Unify memeff code with CUTLASS

ghstack-source-id: f3aa0ec09f8f763d6b7150781e485d1bbdc277d9
Pull Request resolved: https://github.com/fairinternal/xformers/pull/431

__original_commit__ = fairinternal/xformers@0032bbc276ba1696a398e4da8eb4fc5ddfd37477

xformers/csrc/attention/cuda/fmha/debug_utils.h

@@ -61,6 +61,7 @@ struct __string_view {
   char const* data;
   std::size_t size;
 };
+#if __cplusplus >= 201402L
 template <class T>
 constexpr __string_view __get_type_name() {
   char const* p = __PRETTY_FUNCTION__;
@@ -83,6 +84,12 @@ constexpr __string_view __get_type_name() {
   }
   return {};
 }
+#else
+template <class T>
+constexpr __string_view __get_type_name() {
+  return {"unsupported", 11};
+}
+#endif
 
 // Print a given array
 #define PRINT_ACCUM8_T0_L0_START(name, accum, start)  \

xformers/csrc/attention/cuda/fmha/find_default_mma.h

@@ -2,7 +2,7 @@
     \brief Cutlass provides helper template functions to figure out the right
    datastructures to instanciate to run a GEMM with various parameters (see
    `cutlass/gemm/threadblock/default_mma.h`). However, due to template
-   instanciation priority rules, it will only create an MmaMultiStage with
+   instantiation priority rules, it will only create an MmaMultiStage with
    kStages=3 (otherwise creates an MmePipelined - which is not compatible with
    FastF32). kStages=3 uses too much shared memory and we want to use kStages=2,
    so we just copy-pasted some code from `default_mma.h` and