[Mlir-commits] [mlir] [mlir][tosa] Enhance verify checks for PAD Op (PR #137177)

[Peng Sun]

================
@@ -1534,15 +1534,49 @@ LogicalResult tosa::PadOp::verify() {
   if (!inputType || !outputType)
     return success();
 
-  auto paddingRank = cast<tosa::shapeType>(getPadding().getType()).getRank();
+  auto inputRank = inputType.getRank();
+  auto outputRank = outputType.getRank();
+  if (inputRank != outputRank)
+    return emitOpError() << "expect same input and output tensor rank, but got "
+                         << "inputRank: " << inputRank
+                         << ", outputRank: " << outputRank;
+
+  DenseIntElementsAttr paddingAttr;
+  if (!matchPattern(getPadding(), m_Constant(&paddingAttr)))
+    return failure();
+
+  auto paddingValues = paddingAttr.getValues<APInt>();
+  if (paddingValues.size() != static_cast<size_t>(inputRank * 2))
+    return emitOpError() << "padding tensor must have " << inputRank
+                         << " * 2 = " << inputRank * 2 << " elements, but got "
+                         << paddingValues.size();
+
+  auto inputShape = inputType.getShape();
+  auto outputShape = outputType.getShape();
+
+  for (int64_t i = 0; i < inputRank; ++i) {
+    // Skip shape verification for dynamic dims
+    if (inputShape[i] == ShapedType::kDynamic ||
+        outputShape[i] == ShapedType::kDynamic)
+      continue;
+
+    int64_t padStart = paddingValues[i * 2].getSExtValue();
+    int64_t padEnd = paddingValues[i * 2 + 1].getSExtValue();
 
-  if (inputType.getRank() != outputType.getRank())
-    return emitOpError() << "expect same input and output tensor rank.";
+    if (padStart < 0 || padEnd < 0) {
+      return emitOpError() << "padding values must be non-negative, got ["
----------------
psunn wrote:

Thanks Luke.

Update to check both cases sequentially.

There are existing test cases for dynamic padding (e.g., `pad_dyn_padding` in _tosa-to-tensor.mlir_):
```
%0 = tosa.const_shape {values = dense<[-1, 2, 3, 4]> : tensor<4xindex>} : () -> !tosa.shape<4>
%pad_const = "tosa.const"() {values = dense<3.14> : tensor<1xf32>} : () -> tensor<1xf32>
%1 = "tosa.pad"(%arg0, %0, %pad_const)  : (tensor<1x2xf32>, !tosa.shape<4>, tensor<1xf32>)  -> (tensor<?x9xf32>)
```
While the TOSA specification does not permit negative padding values, the pad verifier currently allows `-1` as a placeholder for dynamic padding. We could consider strengthening this within validation pass in the TOSA dialect, if needed.

https://github.com/llvm/llvm-project/pull/137177