[Mlir-commits] [mlir] [mlir][gpu] Add `subgroup_broadcast` op (PR #152808)

[Jakub Kuderski]

================
@@ -171,6 +171,27 @@ struct GPUSubgroupSizeOpToROCDL : ConvertOpToLLVMPattern<gpu::SubgroupSizeOp> {
   const amdgpu::Chipset chipset;
 };
 
+struct GPUSubgroupBroadcastOpToROCDL
+    : public ConvertOpToLLVMPattern<gpu::SubgroupBroadcastOp> {
+  using ConvertOpToLLVMPattern::ConvertOpToLLVMPattern;
+
+  LogicalResult
+  matchAndRewrite(gpu::SubgroupBroadcastOp op, OpAdaptor adaptor,
----------------
kuhar wrote:

We can always decompose vectors to the largest supported scalars, similar to subgroup_reduce

https://github.com/llvm/llvm-project/pull/152808