Use CUDA runtime API to retrieve function pointer to driver API (#1700)

* Query pfn to driver api

* use default for older toolkits

---------

Co-authored-by: shunfans <shunfans@nvidia.com>

include/cute/atom/copy_traits_sm90_im2col.hpp

@@ -40,6 +40,8 @@
 
 #include "cute/algorithm/prefetch.hpp"
 #include "cutlass/fast_math.h"
+#include "cutlass/cuda_host_adapter.hpp"
+
 namespace cute
 {
 
@@ -450,7 +452,7 @@ make_im2col_tma_copy_desc(
   CUtensorMapFloatOOBfill tma_oob_fill    = to_CUtensorMapFloatOOBfill(aux_params.oobfill_);
   CUtensorMapSwizzle      tma_swizzle     = TMA::to_CUtensorMapSwizzle(detail::get_tma_swizzle_bits(smem_swizzle));
 
-  CUresult encode_result = cuTensorMapEncodeIm2col(
+  CUresult encode_result = CUTLASS_CUDA_DRIVER_WRAPPER_CALL(cuTensorMapEncodeIm2col)(
       &tma_desc,
       tma_format,
       num_total_modes,

include/cute/atom/copy_traits_sm90_tma.hpp

@@ -41,6 +41,7 @@
 #include <cute/algorithm/prefetch.hpp>
 
 #include <cute/numeric/integral_ratio.hpp>
+#include <cutlass/cuda_host_adapter.hpp>
 
 namespace cute
 {
@@ -983,7 +984,7 @@ make_tma_copy_desc(Tensor<GEngine,GLayout> const& gtensor,         // The origin
 
     // TMA smem swizzle type
     CUtensorMapSwizzle smem_swizzle = TMA::to_CUtensorMapSwizzle(get_tma_swizzle_bits(swizzle));
-    CUresult result = cuTensorMapEncodeTiled(
+    CUresult result = CUTLASS_CUDA_DRIVER_WRAPPER_CALL(cuTensorMapEncodeTiled)(
         &tma_desc,
         tma_format,
         tma_dim,