[Mlir-commits] [mlir] [MLIR][XeVM] XeVM to LLVM: Add conversion patterns for id ops (PR #162536)

[Sang Ik Lee]

================
@@ -714,6 +714,137 @@ class LLVMLoadStoreToOCLPattern : public OpConversionPattern<OpType> {
   }
 };
 
+//===----------------------------------------------------------------------===//
+// GPU index id operations
+//===----------------------------------------------------------------------===//
+/*
+// Launch Config ops
+//   dimidx - x, y, x - is fixed to i32
+//   return type is set by XeVM type converter
+// get_local_id
+xevm::WorkitemIdXOp;
+xevm::WorkitemIdYOp;
+xevm::WorkitemIdZOp;
+// get_local_size
+xevm::WorkgroupDimXOp;
+xevm::WorkgroupDimYOp;
+xevm::WorkgroupDimZOp;
+// get_group_id
+xevm::WorkgroupIdXOp;
+xevm::WorkgroupIdYOp;
+xevm::WorkgroupIdZOp;
+// get_num_groups
+xevm::GridDimXOp;
+xevm::GridDimYOp;
+xevm::GridDimZOp;
+// get_global_id : to be added if needed
+*/
+
+// Helpers to get the OpenCL function name and dimension argument for each op.
+static std::pair<StringRef, int64_t> getConfig(xevm::WorkitemIdXOp) {
----------------
silee2 wrote:

Agree. Intel GPU, without it's own LLVM backend, this pass is used as an alternative to tablegen based translation with `llvmBuilder`.
Although, XeVM to LLVM currently exist as a conversion,
In reality it is set up as a XeVM to LLVMOCL transform.
It is a transform since the source XeVM is LLVM (extension) dialect and the target is LLVM as well.
For SPIRV calls, it will require a separate transform like XeVM to LLVMSPV. 


https://github.com/llvm/llvm-project/pull/162536