[Mlir-commits] [mlir] [mlir][spirv] Split codegen for float min/max reductions and others (NFC) (PR #69023)

[Jakub Kuderski]

================
@@ -381,9 +393,26 @@ struct VectorReductionPattern final
     if (Value acc = adaptor.getAcc())
       values.push_back(acc);
 
-    // Reduce them.
-    Value result = values.front();
-    for (Value next : llvm::ArrayRef(values).drop_front()) {
+    return values;
+  }
+};
+
+#define VECTOR_REDUCTION_BASE                                                  \
+  VectorReductionPatternBase<VectorReductionPattern<SPIRVUMaxOp, SPIRVUMinOp,  \
+                                                    SPIRVSMaxOp, SPIRVSMinOp>>
----------------
kuhar wrote:

How about something like this:
```c++
template <typename... Args>
struct VectorReductionPattern;

typename <typename... Args>
using CrtpBase = VectorReductionPatternBase<VectorReductionPattern, Args...>;

template <typename... Args>
struct VectorReductionPattern final : CrtpBase<Args...> {
  using Base = CrtpBase<Args...>; 
   using Base::Base;
   // define SPIRVUMaxOp et al. in VectorReductionPatternBase
  ...
};
```

just an idea, I haven't tried to compile this

https://github.com/llvm/llvm-project/pull/69023