Q, K, V, scale, mask) = softmax(mask(Q @ K.T * scale)) @ V.T
 
     TODO: We should be moving to using a indexing map like approach so we
     can generalize which tensor is transposed and which is not.
@@ -579,6 +583,11 @@ def IREELinalgExt_AttentionOp : IREELinalgExt_Op<"attention",
     Value getScale() {
       return getDpsInputOperand(3)->get();
     }
+    std::optional<Value> getMask() {
+      if (getNumDpsInputs() < 5)
+        return std::nullopt;
+      return getDpsInputOperand(4)->get();
+    }
     Value getOutput() {
       return getDpsInitOperand(0)->get();
     }
@@ -659,18 +668,26 @@ def IREELinalgExt_AttentionOp : IREELinalgExt_Op<"attention",
     AffineMap getValueMap() {
       return cast<AffineMap>(getIndexingMapsArray()[2]);
     }
-    AffineMap getOutputMap() {
+    std::optional<AffineMap> getMaskMap() {
+      if (getNumDpsInputs() < 5)
+        return std::nullopt;
       return cast<AffineMap>(getIndexingMapsArray()[3]);
     }
+    AffineMap getOutputMap() {
+      int64_t outputIndex = getNumDpsInputs() - 1;
+      return cast<AffineMap>(getIndexingMapsArray()[outputIndex]);
+    }
     std::optional<AffineMap> getMaxMap() {
       if (getNumResults() < 2)
         return std::nullopt;
-      return cast<AffineMap>(getIndexingMapsArray()[4]);
+      int64_t maxIndex = getNumDpsInputs();
+      return cast<AffineMap>(getIndexingMapsArray()[maxIndex]);
     }
     std::optional<AffineMap> getSumMap() {
       if (getNumResults() < 3)
         return std::nullopt;
-      return cast<AffineMap>(getIndexingMapsArray()[5]);
+      int64_t sumIndex = getNumDpsInputs() + 1;
+      return cast<AffineMap>(getIndexingMapsArray()[sumIndex]);
     }
 
     int64_t getIterationDomainRank() {

compiler/src/iree/compiler/Dialect/LinalgExt/Transforms/DecomposeAttention.cpp

@@ -177,9 +177,49 @@ static Value truncateToF16(Value input, Value output,
   return genericOp.getResult(0);
 }
 
+static Value applyMasking(Value qkSlice, Value mask, OpBuilder &builder) {
+  ShapedType qkType = cast<ShapedType>(qkSlice.getType());
+  Location loc = qkSlice.getLoc();
+
+  // Create a fill op for scale.
+  SmallVector<OpFoldResult> qkDims =
+      tensor::getMixedSizes(builder, loc, qkSlice);
+
+  // Attention_mask is 1.0 for positions we want to attend and 0.0 for
+  // masked positions. this operation will create a tensor which is 0.0 for
+  // positions we want to attend and -10000.0 for masked positions
+  Value c0 = builder.create<arith::ConstantOp>(
+      loc, builder.getZeroAttr(qkType.getElementType()));
+
+  Value cLargeNeg = builder.create<arith::ConstantOp>(
+      loc, builder.getFloatAttr(qkType.getElementType(), -1e6));
+
+  Value empty =
+      builder.create<tensor::EmptyOp>(loc, qkDims, qkType.getElementType());
+  // Create a generic op to multiply the query by the scale.
+  SmallVector<utils::IteratorType> iteratorTypes(2,
+                                                 utils::IteratorType::parallel);
+  auto identityMap = AffineMap::getMultiDimIdentityMap(2, builder.getContext());
+  SmallVector<AffineMap> indexingMaps(3, identityMap);
+  auto applyMaskOp = builder.create<linalg::GenericOp>(
+      loc, TypeRange{empty.getType()}, ValueRange{qkSlice, mask},
+      ValueRange{empty}, indexingMaps, iteratorTypes,
+      [&](OpBuilder &b, Location loc, ValueRange args) {
+        // TODO: Add support to enable i1 throughout.
+        //      currently type propagation force it to i1.
+        Value trunci =
+            b.create<arith::TruncIOp>(loc, builder.getI1Type(), args[1]);
+        Value masking = b.create<arith::SelectOp>(loc, trunci, c0, cLargeNeg);
+        Value result = b.create<arith::AddFOp>(loc, args[0], masking);
+        b.create<linalg::YieldOp>(loc, result);
+      });
+  return applyMaskOp.getResult(0);
+}
+
 static std::tuple<Value, Value, Value>
 createAttentionBody(Value keySlice, Value valueSlice, Value querySlice,
-                    Value outputSlice, Value maxSlice, Value sumSlice,
+                    std::optional<Value> maskSlice, Value outputSlice,
+                    Value maxSlice, Value sumSlice,
                     OpFoldResult sequenceTileLength,
                     OpFoldResult keyValueTileLength, OpFoldResult headDimension,
                     Type elementType, SmallVectorImpl<Operation *> &ops,
@@ -195,6 +235,11 @@ createAttentionBody(Value keySlice, Value valueSlice, Value querySlice,
   Value qkTranspose = computeQKTranspose(querySlice, keySlice, emptySquare,
                                          zero, loc, builder, ops);
 
+  // Apply masking if mask is specified.
+  if (maskSlice.has_value()) {
+    qkTranspose = applyMasking(qkTranspose, maskSlice.value(), builder);
+  }
+
   // Compute current statistics
   Value newMax = computeRowwiseReduction<arith::MaximumFOp>(
       qkTranspose, maxSlice, loc, builder, ops);
@@ -327,9 +372,9 @@ void decomposeTiledAttention(IREE::LinalgExt::AttentionOp tiledAttnOp,
   // iteration of the loop.
   querySlice = scaleQuery(querySlice, scale, rewriter);
   ops.push_back(querySlice.getDefiningOp());
-
+  std::optional<Value> maybeMask = tiledAttnOp.getMask();
   auto [result, newMax, newSum] = createAttentionBody(
-      keySlice, valueSlice, querySlice, tiledResult, max, sum,
+      keySlice, valueSlice, querySlice, maybeMask, tiledResult, max, sum,
       sequenceTileLength, keyValueTileLength, headDimension, elementType, ops,
       tiledAttnOp.getTransposeV(), loc, rewriter);
 

compiler/src/iree/compiler/Dialect/LinalgExt/Transforms/TileAttention.cpp

@@ -79,7 +79,8 @@ static scf::LoopNest createLoopNest(SmallVectorImpl<Value> &ivs, Value lb,
 }
 
 static Value extractSlice(Value key, ArrayRef<int64_t> keyShape,
-                          ArrayRef<Value> ivs, OpFoldResult keyValueTileLength,
+                          ArrayRef<std::optional<Value>> ivs,
+                          OpFoldResult keyValueTileLength,
                           OpFoldResult headDimension, Type elementType,
                           Location loc, OpBuilder &builder,
                           bool swapLastTwoDims = false) {
@@ -90,8 +91,10 @@ static Value extractSlice(Value key, ArrayRef<int64_t> keyShape,
   SmallVector<OpFoldResult> offsets(keyShape.size(), zero);
   sizes[1] = keyValueTileLength;
   sizes[2] = headDimension;
-  if (!ivs.empty()) {
-    offsets[1] = ivs[0];
+  for (auto [idx, iv] : llvm::enumerate(ivs)) {
+    if (!iv.has_value())
+      continue;
+    offsets[idx] = iv.value();
   }
   SmallVector<int64_t> tensorShape{keyShape[1], keyShape[2]};
   if (swapLastTwoDims) {
@@ -181,6 +184,7 @@ IREE::LinalgExt::AttentionOp tileAttention(IREE::LinalgExt::AttentionOp attnOp,
   Location loc = attnOp.getLoc();
   OpBuilder::InsertionGuard guard(rewriter);
   rewriter.setInsertionPoint(attnOp);
+  llvm::outs() << "high level implement tile Attention!\n";
 
   Value query = attnOp.getQuery();
   ShapedType queryType = attnOp.getQueryType();
@@ -253,21 +257,35 @@ IREE::LinalgExt::AttentionOp tileAttention(IREE::LinalgExt::AttentionOp attnOp,
   rewriter.setInsertionPointToStart(loopNest.loops.back().getBody());
 
   // Extract slices
-  Value keySlice = extractSlice(key, keyShape, ivs, keyValueTileLength,
+  SmallVector<std::optional<Value>> kvIvs(keyShape.size(), std::nullopt);
+  kvIvs[1] = ivs[0];
+  Value keySlice = extractSlice(key, keyShape, kvIvs, keyValueTileLength,
                                 headDimension, elementType, loc, rewriter);
   Value valueSlice =
-      extractSlice(value, keyShape, ivs, keyValueTileLength, headDimension,
+      extractSlice(value, keyShape, kvIvs, keyValueTileLength, headDimension,
                    elementType, loc, rewriter, attnOp.getTransposeV());
   Value querySlice = extractSlice(query, queryShape, {}, sequenceTileLength,
                                   headDimension, elementType, loc, rewriter);
 
   Value scale = attnOp.getScale();
+  SmallVector<Value> tiledInputs = {querySlice, keySlice, valueSlice, scale};
+
+  if (attnOp.getMask().has_value()) {
+    Value mask = attnOp.getMask().value();
+    auto maskElType = llvm::cast<ShapedType>(mask.getType()).getElementType();
+    SmallVector<std::optional<Value>> maskIvs(keyShape.size(), std::nullopt);
+    maskIvs[2] = ivs[0];
+    SmallVector<int64_t> maskShape{queryShape[0], queryShape[1], keyShape[1]};
+    Value maskSlice =
+        extractSlice(mask, maskShape, maskIvs, sequenceTileLength,
+                     keyValueTileLength, maskElType, loc, rewriter);
+    tiledInputs.push_back(maskSlice);
+  }
 
   auto tiledAttentionOp = rewriter.create<IREE::LinalgExt::AttentionOp>(
       attnOp.getLoc(),
       SmallVector<Type>{accumulatorF32.getType(), sum.getType(), max.getType()},
-      SmallVector<Value>{querySlice, keySlice, valueSlice, scale},
-      SmallVector<Value>{iterArgResult, iterArgMax, iterArgSum});
+      tiledInputs, SmallVector<Value>{iterArgResult, iterArgMax, iterArgSum});
 
   if (attnOp.getTransposeV())
     tiledAttentionOp.setTransposeVAttr(attnOp.getTransposeVAttr());

compiler/src/iree/compiler/Dialect/LinalgExt/Transforms/TilingInterfaceImpl.cpp

@@ -1716,6 +1716,14 @@ AttentionOp::getTiledImplementation(OpBuilder &builder,
   tiledOperands.emplace_back(getSlice(builder, loc, getKey(), keySlice));
   tiledOperands.emplace_back(getSlice(builder, loc, getValue(), valueSlice));
   tiledOperands.emplace_back(scale);
+
+  std::optional<Value> mask = getMask();
+  if (mask) {
+    SmallVector<Range> maskSlice =
+        getPermutedSlice(*getMaskMap(), offsets, sizes);
+    tiledOperands.emplace_back(getSlice(builder, loc, mask.value(), maskSlice));
+  }
+
   tiledOperands.emplace_back(getSlice(builder, loc, getOutput(), outputSlice));
 
   std::optional<Value> max = getMax();
@@ -1733,13 +1741,14 @@ AttentionOp::getTiledImplementation(OpBuilder &builder,
   }
 
   SmallVector<Type> resultTypes;
+  int64_t resultIndex = getNumDpsInputs();
   if (hasPureTensorSemantics()) {
-    resultTypes.push_back(tiledOperands[4].getType());
+    resultTypes.push_back(tiledOperands[resultIndex].getType());
     if (max) {
-      resultTypes.push_back(tiledOperands[5].getType());
+      resultTypes.push_back(tiledOperands[resultIndex + 1].getType());
     }
     if (sum) {
-      resultTypes.push_back(tiledOperands[6].getType());
+      resultTypes.push_back(tiledOperands[resultIndex + 2].getType());
     }
   }
 

compiler/src/iree/compiler/Dialect/LinalgExt/Utils/IndexingUtils.cpp

@@ -31,7 +31,6 @@ void AttentionOpDetail::inferFromIndexingMaps(
   AffineMap qMap = indexingMaps[0];
   AffineMap kMap = indexingMaps[1];
   AffineMap vMap = indexingMaps[2];
-  AffineMap resMap = indexingMaps[3];
 
   // Q   = B x M x K1
   // K   = B x K2 x K1
@@ -40,7 +39,6 @@ void AttentionOpDetail::inferFromIndexingMaps(
   llvm::SmallDenseSet<int64_t> qSet = findPermutationsIndexingOperand(qMap);
   llvm::SmallDenseSet<int64_t> vSet = findPermutationsIndexingOperand(vMap);
   llvm::SmallDenseSet<int64_t> kSet = findPermutationsIndexingOperand(kMap);
-  llvm::SmallDenseSet<int64_t> resSet = findPermutationsIndexingOperand(resMap);
 
   // B = Q & K & V
   llvm::SmallDenseSet<int64_t> bSet = qSet;
@@ -76,7 +74,7 @@ void AttentionOpDetail::inferFromIndexingMaps(
 
 FailureOr<AttentionOpDetail>
 AttentionOpDetail::get(ArrayRef<AffineMap> indexingMaps) {
-  if (indexingMaps.size() != 4 && indexingMaps.size() != 6) {
+  if (indexingMaps.size() < 4 || indexingMaps.size() > 7) {
     return failure();
   }