[mlir][linalg] Improve hoist padding buffer size computation.

[gysit]

Summary

Adapt the Fourier Motzkin elimination to take into account affine computations happening outside of the cloned loop nest.

Depends On D112713

[gysit]

Cleanup.

[nicolasvasilache]

Very nice!

mlir/lib/Dialect/Linalg/Transforms/HoistPadding.cpp

271

Similar refactoring, comment, checkm re side-effects should be done before landing please.

[gysit]

Address comments by improving comments and making the index slice computation more concise.

[gysit]

Rebase.

mlir/lib/Dialect/Linalg/Transforms/HoistPadding.cpp

 // TODO: extend on a per-need basis.
 static LogicalResult
 foldUpperBoundsIntoConstraintsSet(FlatAffineValueConstraints &constraints,
                                   scf::ForOp outerLimit,
                                   ArrayRef<scf::ForOp> loops) {
   SetVector<Value> toProjectOut;
   for (scf::ForOp loop : loops) {
     auto ub = loop.upperBound();
-    if (isDefinedOutsideOrConstant(outerLimit, ub))
-      continue;
 
     // Compute a backward slice up to, but not including, `outerLimit`.
-    SetVector<Operation *> backwardSlice;
-    getBackwardSlice(ub, &backwardSlice, [&](Operation *op) {
-      return outerLimit->isProperAncestor(op);
+    SetVector<Value> indexEdges;
+    indexEdges.insert(ub);
+    SetVector<Operation *> indexSlice;
+    getBackwardSlice(ub, &indexSlice, [&](Operation *op) {
+      // Continue only along the index operands of the for loop.
+      if (auto forOp = dyn_cast<scf::ForOp>(op)) {
+        if (!indexEdges.contains(forOp.getInductionVar()))
+          return false;
+        indexEdges.insert(forOp.lowerBound());
+        indexEdges.insert(forOp.upperBound());
+        indexEdges.insert(forOp.step());
+        return true;
+      }
+      // All supported operations have one result.
+      assert(op->getNumResults() == 1 &&
+             "expect operations to have one result");
+      if (!indexEdges.contains(op->getResult(0)))
+        return false;
+      for (Value value : op->getOperands())
+        indexEdges.insert(value);
+      return true;
     });
-    backwardSlice.insert(ub.getDefiningOp());
+    indexSlice.insert(ub.getDefiningOp());
 
     // Iterate over all ops in the slice and compose them in the constraints.
-    for (Operation *op : llvm::reverse(backwardSlice)) {
-      if (!isa<scf::ForOp, AffineApplyOp, AffineMinOp>(op))
-        return failure();
-      if (isa<scf::ForOp>(op))
+    for (Operation *op : llvm::reverse(indexSlice)) {
+      if (isa<scf::ForOp, arith::ConstantOp, tensor::DimOp>(op))
         continue;
-      // Ensure there is a
+      if (!isa<AffineApplyOp, AffineMinOp>(op))
+        return failure();
+      // Ensure there is an id.
       auto ensureIdFailed = [&](Value v) {
         if (constraints.containsId(v)) {
           unsigned pos;
           constraints.findId(v, &pos);
           return pos >= constraints.getNumDimIds();
         }
         constraints.appendDimId(v);
         return false;
       };
 
       // Ensure all ids exist and add results for later projection.
       if (llvm::any_of(op->getResults(), ensureIdFailed) ||
           llvm::any_of(op->getOperands(), ensureIdFailed))
         return failure();
 
       // All supported ops have 1 result.
       // TODO: extend when needed.
+      assert(op->getNumResults() == 1 &&
+             "expect operations to have one result");
       toProjectOut.insert(op->getResult(0));
 
       // Compose supported ops.
       if (auto affineApplyOp = dyn_cast<AffineApplyOp>(op)) {
         AffineValueMap avm(affineApplyOp.getAffineMap(),
                            affineApplyOp.getOperands(),
                            affineApplyOp.getResult());
         if (failed(constraints.composeMap(&avm)))

mlir/test/Dialect/Linalg/hoist-padding.mlir

func @dot(%arg0: tensor<?xf32>, %arg1: tensor<?xf32>, %arg2: tensor<f32>)
   %c0 = arith.constant 0 : index
   %1 = tensor.dim %arg0, %c0 : tensor<?xf32>
   %2 = tensor.dim %arg0, %c0 : tensor<?xf32>
   %3 = tensor.dim %arg1, %c0 : tensor<?xf32>
 
   //      CHECK: scf.for %[[I:[0-9a-z]+]] =
   //
   //      CHECK:   %[[MR8:.*]] = affine.min #[[$MIN_REST8]](%[[I]])
-  //      CHECK:   %[[D0:.*]] = affine.apply #[[$DIV4]](%[[MR8]])
   // Init tensor and pack.
-  //      CHECK:   %[[INIT_PACKED_A:.*]] = linalg.init_tensor [%[[D0]], 2, 2] : tensor<?x2x2xf32>
-  //      CHECK:   %[[CAST_INIT_PACKED_A:.*]] = tensor.cast %[[INIT_PACKED_A]] : tensor<?x2x2xf32> to tensor<?x?x2xf32>
+  //      CHECK:   %[[INIT_PACKED_A:.*]] = linalg.init_tensor [2, 2, 2] : tensor<2x2x2xf32>
+  //      CHECK:   %[[CAST_INIT_PACKED_A:.*]] = tensor.cast %[[INIT_PACKED_A]] : tensor<2x2x2xf32> to tensor<?x?x2xf32>
   //      CHECK:   %[[PACKED_A:.*]] = scf.for %[[II:[0-9a-z]+]] = {{.*}} iter_args(%{{.*}} = %[[CAST_INIT_PACKED_A]]) -> (tensor<?x?x2xf32>) {
   //      CHECK:     scf.for %[[III:[0-9a-z]+]] =
   //      CHECK:       tensor.insert_slice %{{.*}} into %{{.*}}[%{{.*}}, %{{.*}}, 0] [1, 1, 2] [1, 1, 1] : tensor<2xf32> into tensor<?x?x2xf32>
   //
-  //      CHECK:   %[[D0_2:.*]] = affine.apply #[[$DIV4]](%[[MR8]])
   // Init tensor and pack.
-  //      CHECK:   %[[INIT_PACKED_B:.*]] = linalg.init_tensor [%[[D0_2]], 2, 2] : tensor<?x2x2xf32>
-  //      CHECK:   %[[CAST_INIT_PACKED_B:.*]] = tensor.cast %[[INIT_PACKED_B]] : tensor<?x2x2xf32> to tensor<?x?x2xf32>
+  //      CHECK:   %[[INIT_PACKED_B:.*]] = linalg.init_tensor [2, 2, 2] : tensor<2x2x2xf32>
+  //      CHECK:   %[[CAST_INIT_PACKED_B:.*]] = tensor.cast %[[INIT_PACKED_B]] : tensor<2x2x2xf32> to tensor<?x?x2xf32>
   //      CHECK:   %[[PACKED_B:.*]] = scf.for %[[II_2:[0-9a-z]+]] = {{.*}} iter_args(%{{.*}} = %[[CAST_INIT_PACKED_B]]) -> (tensor<?x?x2xf32>) {
   //      CHECK:     scf.for %[[III_2:[0-9a-z]+]] =
   //      CHECK:       tensor.insert_slice %{{.*}} into %{{.*}}[%{{.*}}, %{{.*}}, 0] [1, 1, 2] [1, 1, 1] : tensor<2xf32> into tensor<?x?x2xf32>
   // Compute.
   //      CHECK:   scf.for %[[II_3:[0-9a-z]+]] =
   //      CHECK:     scf.for %[[III_3:[0-9a-z]+]] = {{.*}} iter_args(%[[C:.*]] = %{{.*}}) -> (tensor<f32>) {
   //      CHECK:       %[[IDX0:.*]] = affine.apply #[[$DIV4]](%[[II_3]])
   //      CHECK:       %[[IDX1:.*]] = affine.apply #[[$DIV2]](%[[III_3]])

mlir/test/Dialect/Linalg/pad-and-hoist.mlir

%0 = scf.for %arg3 = %c0 to %c24 step %c15 iter_args(%arg4 = %arg2) -> (tensor<24x25xf32>) {
 
     //    CHECK-DOUBLE:  scf.for %[[IV1:[0-9a-zA-Z]*]] =
     %1 = scf.for %arg5 = %c0 to %c25 step %c16 iter_args(%arg6 = %arg4) -> (tensor<24x25xf32>) {
       %2 = affine.min #map0(%arg3)
       %3 = affine.min #map1(%arg5)
       %4 = tensor.extract_slice %arg6[%arg3, %arg5] [%2, %3] [1, 1] : tensor<24x25xf32> to tensor<?x?xf32>
 
       // Packing the second input operand.
-      //    CHECK-DOUBLE:  = linalg.init_tensor [%{{.*}}, 12, 6]
+      //    CHECK-DOUBLE:  = linalg.init_tensor [3, 12, 6]
       //    CHECK-DOUBLE:  = linalg.pad_tensor {{.*}} nofold
 
       //    CHECK-DOUBLE:  scf.for %[[IV2:[0-9a-zA-Z]*]] =
       %5 = scf.for %arg7 = %c0 to %2 step %c5 iter_args(%arg8 = %4) -> (tensor<?x?xf32>) {
 
         //    CHECK-DOUBLE:  scf.for %[[IV3:[0-9a-zA-Z]*]] =
         %7 = scf.for %arg9 = %c0 to %3 step %c6 iter_args(%arg10 = %arg8) -> (tensor<?x?xf32>) {
           %8 = affine.min #map2(%arg7, %2)