[Intel] Remove incorrect permute implementation (#5001)

The existing implementation is specifically designed for AMD.

Signed-off-by: Whitney Tsang <[email protected]>

Author: whitneywhtsang

third_party/intel/lib/TritonIntelGPUToLLVM/Utility.cpp

@@ -103,19 +103,7 @@ Value shuffleIdx(Location loc, RewriterBase &rewriter, Value val, Value i) {
 
 Value permute(Location loc, RewriterBase &rewriter, Value x, Value y,
               Value selector) {
-  auto b = TritonLLVMOpBuilder(loc, rewriter);
-  Value prmt_mask = selector;
-  // convert from nybble mask to byte mask:
-  prmt_mask =
-      b.or_(b.and_(prmt_mask, b.i32_val(0x000000ff)),
-            b.shl(b.and_(prmt_mask, b.i32_val(0x0000ff00)), b.i32_val(8)));
-  prmt_mask =
-      b.or_(b.and_(prmt_mask, b.i32_val(0x000f000f)),
-            b.shl(b.and_(prmt_mask, b.i32_val(0x00f000f0)), b.i32_val(4)));
-  Value args[] = {x, y, prmt_mask};
-  auto op = createLLVMIntrinsicCallOp(rewriter, loc, "llvm.amdgcn.perm", i32_ty,
-                                      args);
-  return op.getResult(0);
+  llvm_unreachable("not implemented");
 }
 
 LLVM::RoundingMode