[Mlir-commits] [mlir] [MLIR][ROCDL] Add conversion for gpu.subgroup_id to ROCDL (PR #136405)

[Jakub Kuderski]

================
@@ -190,6 +201,49 @@ struct GPUShuffleOpLowering : public ConvertOpToLLVMPattern<gpu::ShuffleOp> {
   }
 };
 
+struct GPUSubgroupIdOpToROCDL final
+    : ConvertOpToLLVMPattern<gpu::SubgroupIdOp> {
+  using ConvertOpToLLVMPattern::ConvertOpToLLVMPattern;
+
+  GPUSubgroupIdOpToROCDL(const LLVMTypeConverter &converter,
+                         const mlir::amdgpu::Chipset &chipset)
+      : ConvertOpToLLVMPattern(converter), chipset(chipset) {}
+
+  const mlir::amdgpu::Chipset chipset;
+
+  LogicalResult
+  matchAndRewrite(gpu::SubgroupIdOp op, gpu::SubgroupIdOp::Adaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    auto int32Type = IntegerType::get(rewriter.getContext(), 32);
+    auto loc = op.getLoc();
+    LLVM::IntegerOverflowFlags flags =
+        LLVM::IntegerOverflowFlags::nsw | LLVM::IntegerOverflowFlags::nuw;
+    // w_id.x + w_dim.x * (w_id.y + w_dim.y * w_id.z)) / subgroup_size
----------------
kuhar wrote:

The parentheses seem off -- I think the opening one is missing.
Also, maybe add a comment explaining **why** this should work? I'm thinking something along the lines of 'subgroup ids match linearized block id ranges`.

https://github.com/llvm/llvm-project/pull/136405