[CODEGEN/DRIVER] Tweaks for performance optimization (#193)

python/src/triton.cc

@@ -161,6 +161,7 @@ void init_triton_frontend(py::module &&m) {
   m.def("sqrt", &ir::dispatch::sqrt, ret::reference);
   // internal (debugging only)
   m.def("multiple_of", &ir::dispatch::multiple_of, ret::reference);
+  m.def("max_contiguous", &ir::dispatch::max_contiguous, ret::reference);
   m.def("debug_barrier", &ir::dispatch::debug_barrier, ret::reference);
 }
 

python/triton/language.py

@@ -637,6 +637,14 @@ def multiple_of(input, value, builder=None):
     return frontend.multiple_of(input, value, builder)
 
 
+@builtin
+def max_contiguous(input, value, builder=None):
+    """
+    Let the compiler knows that the `value` first values in :code:`input` are contiguous. 
+    """
+    return frontend.max_contiguous(input, value, builder)
+
+
 # -----------------------
 # Standard library
 # -----------------------