[Mlir-commits] [mlir] [mlir][spirv][tosa] Add remaining TOSA 1.0 SPIR-V TOSA ops (PR #200383)

[Igor Wodiany]

================
@@ -39,6 +58,80 @@ DenseIntElementsAttr getI32TensorArmAttr(ArrayRef<int32_t> values,
       values);
 }
 
+DenseIntElementsAttr getI32TensorArmAttr(ArrayRef<int64_t> values,
+                                         ConversionPatternRewriter &rewriter) {
+  SmallVector<int32_t> i32Values;
+  i32Values.reserve(values.size());
+  for (int64_t value : values)
+    i32Values.push_back(value);
+  return getI32TensorArmAttr(i32Values, rewriter);
+}
+
+FailureOr<DenseElementsAttr>
+convertDenseElementsAttr(DenseElementsAttr values, ShapedType convertedType) {
+  Type convertedElementType = convertedType.getElementType();
+  if (values.getElementType() == convertedElementType)
+    return values.reshape(convertedType);
+
+  // Constant attributes still have the source TOSA element type. Rebuild them
+  // for the converted SPIR-V tensor type, including integer type-converter
+  // changes such as index to i32, i4 to i8, and i48 to i64.
+  auto integerType = dyn_cast<IntegerType>(convertedElementType);
+  if (!integerType)
+    return failure();
+
+  // The SPIR-V ARM tensor type does not represent scalar shape constants
+  // directly, so model tensor<0xindex> as a one-element i32 tensor.
+  if (values.empty() && values.getElementType().isIndex())
+    return DenseIntElementsAttr::get(convertedType, {1});
+
+  DenseElementsAttr convertedValues =
+      values.mapValues(integerType, [&](const APInt &value) {
+        return value.sextOrTrunc(integerType.getWidth());
+      });
+  return convertedValues.reshape(convertedType);
+}
+
+LogicalResult splitConcat(tosa::ConcatOp op, Type resultType, int32_t axis,
----------------
IgWod wrote:

I think some description of what the function does would be helpful

https://github.com/llvm/llvm-project/pull/200383