[Mlir-commits] [mlir] [mlir][tosa] Add more verifiers for the following operators (PR #127923)

[llvmlistbot at llvm.org]

================
@@ -899,6 +964,57 @@ LogicalResult tosa::MatMulOp::inferReturnTypeComponents(
   return success();
 }
 
+LogicalResult MatMulOp::verify() {
+  auto aType = llvm::dyn_cast<ShapedType>(getA().getType());
+  auto bType = llvm::dyn_cast<ShapedType>(getB().getType());
+
+  // Must be shaped tensor types
+  if (!aType) {
+    emitOpError("expect a shaped tensor for input a, got ") << getA().getType();
+    return failure();
+  }
+  if (!bType) {
+    emitOpError("expect a shaped tensor for input b, got ") << getB().getType();
+    return failure();
+  }
+
+  auto aElementType = aType.getElementType();
+  auto bElementType = bType.getElementType();
+
+  auto aQuantizedEType =
+      llvm::dyn_cast<quant::UniformQuantizedType>(aElementType);
+  auto bQuantizedEType =
+      llvm::dyn_cast<quant::UniformQuantizedType>(bElementType);
+
+  if (aQuantizedEType || bQuantizedEType) {
----------------
Jerry-Ge wrote:

>Probably worth breaking the if nesting?

I don't think so. I think the use of nested ifs is more elegant. 

>Check if not both quantized

There're 3 cases where NOT BOTH QUANTIZED is true. 
1. A is non-quanitzed, B is non-quantized
2. A is non-quantized, B is quantized 
3. A is quantized, B is non-quantized
For case 1, it's valid.  For case 2 and 3, they're not valid. 
If we don't use the nested ifs, the checking conditions for case 2 and 3 will be very similar and long which is not very elegant. 

What do you think? Not sure if you have a better way to organize this. 






https://github.com/llvm/llvm-project/pull/127923