[mlir][tensor] Do not use affine ops on non-perfect-tiling unpack cases.

Do not create an Affine ops for expanded size because the affine op is
too complicated which would hit an assertion in affine ops
simplification.

Reviewed By: mravishankar

Differential Revision: https://reviews.llvm.org/D144151

diff --git a/mlir/lib/Dialect/Tensor/IR/TensorTilingInterfaceImpl.cpp b/mlir/lib/Dialect/Tensor/IR/TensorTilingInterfaceImpl.cpp
index e4367e0..ff11a3a 100644 (file)
--- a/mlir/lib/Dialect/Tensor/IR/TensorTilingInterfaceImpl.cpp
+++ b/mlir/lib/Dialect/Tensor/IR/TensorTilingInterfaceImpl.cpp
@@ -318,8 +318,11 @@ static UnpackTileDimInfo getUnpackTileDimInfo(OpBuilder &b, UnPackOp unpackOp,
       ab.add(AV(dim0).bind(lengthMinusOne), AV(dim1).bind(oneAttr));
   info.sourceOffset = firstCoord.quotient;
   info.resultOffset = firstCoord.remainder;
-  info.destExpandedSize =
-      ab.mul(AV(dim0).bind(info.sourceSize), AV(sym0).bind(innerTileSize));
+  // Do not create an Affine ops for expanded size because the affine op is too
+  // complicated which would trigger an issue in affine ops simplification.
+  info.destExpandedSize = b.createOrFold<arith::MulIOp>(
+      loc, getValueOrCreateConstantIndexOp(b, loc, info.sourceSize),
+      getValueOrCreateConstantIndexOp(b, loc, innerTileSize));
   return info;
 }
 

diff --git a/mlir/test/Dialect/Tensor/tiling.mlir b/mlir/test/Dialect/Tensor/tiling.mlir
index f7734fd..b92e5ae 100644 (file)
--- a/mlir/test/Dialect/Tensor/tiling.mlir
+++ b/mlir/test/Dialect/Tensor/tiling.mlir
@@ -391,11 +391,9 @@ transform.sequence failures(propagate) {
 // CHECK-DAG:   #[[MAP0:.+]] = affine_map<(d0) -> (d0 floordiv 32)>
 // CHECK-DAG:   #[[MAP1:.+]] = affine_map<(d0) -> (d0 mod 32)>
 // CHECK-DAG:   #[[MAP2:.+]] = affine_map<(d0) -> ((d0 + 1) floordiv 32 - d0 floordiv 32 + 1)>
-// CHECK-DAG:   #[[MAP3:.+]] = affine_map<(d0) -> (((d0 + 1) floordiv 32) * 32 - (d0 floordiv 32) * 32 + 32)>
 // CHECK-DAG:   #[[MAP4:.+]] = affine_map<(d0) -> (d0 floordiv 16)>
 // CHECK-DAG:   #[[MAP5:.+]] = affine_map<(d0) -> (d0 mod 16)>
 // CHECK-DAG:   #[[MAP6:.+]] = affine_map<(d0) -> ((d0 + 3) floordiv 16 - d0 floordiv 16 + 1)>
-// CHECK-DAG:   #[[MAP7:.+]] = affine_map<(d0) -> (((d0 + 3) floordiv 16) * 16 - (d0 floordiv 16) * 16 + 16)>
 // CHECK:       func.func @NCnc_to_NC
 // CHECK-SAME:    %[[IN:[A-Za-z0-9]+]]:
 // CHECK-SAME:    %[[OUT:[A-Za-z0-9]+]]:
@@ -440,11 +438,9 @@ transform.sequence failures(propagate) {
 // CHECK-DAG:   #[[MAP0:.+]] = affine_map<(d0) -> (d0 floordiv 32)>
 // CHECK-DAG:   #[[MAP1:.+]] = affine_map<(d0) -> (d0 mod 32)>
 // CHECK-DAG:   #[[MAP2:.+]] = affine_map<(d0) -> ((d0 + 1) floordiv 32 - d0 floordiv 32 + 1)>
-// CHECK-DAG:   #[[MAP3:.+]] = affine_map<(d0) -> (((d0 + 1) floordiv 32) * 32 - (d0 floordiv 32) * 32 + 32)>
 // CHECK-DAG:   #[[MAP4:.+]] = affine_map<(d0) -> (d0 floordiv 8)>
 // CHECK-DAG:   #[[MAP5:.+]] = affine_map<(d0) -> (d0 mod 8)>
 // CHECK-DAG:   #[[MAP6:.+]] = affine_map<(d0) -> ((d0 + 3) floordiv 8 - d0 floordiv 8 + 1)>
-// CHECK-DAG:   #[[MAP7:.+]] = affine_map<(d0) -> (((d0 + 3) floordiv 8) * 8 - (d0 floordiv 8) * 8 + 8)>
 // CHECK:       func.func @CKkc_to_KC
 // CHECK-SAME:    %[[IN:[A-Za-z0-9]+]]:
 // CHECK-SAME:    %[[OUT:[A-Za-z0-9]+]]: