[Mlir-commits] [mlir] [mlir][spirv] Add support for VectorAnyINTEL capability (PR #68034)

[Jakub Kuderski]

================
@@ -342,15 +397,35 @@ convertVectorType(const spirv::TargetEnv &targetEnv,
   cast<spirv::CompositeType>(type).getExtensions(extensions, storageClass);
   cast<spirv::CompositeType>(type).getCapabilities(capabilities, storageClass);
 
+  // If the bit-width related capabilities and extensions are not met
+  // for lower bit-width (<32-bit), convert it to 32-bit
+  auto elementType =
+      convertScalarType(targetEnv, options, scalarType, storageClass);
+  if (!elementType)
+    return nullptr;
+  type = VectorType::get(type.getShape(), elementType);
+
+  SmallVector<spirv::Capability, 4> elidedCaps;
+  SmallVector<spirv::Extension, 4> elidedExts;
+
+  // Relax the bitwidth requirements for capabilities and extensions
+  if (options.emulateLT32BitScalarTypes) {
+    elidedCaps.push_back(spirv::Capability::Int8);
+    elidedCaps.push_back(spirv::Capability::Int16);
+    elidedCaps.push_back(spirv::Capability::Float16);
+  }
+  // For capabilities whose requirements were relaxed, relax requirements for
+  // the extensions that were infered by those capabilities (e.g., elidedCaps)
+  for (spirv::Capability cap : elidedCaps) {
+    std::optional<ArrayRef<spirv::Extension>> ext = spirv::getExtensions(cap);
+    if (ext)
----------------
kuhar wrote:

```suggestion
    if (std::optional<ArrayRef<spirv::Extension>> ext = spirv::getExtensions(cap))
```

https://github.com/llvm/llvm-project/pull/68034