[nvgpu] Expose helper code from NVGPU to LLVM lowering (NFC). (#4235)

OpenXLA would like to reuse code from `NVGPUOpPatternBase` for Sparsity
lowering. Currently this is done by patching `NVGPUToLLVMPass.cpp` to
add an extra pattern deriving from `NVGPUOpPatternBase`. Instead, we
would like to move this pattern to the downstream OpenXLA repository.

This change removes the `NVGPUOpPatternBase` base class and exposes the
core lowering code as a utility function (`rewriteAsPtxAsm()`) instead.
The existing patterns now directly inherit from `mlir::OpRewritePattern`
and use this utility function.

I hope that exposing an extra function is acceptable. I tried to balance
it with cleaning up the code a bit (e.g., no more CRTP, but of course
that's subjective).

third_party/nvidia/include/NVGPUToLLVM/NVGPUToLLVMPass.h

@@ -2,6 +2,14 @@
 #define TRITON_CONVERSION_NVGPU_TO_LLVM_PASS_H
 
 #include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "mlir/IR/Operation.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/IR/Value.h"
+#include "mlir/Support/LogicalResult.h"
 
 namespace mlir {
 
@@ -10,6 +18,19 @@ template <typename T> class OperationPass;
 
 namespace triton {
 
+namespace nvgpu {
+
+using Constraints = std::vector<std::string>;
+using OperandsAndConstraints = std::vector<std::pair<Value, std::string>>;
+
+LogicalResult
+rewriteAsPtxAsm(mlir::Operation *op, mlir::PatternRewriter &rewriter,
+                std::string ptxAsm,
+                const OperandsAndConstraints &operandsAndConstraints = {},
+                const Constraints &outputConstraints = {});
+
+} // namespace nvgpu
+
 std::unique_ptr<OperationPass<ModuleOp>> createConvertNVGPUToLLVMPass();
 
 } // namespace triton