[mlir][tosa] Allow unranked input/output tensors in resize ops (llvm#141608)

This commit allows the input/output of the resize op to be unranked to
account for shapes being computed during shape inference.

Author: lhutton1

mlir/lib/Dialect/Tosa/IR/TosaOps.cpp

@@ -2496,16 +2496,6 @@ LogicalResult tosa::ResizeOp::verify() {
   const RankedTensorType outputType =
       llvm::dyn_cast<RankedTensorType>(output.getType());
 
-  if (!inputType)
-    return emitOpError("expect a ranked input tensor");
-  if (!outputType)
-    return emitOpError("expect a ranked output tensor");
-
-  const int64_t oh = outputType.getDimSize(1);
-  const int64_t ow = outputType.getDimSize(2);
-  const int64_t ih = inputType.getDimSize(1);
-  const int64_t iw = inputType.getDimSize(2);
-
   SmallVector<int64_t> scaleValues;
   SmallVector<int64_t> offsetValues;
   SmallVector<int64_t> borderValues;
@@ -2531,6 +2521,16 @@ LogicalResult tosa::ResizeOp::verify() {
   const int64_t borderY = borderValues[0];
   const int64_t borderX = borderValues[1];
 
+  if (!inputType)
+    return success();
+  if (!outputType)
+    return success();
+
+  const int64_t oh = outputType.getDimSize(1);
+  const int64_t ow = outputType.getDimSize(2);
+  const int64_t ih = inputType.getDimSize(1);
+  const int64_t iw = inputType.getDimSize(2);
+
   // Don't check with input height that could be broadcast (ih != 1)
   // since Linalg, a consumer of TOSA, expects broadcasting support
   // in resize to be available. Taking the cautious approach for now,

mlir/test/Dialect/Tosa/ops.mlir

@@ -743,6 +743,26 @@ func.func @test_resize(%arg0: tensor<1x32x32x8xf32>) -> tensor<1x64x64x8xf32> {
   return %1 : tensor<1x64x64x8xf32>
 }
 
+// -----
+// CHECK-LABEL: resize_unranked_output
+func.func @test_resize_unranked_output(%arg0: tensor<1x32x32x8xf32>) -> tensor<*xf32> {
+  %scale = tosa.const_shape { values = dense<[4, 2, 4, 2]> : tensor<4xindex> } : () -> !tosa.shape<4>
+  %offset = tosa.const_shape { values = dense<[-1, -1]> : tensor<2xindex> } : () -> !tosa.shape<2>
+  %border = tosa.const_shape { values = dense<[1, 1]> : tensor<2xindex> } : () -> !tosa.shape<2>
+  %1 = tosa.resize %arg0, %scale, %offset, %border { mode = "BILINEAR" } : (tensor<1x32x32x8xf32>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<*xf32>
+  return %1 : tensor<*xf32>
+}
+
+// -----
+// CHECK-LABEL: resize_unranked_input
+func.func @test_resize_unranked_input(%arg0: tensor<*xf32>) -> tensor<1x64x64x8xf32> {
+  %scale = tosa.const_shape { values = dense<[4, 2, 4, 2]> : tensor<4xindex> } : () -> !tosa.shape<4>
+  %offset = tosa.const_shape { values = dense<[-1, -1]> : tensor<2xindex> } : () -> !tosa.shape<2>
+  %border = tosa.const_shape { values = dense<[1, 1]> : tensor<2xindex> } : () -> !tosa.shape<2>
+  %1 = tosa.resize %arg0, %scale, %offset, %border { mode = "BILINEAR" } : (tensor<*xf32>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<1x64x64x8xf32>
+  return %1 : tensor<1x64x64x8xf32>
+}
+
 // -----
 // CHECK-LABEL: cast
 func.func @test_cast1(%arg0: tensor<13x21x3xi32>) -> tensor<13x21x3xf32> {